From 66862a03ffd86a3f9436742337a4ec1d6807d4c4 Mon Sep 17 00:00:00 2001 From: ftgreat Date: Tue, 28 Mar 2023 18:04:53 +0800 Subject: [PATCH 1/7] added altdiffusion_m18 Signed-off-by: ftgreat --- examples/AltDiffusion/generate.py | 19 +- examples/AltDiffusion/generate_18m.py | 28 + flagai/auto_model/auto_loader.py | 8 +- flagai/model/mm/AltCLIP.py | 1 + flagai/model/mm/AltDiffusion.py | 2 +- flagai/model/mm/AltDiffusion2.py | 1845 +++++++++++++++++ flagai/model/mm/Sampler.py | 2 + flagai/model/mm/autoencoder2.py | 219 ++ flagai/model/mm/dpm_solver/__init__.py | 1 + flagai/model/mm/dpm_solver/dpm_solver.py | 1154 +++++++++++ flagai/model/mm/dpm_solver/sampler.py | 87 + flagai/model/mm/lm/dcn_clip.py | 61 + flagai/model/mm/modules/attention.py | 343 +++ .../mm/modules/diffusionmodules/__init__.py | 0 .../mm/modules/diffusionmodules/model.py | 852 ++++++++ .../modules/diffusionmodules/openaimodel.py | 903 ++++++++ .../mm/modules/diffusionmodules/upscaling.py | 81 + .../model/mm/modules/diffusionmodules/util.py | 270 +++ .../mm/modules/distributions/__init__.py | 0 .../mm/modules/distributions/distributions.py | 92 + flagai/model/mm/modules/ema.py | 81 + flagai/model/mm/modules/encoders/__init__.py | 0 flagai/model/mm/modules/encoders/modules.py | 313 +++ flagai/model/mm/openaimodel.py | 904 ++++++++ flagai/model/mm/utils.py | 111 +- flagai/model/predictor/predictor.py | 113 +- 26 files changed, 7477 insertions(+), 13 deletions(-) create mode 100644 examples/AltDiffusion/generate_18m.py create mode 100644 flagai/model/mm/AltDiffusion2.py create mode 100644 flagai/model/mm/autoencoder2.py create mode 100644 flagai/model/mm/dpm_solver/__init__.py create mode 100644 flagai/model/mm/dpm_solver/dpm_solver.py create mode 100644 flagai/model/mm/dpm_solver/sampler.py create mode 100644 flagai/model/mm/lm/dcn_clip.py create mode 100644 flagai/model/mm/modules/attention.py create mode 100644 flagai/model/mm/modules/diffusionmodules/__init__.py create mode 100644 flagai/model/mm/modules/diffusionmodules/model.py create mode 100644 flagai/model/mm/modules/diffusionmodules/openaimodel.py create mode 100644 flagai/model/mm/modules/diffusionmodules/upscaling.py create mode 100644 flagai/model/mm/modules/diffusionmodules/util.py create mode 100644 flagai/model/mm/modules/distributions/__init__.py create mode 100644 flagai/model/mm/modules/distributions/distributions.py create mode 100644 flagai/model/mm/modules/ema.py create mode 100644 flagai/model/mm/modules/encoders/__init__.py create mode 100644 flagai/model/mm/modules/encoders/modules.py create mode 100644 flagai/model/mm/openaimodel.py diff --git a/examples/AltDiffusion/generate.py b/examples/AltDiffusion/generate.py index 1fa2c88f..f29524bd 100644 --- a/examples/AltDiffusion/generate.py +++ b/examples/AltDiffusion/generate.py @@ -1,6 +1,7 @@ # Copyright © 2022 BAAI. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License") +import sys;sys.path.append("/home/yanzhaodong/FlagAI") import torch from flagai.auto_model.auto_loader import AutoLoader from flagai.model.predictor.predictor import Predictor @@ -13,10 +14,16 @@ model_dir="./checkpoints", use_fp16=False) +loader2 = AutoLoader(task_name="text2img", model_name="AltDiffusion-m9") + model = loader.get_model() -model.eval() -model.to(device) -predictor = Predictor(model) -predictor.predict_generate_images( - "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" -) +import pdb;pdb.set_trace() +for name, param in model.named_parameters(): + if name.startswith("cond_stage_model"): + print(name) +# model.eval() +# model.to(device) +# predictor = Predictor(model) +# predictor.predict_generate_images( +# "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" +# ) diff --git a/examples/AltDiffusion/generate_18m.py b/examples/AltDiffusion/generate_18m.py new file mode 100644 index 00000000..766a0149 --- /dev/null +++ b/examples/AltDiffusion/generate_18m.py @@ -0,0 +1,28 @@ +# Copyright © 2022 BAAI. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License") +import sys;sys.path.append("/home/yanzhaodong/FlagAI") +import torch +from flagai.auto_model.auto_loader import AutoLoader +from flagai.model.predictor.predictor import Predictor + +# Initialize +device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + +loader = AutoLoader(task_name="text2img", #contrastive learning + model_name="AltDiffusion-m18", + model_dir="./checkpoints", + use_fp16=False) +# loader = AutoLoader(task_name="text2img", #contrastive learning +# model_name="AltDiffusion-m18") +model = loader.get_model() +# for name, param in model.named_parameters(): +# if name.startswith("cond_stage_model"): +# print(name) +import pdb;pdb.set_trace() +# model.eval() +# model.to(device) +# predictor = Predictor(model) +# predictor.predict_generate_images_m18( +# "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" +# ) diff --git a/flagai/auto_model/auto_loader.py b/flagai/auto_model/auto_loader.py index 71ac1c44..b17e00cd 100644 --- a/flagai/auto_model/auto_loader.py +++ b/flagai/auto_model/auto_loader.py @@ -65,6 +65,7 @@ def __getattr__(self, name): "cpm3_lm": ("flagai.model.cpm3_model", "CPM3"), "cpm3_train": ("flagai.model.cpm3_train_model", "CPM3"), "diffusion_text2img": ("flagai.model.mm.AltDiffusion", "LatentDiffusion"), + "diffusion2_text2img": ("flagai.model.mm.AltDiffusion2", "LatentDiffusion"), "altclip_txt_img_matching": ("flagai.model.mm.AltCLIP", "AltCLIP"), "evaclip_txt_img_matching": ("flagai.model.mm.eva_clip_model", "EVA_CLIP"), } @@ -121,7 +122,9 @@ def __getattr__(self, name): "altdiffusion": ["flagai.model.mm.diffusion", "LatentDiffusion", "diffusion", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], "altdiffusion-m9": - ["flagai.model.mm.diffusion", "LatentDiffusion", "diffusion", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], + ["flagai.model.mm.diffusion2", "LatentDiffusion", "diffusion", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], + "altdiffusion-m18": + ["flagai.model.mm.Altdiffusion2", "LatentDiffusion", "diffusion2", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], "swinv1-base-patch4-window7-224": ["flagai.model.vision.swinv1", "SwinTransformer", "swinv1", "vision"], "swinv2-base-patch4-window8-256": @@ -198,7 +201,6 @@ def __init__(self, f"For the model_name: {model_name}, these tasks are be supported: {tasks}" ) return - download_path = os.path.join(model_dir, raw_model_name) print("*" * 20, task_name, model_name) model_name_ = self.is_exist_finetuned_model(raw_model_name, task_name) @@ -211,7 +213,7 @@ def __init__(self, **kwargs) if kwargs.get("use_fp16", None): self.model.half() - + if model_type == "nlp": if brief_model_name in ["galactica", ]: self.tokenizer = getattr(LazyImport(MODEL_DICT[model_name][4]), diff --git a/flagai/model/mm/AltCLIP.py b/flagai/model/mm/AltCLIP.py index 9757e2ff..34251ab7 100644 --- a/flagai/model/mm/AltCLIP.py +++ b/flagai/model/mm/AltCLIP.py @@ -81,6 +81,7 @@ def __init__(self, if text_config_dict is None: text_config_dict = {} # when reload the config from local, we need name to select which class should be instanced. + import pdb;pdb.set_trace() self.text_config = STUDENT_CONFIG_DICT[ kwargs['text_config']['model_type']](**kwargs.pop('text_config')) self.num_layers = num_layers diff --git a/flagai/model/mm/AltDiffusion.py b/flagai/model/mm/AltDiffusion.py index e223c4c2..9a6b12cf 100644 --- a/flagai/model/mm/AltDiffusion.py +++ b/flagai/model/mm/AltDiffusion.py @@ -1932,4 +1932,4 @@ def normal_kl(mean1, logvar1, mean2, logvar2): ] return 0.5 * (-1.0 + logvar2 - logvar1 + torch.exp(logvar1 - logvar2) + - ((mean1 - mean2)**2) * torch.exp(-logvar2)) + ((mean1 - mean2)**2) * torch.exp(-logvar2)) \ No newline at end of file diff --git a/flagai/model/mm/AltDiffusion2.py b/flagai/model/mm/AltDiffusion2.py new file mode 100644 index 00000000..4c17b7e4 --- /dev/null +++ b/flagai/model/mm/AltDiffusion2.py @@ -0,0 +1,1845 @@ +import sys +import torch +import torch.nn as nn +import numpy as np +from torch.optim.lr_scheduler import LambdaLR +from einops import rearrange, repeat +from contextlib import contextmanager +from functools import partial +from tqdm import tqdm +from torchvision.utils import make_grid +from pytorch_lightning.utilities.distributed import rank_zero_only +from flagai.model.mm.utils import exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config, log_txt_as_img +from flagai.model.mm.autoencoders import VQModelInterface, IdentityFirstStage, AutoencoderKL +from flagai.model.mm.utils import make_beta_schedule, extract_into_tensor, noise_like +from flagai.model.mm.Sampler import DDIMSampler +from flagai.model.base_model import BaseModel +from torch.cuda.amp import autocast as autocast +import pytorch_lightning as pl + +__conditioning_keys__ = { + 'concat': 'c_concat', + 'crossattn': 'c_crossattn', + 'adm': 'y' +} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def uniform_on_device(r1, r2, shape, device): + return (r1 - r2) * torch.rand(*shape, device=device) + r2 + + +class DDPM(BaseModel): + # classic DDPM with Gaussian diffusion, in image space + def __init__( + self, + unet_config, + timesteps=1000, + beta_schedule="linear", + loss_type="l2", + ckpt_path=None, + ignore_keys=[], + load_only_unet=False, + monitor="val/loss", + use_ema=False, + first_stage_key="image", + image_size=256, + channels=3, + log_every_t=100, + clip_denoised=True, + linear_start=1e-4, + linear_end=2e-2, + cosine_s=8e-3, + given_betas=None, + original_elbo_weight=0., + v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta + l_simple_weight=1., + conditioning_key=None, + parameterization="eps", # all assuming fixed variance schedules + scheduler_config=None, + use_positional_encodings=False, + learn_logvar=False, + logvar_init=0., + **kwargs, + ): + super(DDPM, self).__init__(unet_config, **kwargs) + unet_config.params.update(kwargs) + assert parameterization in [ + "eps", "x0" + ], 'currently only supporting "eps" and "x0"' + self.parameterization = parameterization + print( + f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode" + ) + self.cond_stage_model = None + self.clip_denoised = clip_denoised + self.log_every_t = log_every_t + self.first_stage_key = first_stage_key + self.image_size = image_size # try conv? + self.channels = channels + self.use_positional_encodings = use_positional_encodings + self.model = DiffusionWrapper(unet_config, conditioning_key) + self.use_ema = use_ema + count_params(self.model, verbose=True) + + self.use_scheduler = scheduler_config is not None + if self.use_scheduler: + self.scheduler_config = scheduler_config + + self.v_posterior = v_posterior + self.original_elbo_weight = original_elbo_weight + self.l_simple_weight = l_simple_weight + + if monitor is not None: + self.monitor = monitor + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, + ignore_keys=ignore_keys, + only_model=load_only_unet) + + self.register_schedule(given_betas=given_betas, + beta_schedule=beta_schedule, + timesteps=timesteps, + linear_start=linear_start, + linear_end=linear_end, + cosine_s=cosine_s) + + self.loss_type = loss_type + + self.learn_logvar = learn_logvar + self.logvar = torch.full(fill_value=logvar_init, + size=(self.num_timesteps, )) + if self.learn_logvar: + self.logvar = nn.Parameter(self.logvar, requires_grad=True) + + def register_schedule(self, + given_betas=None, + beta_schedule="linear", + timesteps=1000, + linear_start=1e-4, + linear_end=2e-2, + cosine_s=8e-3): + if exists(given_betas): + betas = given_betas + else: + betas = make_beta_schedule(beta_schedule, + timesteps, + linear_start=linear_start, + linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[ + 0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', + to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', + to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', + to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', + to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', + to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', + to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + # calculations for posterior q(x_{t-1} | x_t, x_0) + posterior_variance = (1 - self.v_posterior) * betas * ( + 1. - alphas_cumprod_prev) / ( + 1. - alphas_cumprod) + self.v_posterior * betas + # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) + self.register_buffer('posterior_variance', + to_torch(posterior_variance)) + # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain + self.register_buffer( + 'posterior_log_variance_clipped', + to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) + self.register_buffer( + 'posterior_mean_coef1', + to_torch(betas * np.sqrt(alphas_cumprod_prev) / + (1. - alphas_cumprod))) + self.register_buffer( + 'posterior_mean_coef2', + to_torch((1. - alphas_cumprod_prev) * np.sqrt(alphas) / + (1. - alphas_cumprod))) + + if self.parameterization == "eps": + lvlb_weights = self.betas**2 / (2 * self.posterior_variance * + to_torch(alphas) * + (1 - self.alphas_cumprod)) + elif self.parameterization == "x0": + lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / ( + 2. * 1 - torch.Tensor(alphas_cumprod)) + else: + raise NotImplementedError("mu not supported") + # TODO how to choose this term + lvlb_weights[0] = lvlb_weights[1] + self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) + assert not torch.isnan(self.lvlb_weights).all() + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.model.parameters()) + self.model_ema.copy_to(self.model) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.model.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + missing, unexpected = self.load_state_dict( + sd, + strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print( + f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys" + ) + if len(missing) > 0: + print(f"Missing Keys: {missing}") + if len(unexpected) > 0: + print(f"Unexpected Keys: {unexpected}") + + def q_mean_variance(self, x_start, t): + """ + Get the distribution q(x_t | x_0). + :param x_start: the [N x C x ...] tensor of noiseless inputs. + :param t: the number of diffusion steps (minus 1). Here, 0 means one step. + :return: A tuple (mean, variance, log_variance), all of x_start's shape. + """ + mean = ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * + x_start) + variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, + x_start.shape) + log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, + t, x_start.shape) + return mean, variance, log_variance + + def predict_start_from_noise(self, x_t, t, noise): + return ( + extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * + x_t - extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, + x_t.shape) * noise) + + def q_posterior(self, x_start, x_t, t): + posterior_mean = ( + extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * + x_start + + extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t) + posterior_variance = extract_into_tensor(self.posterior_variance, t, + x_t.shape) + posterior_log_variance_clipped = extract_into_tensor( + self.posterior_log_variance_clipped, t, x_t.shape) + return posterior_mean, posterior_variance, posterior_log_variance_clipped + + def p_mean_variance(self, x, t, clip_denoised: bool): + model_out = self.model(x, t) + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + if clip_denoised: + x_recon.clamp_(-1., 1.) + + model_mean, posterior_variance, posterior_log_variance = self.q_posterior( + x_start=x_recon, x_t=x, t=t) + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, t, clip_denoised=True, repeat_noise=False): + b, *_, device = *x.shape, x.device + model_mean, _, model_log_variance = self.p_mean_variance( + x=x, t=t, clip_denoised=clip_denoised) + noise = noise_like(x.shape, device, repeat_noise) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape( + b, *((1, ) * (len(x.shape) - 1))) + return model_mean + nonzero_mask * (0.5 * + model_log_variance).exp() * noise + + @torch.no_grad() + def p_sample_loop(self, shape, return_intermediates=False): + device = self.betas.device + b = shape[0] + img = torch.randn(shape, device=device) + intermediates = [img] + for i in tqdm(reversed(range(0, self.num_timesteps)), + desc='Sampling t', + total=self.num_timesteps): + img = self.p_sample(img, + torch.full((b, ), + i, + device=device, + dtype=torch.long), + clip_denoised=self.clip_denoised) + if i % self.log_every_t == 0 or i == self.num_timesteps - 1: + intermediates.append(img) + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, batch_size=16, return_intermediates=False): + image_size = self.image_size + channels = self.channels + return self.p_sample_loop( + (batch_size, channels, image_size, image_size), + return_intermediates=return_intermediates) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * + x_start + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, + t, x_start.shape) * noise) + + def get_loss(self, pred, target, mean=True): + if self.loss_type == 'l1': + loss = (target - pred).abs() + if mean: + loss = loss.mean() + elif self.loss_type == 'l2': + if mean: + loss = torch.nn.functional.mse_loss(target, pred) + else: + loss = torch.nn.functional.mse_loss(target, + pred, + reduction='none') + else: + raise NotImplementedError("unknown loss type '{loss_type}'") + + return loss + + def p_losses(self, x_start, t, noise=None): + + noise = default(noise, lambda: torch.randn_like(x_start)) + # print(f"after get noise: {torch.cuda.memory_allocated()/1024/1024/1024:.2f}GB") + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + # print(f"after sampled noise: {torch.cuda.memory_allocated()/1024/1024/1024:.2f}GB") + model_out = self.model(x_noisy, t) + # print(f"after predict noise: {torch.cuda.memory_allocated()/1024/1024/1024:.2f}GB") + + loss_dict = {} + if self.parameterization == "eps": + target = noise + elif self.parameterization == "x0": + target = x_start + else: + raise NotImplementedError( + f"Paramterization {self.parameterization} not yet supported") + + loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3]) + # print(f"after get loss: {torch.cuda.memory_allocated()/1024/1024/1024:.2f}GB") + + log_prefix = 'train' if self.training else 'val' + + loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()}) + loss_simple = loss.mean() * self.l_simple_weight + + loss_vlb = (self.lvlb_weights[t] * loss).mean() + loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb}) + + loss = loss_simple + self.original_elbo_weight * loss_vlb + + loss_dict.update({f'{log_prefix}/loss': loss}) + + return loss, loss_dict + + def forward(self, x, *args, **kwargs): + # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size + # assert h == img_size and w == img_size, f'height and width of image must be {img_size}' + t = torch.randint(0, + self.num_timesteps, (x.shape[0], ), + device=self.device).long() + return self.p_losses(x, t, *args, **kwargs) + + def get_input(self, batch, k): + # The k here stands for key, and is primarily used to handle the case that batch contains both image and caption + x = batch[k] + # In unsupervised learning, k is not used + #x = batch + if len(x.shape) == 3: + x = x[..., None] + # dataseet中已经处理了图片顺序，不需要再做rearrange + # x = rearrange(x, 'b h w c -> b c h w') + x = x.to(memory_format=torch.contiguous_format).float() + return x + + def shared_step(self, batch): + x = self.get_input(batch, self.first_stage_key) + loss, loss_dict = self(x) + return loss, loss_dict + + def training_step(self, batch, batch_idx): + + loss, loss_dict = self.shared_step(batch) + self.log_dict(loss_dict, + prog_bar=True, + logger=True, + on_step=True, + on_epoch=True) + try: + self.log("global_step", + self.global_step, + prog_bar=True, + logger=True, + on_step=True, + on_epoch=False) + + if self.use_scheduler: + lr = self.optimizers().param_groups[0]['lr'] + self.log('lr_abs', + lr, + prog_bar=True, + logger=True, + on_step=True, + on_epoch=False) + except: + return loss + + return loss + + @torch.no_grad() + def validation_step(self, batch, batch_idx): + _, loss_dict_no_ema = self.shared_step(batch) + with self.ema_scope(): + _, loss_dict_ema = self.shared_step(batch) + loss_dict_ema = { + key + '_ema': loss_dict_ema[key] + for key in loss_dict_ema + } + self.log_dict(loss_dict_no_ema, + prog_bar=False, + logger=True, + on_step=False, + on_epoch=True) + self.log_dict(loss_dict_ema, + prog_bar=False, + logger=True, + on_step=False, + on_epoch=True) + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self.model) + + def _get_rows_from_list(self, samples): + n_imgs_per_row = len(samples) + denoise_grid = rearrange(samples, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + @torch.no_grad() + def log_images(self, + batch, + N=8, + n_row=2, + sample=True, + return_keys=None, + **kwargs): + log = dict() + x = self.get_input(batch, self.first_stage_key) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + x = x.to(self.device)[:N] + log["inputs"] = x + + # get diffusion row + diffusion_row = list() + x_start = x[:n_row] + + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(x_start) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + diffusion_row.append(x_noisy) + + log["diffusion_row"] = self._get_rows_from_list(diffusion_row) + + if sample: + # get denoise row + with self.ema_scope("Plotting"): + samples, denoise_row = self.sample(batch_size=N, + return_intermediates=True) + + log["samples"] = samples + log["denoise_row"] = self._get_rows_from_list(denoise_row) + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + +class LatentDiffusion(DDPM): + """main class""" + + def to(self, device): + self.device = device + self.cond_stage_model.to(device) + super().to(device) + + def __init__(self, + first_stage_config, + cond_stage_config, + num_timesteps_cond=None, + cond_stage_key="image", + cond_stage_trainable=False, + concat_mode=True, + cond_stage_forward=None, + conditioning_key=None, + scale_factor=1.0, + scale_by_std=False, + tokenizer=None, + force_null_conditioning=False, + *args, + **kwargs): + self.device = "cpu" + self.tokenizer = tokenizer + self.force_null_conditioning = force_null_conditioning + self.num_timesteps_cond = default(num_timesteps_cond, 1) + self.scale_by_std = scale_by_std + assert self.num_timesteps_cond <= kwargs['timesteps'] + # for backwards compatibility after implementation of DiffusionWrapper + if conditioning_key is None: + conditioning_key = 'concat' if concat_mode else 'crossattn' + if cond_stage_config == '__is_unconditional__' and not self.force_null_conditioning: + conditioning_key = None + + ckpt_path = kwargs.pop("ckpt_path", None) + reset_ema = kwargs.pop("reset_ema", False) + reset_num_ema_updates = kwargs.pop("reset_num_ema_updates", False) + ignore_keys = kwargs.pop("ignore_keys", []) + super().__init__(conditioning_key=conditioning_key, *args, **kwargs) + self.concat_mode = concat_mode + self.cond_stage_trainable = cond_stage_trainable + self.cond_stage_key = cond_stage_key + self.conditioning_key = conditioning_key + try: + self.num_downs = len( + first_stage_config.params.ddconfig.ch_mult) - 1 + except: + self.num_downs = 0 + if not scale_by_std: + self.scale_factor = scale_factor + else: + self.register_buffer('scale_factor', torch.tensor(scale_factor)) + self.instantiate_first_stage(first_stage_config) + self.instantiate_cond_stage(cond_stage_config) + self.cond_stage_forward = cond_stage_forward + # if self.cond_stage_forward is None: + # self.set_cond_stage_forward() + self.clip_denoised = False + self.bbox_tokenizer = None + + self.restarted_from_ckpt = False + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys) + self.restarted_from_ckpt = True + if reset_ema: + assert self.use_ema + print( + f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.") + self.model_ema = LitEma(self.model) + if reset_num_ema_updates: + print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ") + assert self.use_ema + self.model_ema.reset_num_updates() + + def make_cond_schedule(self, ): + self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long) + ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long() + self.cond_ids[:self.num_timesteps_cond] = ids + + @rank_zero_only + @torch.no_grad() + # def on_train_batch_start(self, batch, batch_idx, dataloader_idx): + def on_train_batch_start(self, batch, batch_idx): + # only for very first batch + if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt: + assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously' + # set rescale weight to 1./std of encodings + print("### USING STD-RESCALING ###") + x = super().get_input(batch, self.first_stage_key) + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + del self.scale_factor + self.register_buffer('scale_factor', 1. / z.flatten().std()) + print(f"setting self.scale_factor to {self.scale_factor}") + print("### USING STD-RESCALING ###") + + def on_train_epoch_start(self) -> None: + # print(self.current_epoch) + if 'set_epoch' in dir(self.data.data["train"]): + self.data.data["train"].set_epoch(self.current_epoch) + # self.data.data["val"].set_epoch(self.current_epoch) + + def register_schedule(self, + given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s) + + self.shorten_cond_schedule = self.num_timesteps_cond > 1 + if self.shorten_cond_schedule: + self.make_cond_schedule() + + def instantiate_first_stage(self, config): + model = instantiate_from_config(config) + self.first_stage_model = model.eval() + self.first_stage_model.train = disabled_train + for param in self.first_stage_model.parameters(): + param.requires_grad = False + + def instantiate_cond_stage(self, config): + if not self.cond_stage_trainable: + if config == "__is_first_stage__": + print("Using first stage also as cond stage.") + self.cond_stage_model = self.first_stage_model + elif config == "__is_unconditional__": + print(f"Training {self.__class__.__name__} as an unconditional model.") + self.cond_stage_model = None + # self.be_unconditional = True + else: + model = instantiate_from_config(config) + self.cond_stage_model = model.eval() + self.cond_stage_model.train = disabled_train + for param in self.cond_stage_model.parameters(): + param.requires_grad = False + else: + assert config != '__is_first_stage__' + assert config != '__is_unconditional__' + model = instantiate_from_config(config) + self.cond_stage_model = model + + def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False): + denoise_row = [] + for zd in tqdm(samples, desc=desc): + denoise_row.append(self.decode_first_stage(zd.to(self.device), + force_not_quantize=force_no_decoder_quantization)) + n_imgs_per_row = len(denoise_row) + denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W + denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + def get_first_stage_encoding(self, encoder_posterior): + if isinstance(encoder_posterior, DiagonalGaussianDistribution): + z = encoder_posterior.sample() + elif isinstance(encoder_posterior, torch.Tensor): + z = encoder_posterior + else: + raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") + return self.scale_factor * z + + def get_learned_conditioning(self, c): + if self.cond_stage_forward is None: + if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): + c = self.cond_stage_model.encode(c) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + else: + c = self.cond_stage_model(c) + else: + assert hasattr(self.cond_stage_model, self.cond_stage_forward) + c = getattr(self.cond_stage_model, self.cond_stage_forward)(c) + return c + + def meshgrid(self, h, w): + y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1) + x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1) + + arr = torch.cat([y, x], dim=-1) + return arr + + def delta_border(self, h, w): + """ + :param h: height + :param w: width + :return: normalized distance to image border, + wtith min distance = 0 at border and max dist = 0.5 at image center + """ + lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2) + arr = self.meshgrid(h, w) / lower_right_corner + dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0] + dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0] + edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0] + return edge_dist + + def get_weighting(self, h, w, Ly, Lx, device): + weighting = self.delta_border(h, w) + weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"], + self.split_input_params["clip_max_weight"], ) + weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device) + + if self.split_input_params["tie_braker"]: + L_weighting = self.delta_border(Ly, Lx) + L_weighting = torch.clip(L_weighting, + self.split_input_params["clip_min_tie_weight"], + self.split_input_params["clip_max_tie_weight"]) + + L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device) + weighting = weighting * L_weighting + return weighting + + def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code + """ + :param x: img of size (bs, c, h, w) + :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1]) + """ + bs, nc, h, w = x.shape + + # number of crops in image + Ly = (h - kernel_size[0]) // stride[0] + 1 + Lx = (w - kernel_size[1]) // stride[1] + 1 + + if uf == 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params) + + weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx)) + + elif uf > 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf), + dilation=1, padding=0, + stride=(stride[0] * uf, stride[1] * uf)) + fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx)) + + elif df > 1 and uf == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df), + dilation=1, padding=0, + stride=(stride[0] // df, stride[1] // df)) + fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx)) + + else: + raise NotImplementedError + + return fold, unfold, normalization, weighting + + @torch.no_grad() + def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False, + cond_key=None, return_original_cond=False, bs=None, return_x=False): + x = super().get_input(batch, k) + if bs is not None: + x = x[:bs] + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + + if self.model.conditioning_key is not None and not self.force_null_conditioning: + + if isinstance(batch, list): + xc = batch[1] + else: + if cond_key is None: + cond_key = self.cond_stage_key + if cond_key != self.first_stage_key: + if cond_key in ['caption', 'coordinates_bbox', "txt"]: + xc = batch[cond_key] + elif cond_key in ['class_label', 'cls']: + xc = batch + else: + xc = super().get_input(batch, cond_key).to(self.device) + else: + xc = x + if not self.cond_stage_trainable or force_c_encode: + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + c = self.get_learned_conditioning(xc.to(self.device)) + else: + c = xc + if bs is not None: + c = c[:bs] + + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + ckey = __conditioning_keys__[self.model.conditioning_key] + c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y} + else: + c = None + xc = None + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + c = {'pos_x': pos_x, 'pos_y': pos_y} + out = [z, c] + if return_first_stage_outputs: + xrec = self.decode_first_stage(z) + out.extend([x, xrec]) + if return_x: + out.extend([x]) + if return_original_cond: + out.append(xc) + return out + + @torch.no_grad() + def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False): + if predict_cids: + if z.dim() == 4: + z = torch.argmax(z.exp(), dim=1).long() + z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None) + z = rearrange(z, 'b h w c -> b c h w').contiguous() + + z = 1. / self.scale_factor * z + return self.first_stage_model.decode(z) + + @torch.no_grad() + def encode_first_stage(self, x): + return self.first_stage_model.encode(x) + + def shared_step(self, batch, **kwargs): + x, c = self.get_input(batch, self.first_stage_key) + loss = self(x, c) + return loss + + def forward(self, x, c, *args, **kwargs): + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + if self.model.conditioning_key is not None: + assert c is not None + if self.cond_stage_trainable: + c = self.get_learned_conditioning(c) + if self.shorten_cond_schedule: # TODO: drop this option + tc = self.cond_ids[t].to(self.device) + c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float())) + return self.p_losses(x, c, t, *args, **kwargs) + + def apply_model(self, x_noisy, t, cond, return_ids=False): + if isinstance(cond, dict): + # hybrid case, cond is expected to be a dict + pass + else: + if not isinstance(cond, list): + cond = [cond] + key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn' + cond = {key: cond} + + x_recon = self.model(x_noisy, t, **cond) + + if isinstance(x_recon, tuple) and not return_ids: + return x_recon[0] + else: + return x_recon + + def _predict_eps_from_xstart(self, x_t, t, pred_xstart): + return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \ + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) + + def _prior_bpd(self, x_start): + """ + Get the prior KL term for the variational lower-bound, measured in + bits-per-dim. + This term can't be optimized, as it only depends on the encoder. + :param x_start: the [N x C x ...] tensor of inputs. + :return: a batch of [N] KL values (in bits), one per batch element. + """ + batch_size = x_start.shape[0] + t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device) + qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t) + kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0) + return mean_flat(kl_prior) / np.log(2.0) + + def p_losses(self, x_start, cond, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_output = self.apply_model(x_noisy, t, cond) + + loss_dict = {} + prefix = 'train' if self.training else 'val' + + if self.parameterization == "x0": + target = x_start + elif self.parameterization == "eps": + target = noise + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError() + + loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3]) + loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()}) + + # intresting thing: torch 1.12 and cu113 will not call a error,but 1.13.1+cu117 + # so I use the code bellow to change t to cpu and fixed the bug easily + logvar_t = self.logvar[t.cpu()].to(self.device) + loss = loss_simple / torch.exp(logvar_t) + logvar_t + # loss = loss_simple / torch.exp(self.logvar) + self.logvar + if self.learn_logvar: + loss_dict.update({f'{prefix}/loss_gamma': loss.mean()}) + loss_dict.update({'logvar': self.logvar.data.mean()}) + + loss = self.l_simple_weight * loss.mean() + + loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3)) + loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean() + loss_dict.update({f'{prefix}/loss_vlb': loss_vlb}) + loss += (self.original_elbo_weight * loss_vlb) + loss_dict.update({f'{prefix}/loss': loss}) + + return loss, loss_dict + + def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False, + return_x0=False, score_corrector=None, corrector_kwargs=None): + t_in = t + model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids) + + if score_corrector is not None: + assert self.parameterization == "eps" + model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs) + + if return_codebook_ids: + model_out, logits = model_out + + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + else: + raise NotImplementedError() + + if clip_denoised: + x_recon.clamp_(-1., 1.) + if quantize_denoised: + x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon) + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + if return_codebook_ids: + return model_mean, posterior_variance, posterior_log_variance, logits + elif return_x0: + return model_mean, posterior_variance, posterior_log_variance, x_recon + else: + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, + return_codebook_ids=False, quantize_denoised=False, return_x0=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None): + b, *_, device = *x.shape, x.device + outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, + return_codebook_ids=return_codebook_ids, + quantize_denoised=quantize_denoised, + return_x0=return_x0, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + if return_codebook_ids: + raise DeprecationWarning("Support dropped.") + model_mean, _, model_log_variance, logits = outputs + elif return_x0: + model_mean, _, model_log_variance, x0 = outputs + else: + model_mean, _, model_log_variance = outputs + + noise = noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + + if return_codebook_ids: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1) + if return_x0: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0 + else: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False, + img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0., + score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None, + log_every_t=None): + if not log_every_t: + log_every_t = self.log_every_t + timesteps = self.num_timesteps + if batch_size is not None: + b = batch_size if batch_size is not None else shape[0] + shape = [batch_size] + list(shape) + else: + b = batch_size = shape[0] + if x_T is None: + img = torch.randn(shape, device=self.device) + else: + img = x_T + intermediates = [] + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation', + total=timesteps) if verbose else reversed( + range(0, timesteps)) + if type(temperature) == float: + temperature = [temperature] * timesteps + + for i in iterator: + ts = torch.full((b,), i, device=self.device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img, x0_partial = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised, return_x0=True, + temperature=temperature[i], noise_dropout=noise_dropout, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + # 以下几行是img2img中mask功能的关键思路 + if mask is not None: + assert x0 is not None + img_orig = self.q_sample(x0, ts) # ts时刻的噪声图，q是前向过程 + img = img_orig * mask + (1. - mask) * img # mask部分用根据cond生成的内容替换，其他部分是原图的噪声图，所以在下一个迭代的时候会mask部分的生成也会condition on 其他部分的原图噪声 + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(x0_partial) + if callback: callback(i) + if img_callback: img_callback(img, i) + return img, intermediates + + @torch.no_grad() + def p_sample_loop(self, cond, shape, return_intermediates=False, + x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, start_T=None, + log_every_t=None): + + if not log_every_t: + log_every_t = self.log_every_t + device = self.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + intermediates = [img] + if timesteps is None: + timesteps = self.num_timesteps + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed( + range(0, timesteps)) + + if mask is not None: + assert x0 is not None + assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match + + for i in iterator: + ts = torch.full((b,), i, device=device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised) + if mask is not None: + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(img) + if callback: callback(i) + if img_callback: img_callback(img, i) + + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None, + verbose=True, timesteps=None, quantize_denoised=False, + mask=None, x0=None, shape=None, **kwargs): + if shape is None: + shape = (batch_size, self.channels, self.image_size, self.image_size) + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + return self.p_sample_loop(cond, + shape, + return_intermediates=return_intermediates, x_T=x_T, + verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised, + mask=mask, x0=x0) + + @torch.no_grad() + def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): + if ddim: + ddim_sampler = DDIMSampler(self) + shape = (self.channels, self.image_size, self.image_size) + samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, + shape, cond, verbose=False, **kwargs) + + else: + samples, intermediates = self.sample(cond=cond, batch_size=batch_size, + return_intermediates=True, **kwargs) + + return samples, intermediates + + @torch.no_grad() + def get_unconditional_conditioning(self, batch_size, null_label=None): + if null_label is not None: + xc = null_label + if isinstance(xc, ListConfig): + xc = list(xc) + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + if hasattr(xc, "to"): + xc = xc.to(self.device) + c = self.get_learned_conditioning(xc) + else: + if self.cond_stage_key in ["class_label", "cls"]: + xc = self.cond_stage_model.get_unconditional_conditioning(batch_size, device=self.device) + return self.get_learned_conditioning(xc) + else: + raise NotImplementedError("todo") + if isinstance(c, list): # in case the encoder gives us a list + for i in range(len(c)): + c[i] = repeat(c[i], '1 ... -> b ...', b=batch_size).to(self.device) + else: + c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device) + return c + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=50, ddim_eta=0., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, + return_first_stage_outputs=True, + force_c_encode=True, + return_original_cond=True, + bs=N) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', "cls"]: + try: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + except KeyError: + # probably no "human_label" in batch + pass + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance( + self.first_stage_model, IdentityFirstStage): + # also display when quantizing x0 while sampling + with ema_scope("Plotting Quantized Denoised"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + quantize_denoised=True) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True, + # quantize_denoised=True) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_x0_quantized"] = x_samples + + if unconditional_guidance_scale > 1.0: + uc = self.get_unconditional_conditioning(N, unconditional_guidance_label) + if self.model.conditioning_key == "crossattn-adm": + uc = {"c_crossattn": [uc], "c_adm": c["c_adm"]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if inpaint: + # make a simple center square + b, h, w = z.shape[0], z.shape[2], z.shape[3] + mask = torch.ones(N, h, w).to(self.device) + # zeros will be filled in + mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0. + mask = mask[:, None, ...] + with ema_scope("Plotting Inpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_inpainting"] = x_samples + log["mask"] = mask + + # outpaint + mask = 1. - mask + with ema_scope("Plotting Outpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_outpainting"] = x_samples + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + # params = [] + # for param in self.model.named_parameters(): + # if 'hypernetwork' in param[0]: + # params.append(param[1]) + if self.frozen_unet: + params = [] + for param in self.model.named_parameters(): + if 'transformer_blocks' and 'attn2' in param[0]: + if 'to_k' or 'to_v' in param[0]: + params.append(param[1]) + else: + params = list(self.model.parameters()) + if self.cond_stage_trainable: + print(f"{self.__class__.__name__}: Also optimizing conditioner params!") + # conda_stage_model_params = [] + # for param in self.cond_stage_model.named_parameters(): + # if 'transformation' in param[0]: + # conda_stage_model_params.append(param[1]) + # params = params + conda_stage_model_params + params = params + list(self.cond_stage_model.parameters()) + if self.learn_logvar: + print('Diffusion model optimizing logvar') + params.append(self.logvar) + opt = torch.optim.AdamW(params, lr=lr) + if self.use_scheduler: + assert 'target' in self.scheduler_config + scheduler = instantiate_from_config(self.scheduler_config) + + print("Setting up LambdaLR scheduler...") + scheduler = [ + { + 'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule), + 'interval': 'step', + 'frequency': 1 + }] + return [opt], scheduler + return opt + + @torch.no_grad() + def to_rgb(self, x): + x = x.float() + if not hasattr(self, "colorize"): + self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x) + x = nn.functional.conv2d(x, weight=self.colorize) + x = 2. * (x - x.min()) / (x.max() - x.min()) - 1. + return x + + +class DiffusionWrapper(pl.LightningModule): + def __init__(self, diff_model_config, conditioning_key, frozen_unet=False): + super().__init__() + self.sequential_cross_attn = diff_model_config.pop("sequential_crossattn", False) + self.diffusion_model = instantiate_from_config(diff_model_config) + # for param in self.diffusion_model.named_parameters(): + # if 'hypernetwork' in param[0]: + # param[1].requires_grad = True + # else: + # param[1].requires_grad = False + if frozen_unet: + for param in self.diffusion_model.named_parameters(): + if 'transformer_blocks' in param[0] and 'attn2' in param[0]: + if 'to_k' or 'to_v' in param[0]: + param[1].requires_grad = True + else: + param[1].requires_grad = False + else: + param[1].requires_grad = False + self.conditioning_key = conditioning_key + assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm', 'hybrid-adm', 'crossattn-adm'] + + def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None): + if self.conditioning_key is None: + out = self.diffusion_model(x, t) + elif self.conditioning_key == 'concat': + xc = torch.cat([x] + c_concat, dim=1) + out = self.diffusion_model(xc, t) + elif self.conditioning_key == 'crossattn': + if not self.sequential_cross_attn: + cc = torch.cat(c_crossattn, 1) + else: + cc = c_crossattn + out = self.diffusion_model(x, t, context=cc) + elif self.conditioning_key == 'hybrid': + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc) + elif self.conditioning_key == 'hybrid-adm': + assert c_adm is not None + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, y=c_adm) + elif self.conditioning_key == 'crossattn-adm': + assert c_adm is not None + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(x, t, context=cc, y=c_adm) + elif self.conditioning_key == 'adm': + cc = c_crossattn[0] + out = self.diffusion_model(x, t, y=cc) + else: + raise NotImplementedError() + + return out + + +class LatentUpscaleDiffusion(LatentDiffusion): + def __init__(self, *args, low_scale_config, low_scale_key="LR", noise_level_key=None, **kwargs): + super().__init__(*args, **kwargs) + # assumes that neither the cond_stage nor the low_scale_model contain trainable params + assert not self.cond_stage_trainable + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + self.noise_level_key = noise_level_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False): + if not log_mode: + z, c = super().get_input(batch, k, force_c_encode=True, bs=bs) + else: + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + x_low = batch[self.low_scale_key][:bs] + x_low = rearrange(x_low, 'b h w c -> b c h w') + x_low = x_low.to(memory_format=torch.contiguous_format).float() + zx, noise_level = self.low_scale_model(x_low) + if self.noise_level_key is not None: + # get noise level from batch instead, e.g. when extracting a custom noise level for bsr + raise NotImplementedError('TODO') + + all_conds = {"c_concat": [zx], "c_crossattn": [c], "c_adm": noise_level} + if log_mode: + # TODO: maybe disable if too expensive + x_low_rec = self.low_scale_model.decode(zx) + return z, all_conds, x, xrec, xc, x_low, x_low_rec, noise_level + return z, all_conds + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True, + unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc, x_low, x_low_rec, noise_level = self.get_input(batch, self.first_stage_key, bs=N, + log_mode=True) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + log["x_lr"] = x_low + log[f"x_lr_rec_@noise_levels{'-'.join(map(lambda x: str(x), list(noise_level.cpu().numpy())))}"] = x_low_rec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label) + # TODO explore better "unconditional" choices for the other keys + # maybe guide away from empty text label and highest noise level and maximally degraded zx? + uc = dict() + for k in c: + if k == "c_crossattn": + assert isinstance(c[k], list) and len(c[k]) == 1 + uc[k] = [uc_tmp] + elif k == "c_adm": # todo: only run with text-based guidance? + assert isinstance(c[k], torch.Tensor) + #uc[k] = torch.ones_like(c[k]) * self.low_scale_model.max_noise_level + uc[k] = c[k] + elif isinstance(c[k], list): + uc[k] = [c[k][i] for i in range(len(c[k]))] + else: + uc[k] = c[k] + + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + return log + + +class LatentFinetuneDiffusion(LatentDiffusion): + """ + Basis for different finetunas, such as inpainting or depth2image + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys: tuple, + finetune_keys=("model.diffusion_model.input_blocks.0.0.weight", + "model_ema.diffusion_modelinput_blocks00weight" + ), + keep_finetune_dims=4, + # if model was trained without concat mode before and we would like to keep these channels + c_concat_log_start=None, # to log reconstruction of c_concat codes + c_concat_log_end=None, + *args, **kwargs + ): + ckpt_path = kwargs.pop("ckpt_path", None) + ignore_keys = kwargs.pop("ignore_keys", list()) + super().__init__(*args, **kwargs) + self.finetune_keys = finetune_keys + self.concat_keys = concat_keys + self.keep_dims = keep_finetune_dims + self.c_concat_log_start = c_concat_log_start + self.c_concat_log_end = c_concat_log_end + if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint' + if exists(ckpt_path): + self.init_from_ckpt(ckpt_path, ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + + # make it explicit, finetune by including extra input channels + if exists(self.finetune_keys) and k in self.finetune_keys: + new_entry = None + for name, param in self.named_parameters(): + if name in self.finetune_keys: + print( + f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only") + new_entry = torch.zeros_like(param) # zero init + assert exists(new_entry), 'did not find matching parameter to modify' + new_entry[:, :self.keep_dims, ...] = sd[k] + sd[k] = new_entry + + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys: {missing}") + if len(unexpected) > 0: + print(f"Unexpected Keys: {unexpected}") + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True) + c_cat, c = c["c_concat"][0], c["c_crossattn"][0] + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if not (self.c_concat_log_start is None and self.c_concat_log_end is None): + log["c_concat_decoded"] = self.decode_first_stage(c_cat[:, self.c_concat_log_start:self.c_concat_log_end]) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label) + uc_cat = c_cat + uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc_full, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + return log + + +class LatentInpaintDiffusion(LatentFinetuneDiffusion): + """ + can either run as pure inpainting model (only concat mode) or with mixed conditionings, + e.g. mask as concat and text via cross-attn. + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys=("mask", "masked_image"), + masked_image_key="masked_image", + *args, **kwargs + ): + super().__init__(concat_keys, *args, **kwargs) + self.masked_image_key = masked_image_key + assert self.masked_image_key in concat_keys + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + c_cat = list() + for ck in self.concat_keys: + cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + bchw = z.shape + if ck != self.masked_image_key: + cc = torch.nn.functional.interpolate(cc, size=bchw[-2:]) + else: + cc = self.get_first_stage_encoding(self.encode_first_stage(cc)) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super(LatentInpaintDiffusion, self).log_images(*args, **kwargs) + log["masked_image"] = rearrange(args[0]["masked_image"], + 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + return log + + +class LatentDepth2ImageDiffusion(LatentFinetuneDiffusion): + """ + condition on monocular depth estimation + """ + + def __init__(self, depth_stage_config, concat_keys=("midas_in",), *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.depth_model = instantiate_from_config(depth_stage_config) + self.depth_stage_key = concat_keys[0] + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for depth2img' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + c_cat = list() + for ck in self.concat_keys: + cc = batch[ck] + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + cc = self.depth_model(cc) + cc = torch.nn.functional.interpolate( + cc, + size=z.shape[2:], + mode="bicubic", + align_corners=False, + ) + + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + cc = 2. * (cc - depth_min) / (depth_max - depth_min + 0.001) - 1. + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + depth = self.depth_model(args[0][self.depth_stage_key]) + depth_min, depth_max = torch.amin(depth, dim=[1, 2, 3], keepdim=True), \ + torch.amax(depth, dim=[1, 2, 3], keepdim=True) + log["depth"] = 2. * (depth - depth_min) / (depth_max - depth_min) - 1. + return log + + +class LatentUpscaleFinetuneDiffusion(LatentFinetuneDiffusion): + """ + condition on low-res image (and optionally on some spatial noise augmentation) + """ + def __init__(self, concat_keys=("lr",), reshuffle_patch_size=None, + low_scale_config=None, low_scale_key=None, *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.reshuffle_patch_size = reshuffle_patch_size + self.low_scale_model = None + if low_scale_config is not None: + print("Initializing a low-scale model") + assert exists(low_scale_key) + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for upscaling-ft' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + # optionally make spatial noise_level here + c_cat = list() + noise_level = None + for ck in self.concat_keys: + cc = batch[ck] + cc = rearrange(cc, 'b h w c -> b c h w') + if exists(self.reshuffle_patch_size): + assert isinstance(self.reshuffle_patch_size, int) + cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w', + p1=self.reshuffle_patch_size, p2=self.reshuffle_patch_size) + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + if exists(self.low_scale_model) and ck == self.low_scale_key: + cc, noise_level = self.low_scale_model(cc) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + if exists(noise_level): + all_conds = {"c_concat": [c_cat], "c_crossattn": [c], "c_adm": noise_level} + else: + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + log["lr"] = rearrange(args[0]["lr"], 'b h w c -> b c h w') + return log \ No newline at end of file diff --git a/flagai/model/mm/Sampler.py b/flagai/model/mm/Sampler.py index 6af3a099..063ee708 100644 --- a/flagai/model/mm/Sampler.py +++ b/flagai/model/mm/Sampler.py @@ -465,3 +465,5 @@ def get_x_prev_and_pred_x0(e_t, index): x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) return x_prev, pred_x0, e_t + + diff --git a/flagai/model/mm/autoencoder2.py b/flagai/model/mm/autoencoder2.py new file mode 100644 index 00000000..90bf7a94 --- /dev/null +++ b/flagai/model/mm/autoencoder2.py @@ -0,0 +1,219 @@ +import torch +import pytorch_lightning as pl +import torch.nn.functional as F +from contextlib import contextmanager + +from flagai.model.mm.modules.diffusionmodules.model import Encoder, Decoder +from flagai.model.mm.modules.distributions.distributions import DiagonalGaussianDistribution + +from flagai.model.mm.utils import instantiate_from_config +from flagai.model.mm.modules.ema import LitEma + + +class AutoencoderKL(pl.LightningModule): + def __init__(self, + ddconfig, + lossconfig, + embed_dim, + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None, + monitor=None, + ema_decay=None, + learn_logvar=False + ): + super().__init__() + self.learn_logvar = learn_logvar + self.image_key = image_key + self.encoder = Encoder(**ddconfig) + self.decoder = Decoder(**ddconfig) + self.loss = instantiate_from_config(lossconfig) + assert ddconfig["double_z"] + self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1) + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) + self.embed_dim = embed_dim + if colorize_nlabels is not None: + assert type(colorize_nlabels)==int + self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) + if monitor is not None: + self.monitor = monitor + + self.use_ema = ema_decay is not None + if self.use_ema: + self.ema_decay = ema_decay + assert 0. < ema_decay < 1. + self.model_ema = LitEma(self, decay=ema_decay) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list()): + sd = torch.load(path, map_location="cpu")["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + self.load_state_dict(sd, strict=False) + print(f"Restored from {path}") + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.parameters()) + self.model_ema.copy_to(self) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self) + + def encode(self, x): + h = self.encoder(x) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + return posterior + + def decode(self, z): + z = self.post_quant_conv(z) + dec = self.decoder(z) + return dec + + def forward(self, input, sample_posterior=True): + posterior = self.encode(input) + if sample_posterior: + z = posterior.sample() + else: + z = posterior.mode() + dec = self.decode(z) + return dec, posterior + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float() + return x + + def training_step(self, batch, batch_idx, optimizer_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + + if optimizer_idx == 0: + # train encoder+decoder+logvar + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return aeloss + + if optimizer_idx == 1: + # train the discriminator + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + + self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return discloss + + def validation_step(self, batch, batch_idx): + log_dict = self._validation_step(batch, batch_idx) + with self.ema_scope(): + log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema") + return log_dict + + def _validation_step(self, batch, batch_idx, postfix=""): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"]) + self.log_dict(log_dict_ae) + self.log_dict(log_dict_disc) + return self.log_dict + + def configure_optimizers(self): + lr = self.learning_rate + ae_params_list = list(self.encoder.parameters()) + list(self.decoder.parameters()) + list( + self.quant_conv.parameters()) + list(self.post_quant_conv.parameters()) + if self.learn_logvar: + print(f"{self.__class__.__name__}: Learning logvar") + ae_params_list.append(self.loss.logvar) + opt_ae = torch.optim.Adam(ae_params_list, + lr=lr, betas=(0.5, 0.9)) + opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), + lr=lr, betas=(0.5, 0.9)) + return [opt_ae, opt_disc], [] + + def get_last_layer(self): + return self.decoder.conv_out.weight + + @torch.no_grad() + def log_images(self, batch, only_inputs=False, log_ema=False, **kwargs): + log = dict() + x = self.get_input(batch, self.image_key) + x = x.to(self.device) + if not only_inputs: + xrec, posterior = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec.shape[1] > 3 + x = self.to_rgb(x) + xrec = self.to_rgb(xrec) + log["samples"] = self.decode(torch.randn_like(posterior.sample())) + log["reconstructions"] = xrec + if log_ema or self.use_ema: + with self.ema_scope(): + xrec_ema, posterior_ema = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec_ema.shape[1] > 3 + xrec_ema = self.to_rgb(xrec_ema) + log["samples_ema"] = self.decode(torch.randn_like(posterior_ema.sample())) + log["reconstructions_ema"] = xrec_ema + log["inputs"] = x + return log + + def to_rgb(self, x): + assert self.image_key == "segmentation" + if not hasattr(self, "colorize"): + self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) + x = F.conv2d(x, weight=self.colorize) + x = 2.*(x-x.min())/(x.max()-x.min()) - 1. + return x + + +class IdentityFirstStage(torch.nn.Module): + def __init__(self, *args, vq_interface=False, **kwargs): + self.vq_interface = vq_interface + super().__init__() + + def encode(self, x, *args, **kwargs): + return x + + def decode(self, x, *args, **kwargs): + return x + + def quantize(self, x, *args, **kwargs): + if self.vq_interface: + return x, None, [None, None, None] + return x + + def forward(self, x, *args, **kwargs): + return x + diff --git a/flagai/model/mm/dpm_solver/__init__.py b/flagai/model/mm/dpm_solver/__init__.py new file mode 100644 index 00000000..7427f38c --- /dev/null +++ b/flagai/model/mm/dpm_solver/__init__.py @@ -0,0 +1 @@ +from .sampler import DPMSolverSampler \ No newline at end of file diff --git a/flagai/model/mm/dpm_solver/dpm_solver.py b/flagai/model/mm/dpm_solver/dpm_solver.py new file mode 100644 index 00000000..095e5ba3 --- /dev/null +++ b/flagai/model/mm/dpm_solver/dpm_solver.py @@ -0,0 +1,1154 @@ +import torch +import torch.nn.functional as F +import math +from tqdm import tqdm + + +class NoiseScheduleVP: + def __init__( + self, + schedule='discrete', + betas=None, + alphas_cumprod=None, + continuous_beta_0=0.1, + continuous_beta_1=20., + ): + """Create a wrapper class for the forward SDE (VP type). + *** + Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t. + We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images. + *** + The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ). + We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper). + Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have: + log_alpha_t = self.marginal_log_mean_coeff(t) + sigma_t = self.marginal_std(t) + lambda_t = self.marginal_lambda(t) + Moreover, as lambda(t) is an invertible function, we also support its inverse function: + t = self.inverse_lambda(lambda_t) + =============================================================== + We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]). + 1. For discrete-time DPMs: + For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by: + t_i = (i + 1) / N + e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1. + We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3. + Args: + betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details) + alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details) + Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`. + **Important**: Please pay special attention for the args for `alphas_cumprod`: + The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that + q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ). + Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have + alpha_{t_n} = \sqrt{\hat{alpha_n}}, + and + log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}). + 2. For continuous-time DPMs: + We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise + schedule are the default settings in DDPM and improved-DDPM: + Args: + beta_min: A `float` number. The smallest beta for the linear schedule. + beta_max: A `float` number. The largest beta for the linear schedule. + cosine_s: A `float` number. The hyperparameter in the cosine schedule. + cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule. + T: A `float` number. The ending time of the forward process. + =============================================================== + Args: + schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs, + 'linear' or 'cosine' for continuous-time DPMs. + Returns: + A wrapper object of the forward SDE (VP type). + + =============================================================== + Example: + # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', betas=betas) + # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod) + # For continuous-time DPMs (VPSDE), linear schedule: + >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.) + """ + + if schedule not in ['discrete', 'linear', 'cosine']: + raise ValueError( + "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format( + schedule)) + + self.schedule = schedule + if schedule == 'discrete': + if betas is not None: + log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0) + else: + assert alphas_cumprod is not None + log_alphas = 0.5 * torch.log(alphas_cumprod) + self.total_N = len(log_alphas) + self.T = 1. + self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1)) + self.log_alpha_array = log_alphas.reshape((1, -1,)) + else: + self.total_N = 1000 + self.beta_0 = continuous_beta_0 + self.beta_1 = continuous_beta_1 + self.cosine_s = 0.008 + self.cosine_beta_max = 999. + self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.)) + self.schedule = schedule + if schedule == 'cosine': + # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T. + # Note that T = 0.9946 may be not the optimal setting. However, we find it works well. + self.T = 0.9946 + else: + self.T = 1. + + def marginal_log_mean_coeff(self, t): + """ + Compute log(alpha_t) of a given continuous-time label t in [0, T]. + """ + if self.schedule == 'discrete': + return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device), + self.log_alpha_array.to(t.device)).reshape((-1)) + elif self.schedule == 'linear': + return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0 + elif self.schedule == 'cosine': + log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.)) + log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0 + return log_alpha_t + + def marginal_alpha(self, t): + """ + Compute alpha_t of a given continuous-time label t in [0, T]. + """ + return torch.exp(self.marginal_log_mean_coeff(t)) + + def marginal_std(self, t): + """ + Compute sigma_t of a given continuous-time label t in [0, T]. + """ + return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t))) + + def marginal_lambda(self, t): + """ + Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T]. + """ + log_mean_coeff = self.marginal_log_mean_coeff(t) + log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff)) + return log_mean_coeff - log_std + + def inverse_lambda(self, lamb): + """ + Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t. + """ + if self.schedule == 'linear': + tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + Delta = self.beta_0 ** 2 + tmp + return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0) + elif self.schedule == 'discrete': + log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb) + t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]), + torch.flip(self.t_array.to(lamb.device), [1])) + return t.reshape((-1,)) + else: + log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + t = t_fn(log_alpha) + return t + + +def model_wrapper( + model, + noise_schedule, + model_type="noise", + model_kwargs={}, + guidance_type="uncond", + condition=None, + unconditional_condition=None, + guidance_scale=1., + classifier_fn=None, + classifier_kwargs={}, +): + """Create a wrapper function for the noise prediction model. + DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to + firstly wrap the model function to a noise prediction model that accepts the continuous time as the input. + We support four types of the diffusion model by setting `model_type`: + 1. "noise": noise prediction model. (Trained by predicting noise). + 2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0). + 3. "v": velocity prediction model. (Trained by predicting the velocity). + The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2]. + [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models." + arXiv preprint arXiv:2202.00512 (2022). + [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models." + arXiv preprint arXiv:2210.02303 (2022). + + 4. "score": marginal score function. (Trained by denoising score matching). + Note that the score function and the noise prediction model follows a simple relationship: + ``` + noise(x_t, t) = -sigma_t * score(x_t, t) + ``` + We support three types of guided sampling by DPMs by setting `guidance_type`: + 1. "uncond": unconditional sampling by DPMs. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + 2. "classifier": classifier guidance sampling [3] by DPMs and another classifier. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + The input `classifier_fn` has the following format: + `` + classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond) + `` + [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis," + in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794. + 3. "classifier-free": classifier-free guidance sampling by conditional DPMs. + The input `model` has the following format: + `` + model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score + `` + And if cond == `unconditional_condition`, the model output is the unconditional DPM output. + [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance." + arXiv preprint arXiv:2207.12598 (2022). + + The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999) + or continuous-time labels (i.e. epsilon to T). + We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise: + `` + def model_fn(x, t_continuous) -> noise: + t_input = get_model_input_time(t_continuous) + return noise_pred(model, x, t_input, **model_kwargs) + `` + where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver. + =============================================================== + Args: + model: A diffusion model with the corresponding format described above. + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + model_type: A `str`. The parameterization type of the diffusion model. + "noise" or "x_start" or "v" or "score". + model_kwargs: A `dict`. A dict for the other inputs of the model function. + guidance_type: A `str`. The type of the guidance for sampling. + "uncond" or "classifier" or "classifier-free". + condition: A pytorch tensor. The condition for the guided sampling. + Only used for "classifier" or "classifier-free" guidance type. + unconditional_condition: A pytorch tensor. The condition for the unconditional sampling. + Only used for "classifier-free" guidance type. + guidance_scale: A `float`. The scale for the guided sampling. + classifier_fn: A classifier function. Only used for the classifier guidance. + classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function. + Returns: + A noise prediction model that accepts the noised data and the continuous time as the inputs. + """ + + def get_model_input_time(t_continuous): + """ + Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time. + For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N]. + For continuous-time DPMs, we just use `t_continuous`. + """ + if noise_schedule.schedule == 'discrete': + return (t_continuous - 1. / noise_schedule.total_N) * 1000. + else: + return t_continuous + + def noise_pred_fn(x, t_continuous, cond=None): + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + t_input = get_model_input_time(t_continuous) + if cond is None: + output = model(x, t_input, **model_kwargs) + else: + output = model(x, t_input, cond, **model_kwargs) + if model_type == "noise": + return output + elif model_type == "x_start": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims) + elif model_type == "v": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x + elif model_type == "score": + sigma_t = noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return -expand_dims(sigma_t, dims) * output + + def cond_grad_fn(x, t_input): + """ + Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t). + """ + with torch.enable_grad(): + x_in = x.detach().requires_grad_(True) + log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs) + return torch.autograd.grad(log_prob.sum(), x_in)[0] + + def model_fn(x, t_continuous): + """ + The noise predicition model function that is used for DPM-Solver. + """ + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + if guidance_type == "uncond": + return noise_pred_fn(x, t_continuous) + elif guidance_type == "classifier": + assert classifier_fn is not None + t_input = get_model_input_time(t_continuous) + cond_grad = cond_grad_fn(x, t_input) + sigma_t = noise_schedule.marginal_std(t_continuous) + noise = noise_pred_fn(x, t_continuous) + return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad + elif guidance_type == "classifier-free": + if guidance_scale == 1. or unconditional_condition is None: + return noise_pred_fn(x, t_continuous, cond=condition) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t_continuous] * 2) + c_in = torch.cat([unconditional_condition, condition]) + noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2) + return noise_uncond + guidance_scale * (noise - noise_uncond) + + assert model_type in ["noise", "x_start", "v"] + assert guidance_type in ["uncond", "classifier", "classifier-free"] + return model_fn + + +class DPM_Solver: + def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.): + """Construct a DPM-Solver. + We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0"). + If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver). + If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++). + In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True. + The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales. + Args: + model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]): + `` + def model_fn(x, t_continuous): + return noise + `` + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model. + thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1]. + max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding. + + [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b. + """ + self.model = model_fn + self.noise_schedule = noise_schedule + self.predict_x0 = predict_x0 + self.thresholding = thresholding + self.max_val = max_val + + def noise_prediction_fn(self, x, t): + """ + Return the noise prediction model. + """ + return self.model(x, t) + + def data_prediction_fn(self, x, t): + """ + Return the data prediction model (with thresholding). + """ + noise = self.noise_prediction_fn(x, t) + dims = x.dim() + alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t) + x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims) + if self.thresholding: + p = 0.995 # A hyperparameter in the paper of "Imagen" [1]. + s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1) + s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims) + x0 = torch.clamp(x0, -s, s) / s + return x0 + + def model_fn(self, x, t): + """ + Convert the model to the noise prediction model or the data prediction model. + """ + if self.predict_x0: + return self.data_prediction_fn(x, t) + else: + return self.noise_prediction_fn(x, t) + + def get_time_steps(self, skip_type, t_T, t_0, N, device): + """Compute the intermediate time steps for sampling. + Args: + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + N: A `int`. The total number of the spacing of the time steps. + device: A torch device. + Returns: + A pytorch tensor of the time steps, with the shape (N + 1,). + """ + if skip_type == 'logSNR': + lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device)) + lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device)) + logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device) + return self.noise_schedule.inverse_lambda(logSNR_steps) + elif skip_type == 'time_uniform': + return torch.linspace(t_T, t_0, N + 1).to(device) + elif skip_type == 'time_quadratic': + t_order = 2 + t = torch.linspace(t_T ** (1. / t_order), t_0 ** (1. / t_order), N + 1).pow(t_order).to(device) + return t + else: + raise ValueError( + "Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type)) + + def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device): + """ + Get the order of each step for sampling by the singlestep DPM-Solver. + We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast". + Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is: + - If order == 1: + We take `steps` of DPM-Solver-1 (i.e. DDIM). + - If order == 2: + - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of DPM-Solver-2. + - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If order == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2. + ============================================ + Args: + order: A `int`. The max order for the solver (2 or 3). + steps: A `int`. The total number of function evaluations (NFE). + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + device: A torch device. + Returns: + orders: A list of the solver order of each step. + """ + if order == 3: + K = steps // 3 + 1 + if steps % 3 == 0: + orders = [3, ] * (K - 2) + [2, 1] + elif steps % 3 == 1: + orders = [3, ] * (K - 1) + [1] + else: + orders = [3, ] * (K - 1) + [2] + elif order == 2: + if steps % 2 == 0: + K = steps // 2 + orders = [2, ] * K + else: + K = steps // 2 + 1 + orders = [2, ] * (K - 1) + [1] + elif order == 1: + K = 1 + orders = [1, ] * steps + else: + raise ValueError("'order' must be '1' or '2' or '3'.") + if skip_type == 'logSNR': + # To reproduce the results in DPM-Solver paper + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device) + else: + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[ + torch.cumsum(torch.tensor([0, ] + orders)).to(device)] + return timesteps_outer, orders + + def denoise_to_zero_fn(self, x, s): + """ + Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization. + """ + return self.data_prediction_fn(x, s) + + def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False): + """ + DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + if self.predict_x0: + phi_1 = torch.expm1(-h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + else: + phi_1 = torch.expm1(h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + + def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, + solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-2 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the second-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 0.5 + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + s1 = ns.inverse_lambda(lambda_s1) + log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff( + s1), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t) + alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_1 = torch.expm1(-h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * ( + model_s1 - model_s) + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_1 = torch.expm1(h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s) + ) + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1} + else: + return x_t + + def singlestep_dpm_solver_third_update(self, x, s, t, r1=1. / 3., r2=2. / 3., model_s=None, model_s1=None, + return_intermediate=False, solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-3 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`). + If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 1. / 3. + if r2 is None: + r2 = 2. / 3. + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + lambda_s2 = lambda_s + r2 * h + s1 = ns.inverse_lambda(lambda_s1) + s2 = ns.inverse_lambda(lambda_s2) + log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff( + s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std( + s2), ns.marginal_std(t) + alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_12 = torch.expm1(-r2 * h) + phi_1 = torch.expm1(-h) + phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1. + phi_2 = phi_1 / h + 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(sigma_s2 / sigma_s, dims) * x + - expand_dims(alpha_s2 * phi_12, dims) * model_s + + r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + expand_dims(alpha_t * phi_2, dims) * D1 + - expand_dims(alpha_t * phi_3, dims) * D2 + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_12 = torch.expm1(r2 * h) + phi_1 = torch.expm1(h) + phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1. + phi_2 = phi_1 / h - 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x + - expand_dims(sigma_s2 * phi_12, dims) * model_s + - r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - expand_dims(sigma_t * phi_2, dims) * D1 + - expand_dims(sigma_t * phi_3, dims) * D2 + ) + + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2} + else: + return x_t + + def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"): + """ + Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + ns = self.noise_schedule + dims = x.dim() + model_prev_1, model_prev_0 = model_prev_list + t_prev_1, t_prev_0 = t_prev_list + lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda( + t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0 = h_0 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + if self.predict_x0: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0 + ) + else: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0 + ) + return x_t + + def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'): + """ + Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + model_prev_2, model_prev_1, model_prev_0 = model_prev_list + t_prev_2, t_prev_1, t_prev_0 = t_prev_list + lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda( + t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_1 = lambda_prev_1 - lambda_prev_2 + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0, r1 = h_0 / h, h_1 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2) + D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1) + D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1) + if self.predict_x0: + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1 + - expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h ** 2 - 0.5), dims) * D2 + ) + else: + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1 + - expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h ** 2 - 0.5), dims) * D2 + ) + return x_t + + def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None, + r2=None): + """ + Singlestep DPM-Solver with the order `order` from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + r1: A `float`. The hyperparameter of the second-order or third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate) + elif order == 2: + return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1) + elif order == 3: + return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1, r2=r2) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'): + """ + Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1]) + elif order == 2: + return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + elif order == 3: + return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, + solver_type='dpm_solver'): + """ + The adaptive step size solver based on singlestep DPM-Solver. + Args: + x: A pytorch tensor. The initial value at time `t_T`. + order: A `int`. The (higher) order of the solver. We only support order == 2 or 3. + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + h_init: A `float`. The initial step size (for logSNR). + atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1]. + rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05. + theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1]. + t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the + current time and `t_0` is less than `t_err`. The default setting is 1e-5. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_0: A pytorch tensor. The approximated solution at time `t_0`. + [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021. + """ + ns = self.noise_schedule + s = t_T * torch.ones((x.shape[0],)).to(x) + lambda_s = ns.marginal_lambda(s) + lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x)) + h = h_init * torch.ones_like(s).to(x) + x_prev = x + nfe = 0 + if order == 2: + r1 = 0.5 + lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + solver_type=solver_type, + **kwargs) + elif order == 3: + r1, r2 = 1. / 3., 2. / 3. + lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + return_intermediate=True, + solver_type=solver_type) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2, + solver_type=solver_type, + **kwargs) + else: + raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order)) + while torch.abs((s - t_0)).mean() > t_err: + t = ns.inverse_lambda(lambda_s + h) + x_lower, lower_noise_kwargs = lower_update(x, s, t) + x_higher = higher_update(x, s, t, **lower_noise_kwargs) + delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev))) + norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)) + E = norm_fn((x_higher - x_lower) / delta).max() + if torch.all(E <= 1.): + x = x_higher + s = t + x_prev = x_lower + lambda_s = ns.marginal_lambda(s) + h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s) + nfe += order + print('adaptive solver nfe', nfe) + return x + + def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform', + method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver', + atol=0.0078, rtol=0.05, + ): + """ + Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`. + ===================================================== + We support the following algorithms for both noise prediction model and data prediction model: + - 'singlestep': + Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver. + We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps). + The total number of function evaluations (NFE) == `steps`. + Given a fixed NFE == `steps`, the sampling procedure is: + - If `order` == 1: + - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2. + - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If `order` == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2. + - 'multistep': + Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`. + We initialize the first `order` values by lower order multistep solvers. + Given a fixed NFE == `steps`, the sampling procedure is: + Denote K = steps. + - If `order` == 1: + - We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2. + - If `order` == 3: + - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3. + - 'singlestep_fixed': + Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3). + We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE. + - 'adaptive': + Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper). + We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`. + You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs + (NFE) and the sample quality. + - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2. + - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3. + ===================================================== + Some advices for choosing the algorithm: + - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs: + Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3, + skip_type='time_uniform', method='singlestep') + - For **guided sampling with large guidance scale** by DPMs: + Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2, + skip_type='time_uniform', method='multistep') + We support three types of `skip_type`: + - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images** + - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**. + - 'time_quadratic': quadratic time for the time steps. + ===================================================== + Args: + x: A pytorch tensor. The initial value at time `t_start` + e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution. + steps: A `int`. The total number of function evaluations (NFE). + t_start: A `float`. The starting time of the sampling. + If `T` is None, we use self.noise_schedule.T (default is 1.0). + t_end: A `float`. The ending time of the sampling. + If `t_end` is None, we use 1. / self.noise_schedule.total_N. + e.g. if total_N == 1000, we have `t_end` == 1e-3. + For discrete-time DPMs: + - We recommend `t_end` == 1. / self.noise_schedule.total_N. + For continuous-time DPMs: + - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15. + order: A `int`. The order of DPM-Solver. + skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'. + method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'. + denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step. + Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1). + This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and + score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID + for diffusion models sampling by diffusion SDEs for low-resolutional images + (such as CIFAR-10). However, we observed that such trick does not matter for + high-resolutional images. As it needs an additional NFE, we do not recommend + it for high-resolutional images. + lower_order_final: A `bool`. Whether to use lower order solvers at the final steps. + Only valid for `method=multistep` and `steps < 15`. We empirically find that + this trick is a key to stabilizing the sampling by DPM-Solver with very few steps + (especially for steps <= 10). So we recommend to set it to be `True`. + solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`. + atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + Returns: + x_end: A pytorch tensor. The approximated solution at time `t_end`. + """ + t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end + t_T = self.noise_schedule.T if t_start is None else t_start + device = x.device + if method == 'adaptive': + with torch.no_grad(): + x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, + solver_type=solver_type) + elif method == 'multistep': + assert steps >= order + timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device) + assert timesteps.shape[0] - 1 == steps + with torch.no_grad(): + vec_t = timesteps[0].expand((x.shape[0])) + model_prev_list = [self.model_fn(x, vec_t)] + t_prev_list = [vec_t] + # Init the first `order` values by lower order multistep DPM-Solver. + for init_order in tqdm(range(1, order), desc="DPM init order"): + vec_t = timesteps[init_order].expand(x.shape[0]) + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order, + solver_type=solver_type) + model_prev_list.append(self.model_fn(x, vec_t)) + t_prev_list.append(vec_t) + # Compute the remaining values by `order`-th order multistep DPM-Solver. + for step in tqdm(range(order, steps + 1), desc="DPM multistep"): + vec_t = timesteps[step].expand(x.shape[0]) + if lower_order_final and steps < 15: + step_order = min(order, steps + 1 - step) + else: + step_order = order + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order, + solver_type=solver_type) + for i in range(order - 1): + t_prev_list[i] = t_prev_list[i + 1] + model_prev_list[i] = model_prev_list[i + 1] + t_prev_list[-1] = vec_t + # We do not need to evaluate the final model value. + if step < steps: + model_prev_list[-1] = self.model_fn(x, vec_t) + elif method in ['singlestep', 'singlestep_fixed']: + if method == 'singlestep': + timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order, + skip_type=skip_type, + t_T=t_T, t_0=t_0, + device=device) + elif method == 'singlestep_fixed': + K = steps // order + orders = [order, ] * K + timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device) + for i, order in enumerate(orders): + t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1] + timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(), + N=order, device=device) + lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner) + vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0]) + h = lambda_inner[-1] - lambda_inner[0] + r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h + r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h + x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2) + if denoise_to_zero: + x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0) + return x + + +############################################################# +# other utility functions +############################################################# + +def interpolate_fn(x, xp, yp): + """ + A piecewise linear function y = f(x), using xp and yp as keypoints. + We implement f(x) in a differentiable way (i.e. applicable for autograd). + The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.) + Args: + x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver). + xp: PyTorch tensor with shape [C, K], where K is the number of keypoints. + yp: PyTorch tensor with shape [C, K]. + Returns: + The function values f(x), with shape [N, C]. + """ + N, K = x.shape[0], xp.shape[1] + all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2) + sorted_all_x, x_indices = torch.sort(all_x, dim=2) + x_idx = torch.argmin(x_indices, dim=2) + cand_start_idx = x_idx - 1 + start_idx = torch.where( + torch.eq(x_idx, 0), + torch.tensor(1, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1) + start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2) + end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2) + start_idx2 = torch.where( + torch.eq(x_idx, 0), + torch.tensor(0, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1) + start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2) + end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2) + cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x) + return cand + + +def expand_dims(v, dims): + """ + Expand the tensor `v` to the dim `dims`. + Args: + `v`: a PyTorch tensor with shape [N]. + `dim`: a `int`. + Returns: + a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`. + """ + return v[(...,) + (None,) * (dims - 1)] \ No newline at end of file diff --git a/flagai/model/mm/dpm_solver/sampler.py b/flagai/model/mm/dpm_solver/sampler.py new file mode 100644 index 00000000..7d137b8c --- /dev/null +++ b/flagai/model/mm/dpm_solver/sampler.py @@ -0,0 +1,87 @@ +"""SAMPLING ONLY.""" +import torch + +from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver + + +MODEL_TYPES = { + "eps": "noise", + "v": "v" +} + + +class DPMSolverSampler(object): + def __init__(self, model, **kwargs): + super().__init__() + self.model = model + to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device) + self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod)) + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + + print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}') + + device = self.model.betas.device + if x_T is None: + img = torch.randn(size, device=device) + else: + img = x_T + + ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod) + + model_fn = model_wrapper( + lambda x, t, c: self.model.apply_model(x, t, c), + ns, + model_type=MODEL_TYPES[self.model.parameterization], + guidance_type="classifier-free", + condition=conditioning, + unconditional_condition=unconditional_conditioning, + guidance_scale=unconditional_guidance_scale, + ) + + dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False) + x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, lower_order_final=True) + + return x.to(device), None \ No newline at end of file diff --git a/flagai/model/mm/lm/dcn_clip.py b/flagai/model/mm/lm/dcn_clip.py new file mode 100644 index 00000000..f91b2326 --- /dev/null +++ b/flagai/model/mm/lm/dcn_clip.py @@ -0,0 +1,61 @@ +import torch +import torch.nn as nn +from transformers import AltCLIPProcessor +from flagai.model.mm.AltCLIP import AltCLIP as AltCLIPModel + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + +class LayerNorm(nn.Module): + + def __init__(self, hidden_size, eps=1e-12): + """Construct a layernorm module in the TF style (epsilon inside the square root). + """ + super(LayerNorm, self).__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.bias = nn.Parameter(torch.zeros(hidden_size)) + self.variance_epsilon = eps + + def forward(self, x): + u = x.mean(-1, keepdim=True) + s = (x - u).pow(2).mean(-1, keepdim=True) + x = (x - u) / torch.sqrt(s + self.variance_epsilon) + return self.weight * x + self.bias + + +class DCNCLIP30M1024(AbstractEncoder): + def __init__(self, device="cuda", max_length=77, ckpt_path=None, download_path=None): + super().__init__() + self.device = device + self.max_length = max_length + + self.ch_clip_model = AltCLIPModel.from_pretrain(ckpt_path) + print("^_^ Using the right lm model!!!") + self.ch_clip_model = self.ch_clip_model.eval() + print("Language model Loaded!!!^_^") + + for param in self.ch_clip_model.parameters(): + param.requires_grad = False + + self.processor = AltCLIPProcessor.from_pretrain(ckpt_path) + self.tokenizer = self.processor.tokenizer + + self.text_encoder = self.ch_clip_model.text_model + + def forward(self, text): + + tokens = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=False,return_overflowing_tokens=False, padding="max_length", return_tensors="pt").to(self.text_encoder.device) + + z = self.text_encoder(**tokens)['penultimate_hidden_state'] + + return z + + def encode(self, text): + z = self(text) + if z.ndim==2: + z = z[:, None, :] + return z diff --git a/flagai/model/mm/modules/attention.py b/flagai/model/mm/modules/attention.py new file mode 100644 index 00000000..3dc97df0 --- /dev/null +++ b/flagai/model/mm/modules/attention.py @@ -0,0 +1,343 @@ +from inspect import isfunction +import math +import torch +import torch.nn.functional as F +from torch import nn, einsum +from einops import rearrange, repeat +from typing import Optional, Any + +from flagai.model.mm.modules.diffusionmodules.util import checkpoint + + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + + +def exists(val): + return val is not None + + +def uniq(arr): + return{el: True for el in arr}.keys() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + + +# feedforward +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def Normalize(in_channels): + return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = rearrange(q, 'b c h w -> b (h w) c') + k = rearrange(k, 'b c h w -> b c (h w)') + w_ = torch.einsum('bij,bjk->bik', q, k) + + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, 'b c h w -> b c (h w)') + w_ = rearrange(w_, 'b i j -> b j i') + h_ = torch.einsum('bij,bjk->bik', v, w_) + h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h) + h_ = self.proj_out(h_) + + return x+h_ + + +class CrossAttention(nn.Module): + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.scale = dim_head ** -0.5 + self.heads = heads + + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), + nn.Dropout(dropout) + ) + + def forward(self, x, context=None, mask=None): + h = self.heads + + # 注释写的是有context的情况，没有context的时候就是self attention + # q:(batch_size, 4096, 320) x: (batch_size, 4096, 320) + q = self.to_q(x) + context = default(context, x) + if type(context) is list: + context_k = context[0] + context_v = context[1] + else: + context_k = context + context_v = context + # k: (batch_size, 77, 320) v:(batch_size, 77, 320) + # contextL (batch_size, 77, 1024) + k = self.to_k(context_k) + v = self.to_v(context_v) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + del q, k + + if exists(mask): + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', sim, v) + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + return self.to_out(out) + + +class MemoryEfficientCrossAttention(nn.Module): + # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0): + super().__init__() + print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using " + f"{heads} heads.") + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.heads = heads + self.dim_head = dim_head + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)) + self.attention_op: Optional[Any] = None + + def forward(self, x, context=None, mask=None): + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + return self.to_out(out) + + +class BasicTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention + } + def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, + disable_self_attn=False): + super().__init__() + attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax" + assert attn_mode in self.ATTENTION_MODES + attn_cls = self.ATTENTION_MODES[attn_mode] + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None) # is a self-attention if not self.disable_self_attn + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, + heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + # self.checkpoint = False + + def forward(self, x, context=None): + return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint) + + def _forward(self, x, context=None): + x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x + x = self.attn2(self.norm2(x), context=context) + x + x = self.ff(self.norm3(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + def __init__(self, in_channels, n_heads, d_head, + depth=1, dropout=0., context_dim=None, + disable_self_attn=False, use_linear=False, + use_checkpoint=True): + super().__init__() + if exists(context_dim) and not isinstance(context_dim, list): + context_dim = [context_dim] + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = Normalize(in_channels) + if not use_linear: + self.proj_in = nn.Conv2d(in_channels, + inner_dim, + kernel_size=1, + stride=1, + padding=0) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + self.transformer_blocks = nn.ModuleList( + [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], + disable_self_attn=disable_self_attn, checkpoint=use_checkpoint) + for d in range(depth)] + ) + if not use_linear: + self.proj_out = zero_module(nn.Conv2d(inner_dim, + in_channels, + kernel_size=1, + stride=1, + padding=0)) + else: + self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) + self.use_linear = use_linear + + def forward(self, x, context=None): + # note: if no context is given, cross-attention defaults to self-attention + if not isinstance(context, list): + context = [context] + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'b c h w -> b (h w) c').contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + x = block(x, context=context[i]) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in + diff --git a/flagai/model/mm/modules/diffusionmodules/__init__.py b/flagai/model/mm/modules/diffusionmodules/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/flagai/model/mm/modules/diffusionmodules/model.py b/flagai/model/mm/modules/diffusionmodules/model.py new file mode 100644 index 00000000..3c554b13 --- /dev/null +++ b/flagai/model/mm/modules/diffusionmodules/model.py @@ -0,0 +1,852 @@ +# pytorch_diffusion + derived encoder decoder +import math +import torch +import torch.nn as nn +import numpy as np +from einops import rearrange +from typing import Optional, Any + +from flagai.model.mm.modules.attention import MemoryEfficientCrossAttention + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + print("No module 'xformers'. Proceeding without it.") + + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0,1,0,0)) + return emb + + +def nonlinearity(x): + # swish + return x*torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=2, + padding=0) + + def forward(self, x): + if self.with_conv: + pad = (0,1,0,1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + + +class ResnetBlock(nn.Module): + def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, + dropout, temb_channels=512): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, + out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + else: + self.nin_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x+h + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = q.reshape(b,c,h*w) + q = q.permute(0,2,1) # b,hw,c + k = k.reshape(b,c,h*w) # b,c,hw + w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b,c,h*w) + w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b,c,h,w) + + h_ = self.proj_out(h_) + + return x+h_ + +class MemoryEfficientAttnBlock(nn.Module): + """ + Uses xformers efficient implementation, + see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + Note: this is a single-head self-attention operation + """ + # + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.attention_op: Optional[Any] = None + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + B, C, H, W = q.shape + q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v)) + + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(B, t.shape[1], 1, C) + .permute(0, 2, 1, 3) + .reshape(B * 1, t.shape[1], C) + .contiguous(), + (q, k, v), + ) + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + out = ( + out.unsqueeze(0) + .reshape(B, 1, out.shape[1], C) + .permute(0, 2, 1, 3) + .reshape(B, out.shape[1], C) + ) + out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C) + out = self.proj_out(out) + return x+out + + +class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention): + def forward(self, x, context=None, mask=None): + b, c, h, w = x.shape + x = rearrange(x, 'b c h w -> b (h w) c') + out = super().forward(x, context=context, mask=mask) + out = rearrange(out, 'b (h w) c -> b c h w', h=h, w=w, c=c) + return x + out + + +def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None): + assert attn_type in ["vanilla", "vanilla-xformers", "memory-efficient-cross-attn", "linear", "none"], f'attn_type {attn_type} unknown' + if XFORMERS_IS_AVAILBLE and attn_type == "vanilla": + attn_type = "vanilla-xformers" + print(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + assert attn_kwargs is None + return AttnBlock(in_channels) + elif attn_type == "vanilla-xformers": + print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...") + return MemoryEfficientAttnBlock(in_channels) + elif type == "memory-efficient-cross-attn": + attn_kwargs["query_dim"] = in_channels + return MemoryEfficientCrossAttentionWrapper(**attn_kwargs) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + raise NotImplementedError() + + +class Model(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch*4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList([ + torch.nn.Linear(self.ch, + self.temb_ch), + torch.nn.Linear(self.temb_ch, + self.temb_ch), + ]) + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + skip_in = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + if i_block == self.num_res_blocks: + skip_in = ch*in_ch_mult[i_level] + block.append(ResnetBlock(in_channels=block_in+skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x, t=None, context=None): + #assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla", + **ignore_kwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + 2*z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # timestep embedding + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class Decoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + attn_type="vanilla", **ignorekwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,)+tuple(ch_mult) + block_in = ch*ch_mult[self.num_resolutions-1] + curr_res = resolution // 2**(self.num_resolutions-1) + self.z_shape = (1,z_channels,curr_res,curr_res) + print("Working with z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape))) + + # z to block_in + self.conv_in = torch.nn.Conv2d(z_channels, + block_in, + kernel_size=3, + stride=1, + padding=1) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, z): + #assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block](h, temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + if self.tanh_out: + h = torch.tanh(h) + return h + + +class SimpleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, *args, **kwargs): + super().__init__() + self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), + ResnetBlock(in_channels=in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=2 * in_channels, + out_channels=4 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=4 * in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + nn.Conv2d(2*in_channels, in_channels, 1), + Upsample(in_channels, with_conv=True)]) + # end + self.norm_out = Normalize(in_channels) + self.conv_out = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + for i, layer in enumerate(self.model): + if i in [1,2,3]: + x = layer(x, None) + else: + x = layer(x) + + h = self.norm_out(x) + h = nonlinearity(h) + x = self.conv_out(h) + return x + + +class UpsampleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, + ch_mult=(2,2), dropout=0.0): + super().__init__() + # upsampling + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + block_in = in_channels + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.res_blocks = nn.ModuleList() + self.upsample_blocks = nn.ModuleList() + for i_level in range(self.num_resolutions): + res_block = [] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + res_block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + self.res_blocks.append(nn.ModuleList(res_block)) + if i_level != self.num_resolutions - 1: + self.upsample_blocks.append(Upsample(block_in, True)) + curr_res = curr_res * 2 + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # upsampling + h = x + for k, i_level in enumerate(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.res_blocks[i_level][i_block](h, None) + if i_level != self.num_resolutions - 1: + h = self.upsample_blocks[k](h) + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class LatentRescaler(nn.Module): + def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2): + super().__init__() + # residual block, interpolate, residual block + self.factor = factor + self.conv_in = nn.Conv2d(in_channels, + mid_channels, + kernel_size=3, + stride=1, + padding=1) + self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + self.attn = AttnBlock(mid_channels) + self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + + self.conv_out = nn.Conv2d(mid_channels, + out_channels, + kernel_size=1, + ) + + def forward(self, x): + x = self.conv_in(x) + for block in self.res_block1: + x = block(x, None) + x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor)))) + x = self.attn(x) + for block in self.res_block2: + x = block(x, None) + x = self.conv_out(x) + return x + + +class MergedRescaleEncoder(nn.Module): + def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, + ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + intermediate_chn = ch * ch_mult[-1] + self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult, + z_channels=intermediate_chn, double_z=False, resolution=resolution, + attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv, + out_ch=None) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn, + mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth) + + def forward(self, x): + x = self.encoder(x) + x = self.rescaler(x) + return x + + +class MergedRescaleDecoder(nn.Module): + def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8), + dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + tmp_chn = z_channels*ch_mult[-1] + self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout, + resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks, + ch_mult=ch_mult, resolution=resolution, ch=ch) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn, + out_channels=tmp_chn, depth=rescale_module_depth) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Upsampler(nn.Module): + def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2): + super().__init__() + assert out_size >= in_size + num_blocks = int(np.log2(out_size//in_size))+1 + factor_up = 1.+ (out_size % in_size) + print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}") + self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels, + out_channels=in_channels) + self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2, + attn_resolutions=[], in_channels=None, ch=in_channels, + ch_mult=[ch_mult for _ in range(num_blocks)]) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Resize(nn.Module): + def __init__(self, in_channels=None, learned=False, mode="bilinear"): + super().__init__() + self.with_conv = learned + self.mode = mode + if self.with_conv: + print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode") + raise NotImplementedError() + assert in_channels is not None + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=4, + stride=2, + padding=1) + + def forward(self, x, scale_factor=1.0): + if scale_factor==1.0: + return x + else: + x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor) + return x diff --git a/flagai/model/mm/modules/diffusionmodules/openaimodel.py b/flagai/model/mm/modules/diffusionmodules/openaimodel.py new file mode 100644 index 00000000..cee042d8 --- /dev/null +++ b/flagai/model/mm/modules/diffusionmodules/openaimodel.py @@ -0,0 +1,903 @@ +from abc import abstractmethod +import math +import inspect +import numpy as np +import torch as th +import torch.nn as nn +import torch.nn.functional as F +from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_ + +from ldm.modules.diffusionmodules.util import ( + checkpoint, + conv_nd, + linear, + avg_pool_nd, + zero_module, + normalization, + timestep_embedding, +) +from ldm.modules.attention import SpatialTransformer +from ldm.util import exists + + +# dummy replace +def convert_module_to_f16(x): + pass + +def convert_module_to_f32(x): + pass + + +## go +class AttentionPool2d(nn.Module): + """ + Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py + """ + + def __init__( + self, + spacial_dim: int, + embed_dim: int, + num_heads_channels: int, + output_dim: int = None, + ): + super().__init__() + self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5) + self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) + self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) + self.num_heads = embed_dim // num_heads_channels + self.attention = QKVAttention(self.num_heads) + + def forward(self, x): + b, c, *_spatial = x.shape + x = x.reshape(b, c, -1) # NC(HW) + x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1) + x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1) + x = self.qkv_proj(x) + x = self.attention(x) + x = self.c_proj(x) + return x[:, :, 0] + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward(self, x, emb, context=None, heypernetwork=None): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate( + x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" + ) + else: + x = F.interpolate(x, scale_factor=2, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + +class TransposedUpsample(nn.Module): + 'Learned 2x upsampling without padding' + def __init__(self, channels, out_channels=None, ks=5): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + + self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) + + def forward(self,x): + return self.up(x) + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=padding + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + return checkpoint( + self._forward, (x, emb), self.parameters(), self.use_checkpoint + ) + + + def _forward(self, x, emb): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels, + num_heads=1, + num_head_channels=-1, + use_checkpoint=False, + use_new_attention_order=False, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels) + self.qkv = conv_nd(1, channels, channels * 3, 1) + if use_new_attention_order: + # split qkv before split heads + self.attention = QKVAttention(self.num_heads) + else: + # split heads before split qkv + self.attention = QKVAttentionLegacy(self.num_heads) + + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x): + return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! + #return pt_checkpoint(self._forward, x) # pytorch + + def _forward(self, x): + b, c, *spatial = x.shape + x = x.reshape(b, c, -1) + qkv = self.qkv(self.norm(x)) + h = self.attention(qkv) + h = self.proj_out(h) + return (x + h).reshape(b, c, *spatial) + + +def count_flops_attn(model, _x, y): + """ + A counter for the `thop` package to count the operations in an + attention operation. + Meant to be used like: + macs, params = thop.profile( + model, + inputs=(inputs, timestamps), + custom_ops={QKVAttention: QKVAttention.count_flops}, + ) + """ + b, c, *spatial = y[0].shape + num_spatial = int(np.prod(spatial)) + # We perform two matmuls with the same number of ops. + # The first computes the weight matrix, the second computes + # the combination of the value vectors. + matmul_ops = 2 * b * (num_spatial ** 2) * c + model.total_ops += th.DoubleTensor([matmul_ops]) + + +class QKVAttentionLegacy(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention and splits in a different order. + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.chunk(3, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", + (q * scale).view(bs * self.n_heads, ch, length), + (k * scale).view(bs * self.n_heads, ch, length), + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + +class HypernetworkModule(nn.Module): + multiplier = 1.0 + activation_dict = { + "linear": nn.Identity, + "relu": nn.ReLU, + "leakyrelu": nn.LeakyReLU, + "elu": nn.ELU, + "swish": nn.Hardswish, + "tanh": nn.Tanh, + "sigmoid": nn.Sigmoid, + } + activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'}) + + def __init__(self, dim=1024, state_dict=None, layer_structure=[1,2,2,1], activation_func="relu", weight_init='Normal', + add_layer_norm=False, use_dropout=False, activate_output=False, last_layer_dropout=True): + super().__init__() + + assert layer_structure is not None, "layer_structure must not be None" + assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!" + assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!" + + linears = [] + for i in range(len(layer_structure) - 1): + + # Add a fully-connected layer + linears.append(nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) + + # Add an activation func except last layer + if activation_func == "linear" or activation_func is None or (i >= len(layer_structure) - 2 and not activate_output): + pass + elif activation_func in self.activation_dict: + linears.append(self.activation_dict[activation_func]()) + else: + raise RuntimeError(f'hypernetwork uses an unsupported activation function: {activation_func}') + + # Add layer normalization + if add_layer_norm: + linears.append(nn.LayerNorm(int(dim * layer_structure[i+1]))) + + # Add dropout except last layer + if use_dropout and (i < len(layer_structure) - 3 or last_layer_dropout and i < len(layer_structure) - 2): + linears.append(nn.Dropout(p=0.3)) + + self.project_in = nn.Sequential(*[nn.Linear(768, 1024), nn.LayerNorm(1024)]) + self.linear = nn.Sequential(*linears) + + if state_dict is not None: + self.fix_old_state_dict(state_dict) + self.load_state_dict(state_dict) + else: + for layer in self.linear: + if type(layer) == nn.Linear or type(layer) == nn.LayerNorm: + w, b = layer.weight.data, layer.bias.data + if weight_init == "Normal" or type(layer) == nn.LayerNorm: + normal_(w, mean=0.0, std=0.01) + normal_(b, mean=0.0, std=0) + elif weight_init == 'XavierUniform': + xavier_uniform_(w) + zeros_(b) + elif weight_init == 'XavierNormal': + xavier_normal_(w) + zeros_(b) + elif weight_init == 'KaimingUniform': + kaiming_uniform_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu') + zeros_(b) + elif weight_init == 'KaimingNormal': + kaiming_normal_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu') + zeros_(b) + else: + raise KeyError(f"Key {weight_init} is not defined as initialization!") + # self.to(devices.device) + + def fix_old_state_dict(self, state_dict): + changes = { + 'linear1.bias': 'linear.0.bias', + 'linear1.weight': 'linear.0.weight', + 'linear2.bias': 'linear.1.bias', + 'linear2.weight': 'linear.1.weight', + } + + for fr, to in changes.items(): + x = state_dict.get(fr, None) + if x is None: + continue + + del state_dict[fr] + state_dict[to] = x + + def forward(self, x): + x = self.project_in(x) + return x + self.linear(x) * self.multiplier + # return self.linear(x) * self.multiplier + + def trainables(self): + layer_structure = [] + for layer in self.linear: + if type(layer) == nn.Linear or type(layer) == nn.LayerNorm: + layer_structure += [layer.weight, layer.bias] + return layer_structure + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param use_new_attention_order: use a different attention pattern for potentially + increased efficiency. + """ + + def __init__( + self, + image_size, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + num_classes=None, + use_checkpoint=False, + use_fp16=False, + num_heads=-1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + resblock_updown=False, + use_new_attention_order=False, + use_spatial_transformer=False, # custom transformer support + transformer_depth=1, # custom transformer support + context_dim=None, # custom transformer support + n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model + legacy=True, + disable_self_attentions=None, + num_attention_blocks=None, + disable_middle_self_attn=False, + use_linear_in_transformer=False, + use_hypernetwork=False + ): + super().__init__() + if use_spatial_transformer: + assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' + + if context_dim is not None: + assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' + from omegaconf.listconfig import ListConfig + if type(context_dim) == ListConfig: + context_dim = list(context_dim) + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.image_size = image_size + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError("provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult") + self.num_res_blocks = num_res_blocks + if disable_self_attentions is not None: + # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) + print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set.") + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.dtype = th.float16 if use_fp16 else th.float32 + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.predict_codebook_ids = n_embed is not None + self.use_hypernetwork = use_hypernetwork + + if self.use_hypernetwork: + self.hypernetwork_layers = nn.Sequential(*[HypernetworkModule(), HypernetworkModule()]) + else: + self.hypernetwork = None + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + print("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + else: + raise ValueError() + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(self.num_res_blocks[level] + 1): + ich = input_block_chans.pop() + layers = [ + ResBlock( + ch + ich, + time_embed_dim, + dropout, + out_channels=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or i < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads_upsample, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + if level and i == self.num_res_blocks[level]: + out_ch = ch + layers.append( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + if self.predict_codebook_ids: + self.id_predictor = nn.Sequential( + normalization(ch), + conv_nd(dims, model_channels, n_embed, 1), + #nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits + ) + + # hypernetwork = Hypernetwork() + # self.input_blocks.Sp.crossattention.hypernetwork = self.h + + def convert_to_fp16(self): + """ + Convert the torso of the model to float16. + """ + self.input_blocks.apply(convert_module_to_f16) + self.middle_block.apply(convert_module_to_f16) + self.output_blocks.apply(convert_module_to_f16) + + def convert_to_fp32(self): + """ + Convert the torso of the model to float32. + """ + self.input_blocks.apply(convert_module_to_f32) + self.middle_block.apply(convert_module_to_f32) + self.output_blocks.apply(convert_module_to_f32) + + def forward(self, x, timesteps=None, context=None, y=None,**kwargs): + """ + Apply the model to an input batch. + :param x: an [N x C x ...] Tensor of inputs. + :param timesteps: a 1-D batch of timesteps. + :param context: conditioning plugged in via crossattn + :param y: an [N] Tensor of labels, if class-conditional. + :return: an [N x C x ...] Tensor of outputs. + """ + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + if self.use_hypernetwork: + context_k = self.hypernetwork_layers[0](context) + context_v = self.hypernetwork_layers[1](context) + context = [context_k, context_v] + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) + + h = x.type(self.dtype) + for module in self.input_blocks: + # h:(4,4,64,64) emb:(4,1280) context:(4,77,1024) + # import pdb + # pdb.set_trace() + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, context) + h = h.type(x.dtype) + if self.predict_codebook_ids: + return self.id_predictor(h) + else: + return self.out(h) diff --git a/flagai/model/mm/modules/diffusionmodules/upscaling.py b/flagai/model/mm/modules/diffusionmodules/upscaling.py new file mode 100644 index 00000000..03816662 --- /dev/null +++ b/flagai/model/mm/modules/diffusionmodules/upscaling.py @@ -0,0 +1,81 @@ +import torch +import torch.nn as nn +import numpy as np +from functools import partial + +from ldm.modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule +from ldm.util import default + + +class AbstractLowScaleModel(nn.Module): + # for concatenating a downsampled image to the latent representation + def __init__(self, noise_schedule_config=None): + super(AbstractLowScaleModel, self).__init__() + if noise_schedule_config is not None: + self.register_schedule(**noise_schedule_config) + + def register_schedule(self, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def forward(self, x): + return x, None + + def decode(self, x): + return x + + +class SimpleImageConcat(AbstractLowScaleModel): + # no noise level conditioning + def __init__(self): + super(SimpleImageConcat, self).__init__(noise_schedule_config=None) + self.max_noise_level = 0 + + def forward(self, x): + # fix to constant noise level + return x, torch.zeros(x.shape[0], device=x.device).long() + + +class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel): + def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False): + super().__init__(noise_schedule_config=noise_schedule_config) + self.max_noise_level = max_noise_level + + def forward(self, x, noise_level=None): + if noise_level is None: + noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long() + else: + assert isinstance(noise_level, torch.Tensor) + z = self.q_sample(x, noise_level) + return z, noise_level + + + diff --git a/flagai/model/mm/modules/diffusionmodules/util.py b/flagai/model/mm/modules/diffusionmodules/util.py new file mode 100644 index 00000000..5b09d609 --- /dev/null +++ b/flagai/model/mm/modules/diffusionmodules/util.py @@ -0,0 +1,270 @@ +# adopted from +# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# and +# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +# and +# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py +# +# thanks! + + +import os +import math +import torch +import torch.nn as nn +import numpy as np +from einops import repeat + +from flagai.model.mm.utils import instantiate_from_config + + +def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if schedule == "linear": + betas = ( + torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2 + ) + + elif schedule == "cosine": + timesteps = ( + torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s + ) + alphas = timesteps / (1 + cosine_s) * np.pi / 2 + alphas = torch.cos(alphas).pow(2) + alphas = alphas / alphas[0] + betas = 1 - alphas[1:] / alphas[:-1] + betas = np.clip(betas, a_min=0, a_max=0.999) + + elif schedule == "sqrt_linear": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) + elif schedule == "sqrt": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5 + else: + raise ValueError(f"schedule '{schedule}' unknown.") + return betas.numpy() + + +def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): + if ddim_discr_method == 'uniform': + c = num_ddpm_timesteps // num_ddim_timesteps + ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c))) + elif ddim_discr_method == 'quad': + ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int) + else: + raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"') + + # assert ddim_timesteps.shape[0] == num_ddim_timesteps + # add one to get the final alpha values right (the ones from first scale to data during sampling) + steps_out = ddim_timesteps + 1 + if verbose: + print(f'Selected timesteps for ddim sampler: {steps_out}') + return steps_out + + +def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): + # select alphas for computing the variance schedule + alphas = alphacums[ddim_timesteps] + alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) + + # according the the formula provided in https://arxiv.org/abs/2010.02502 + sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) + if verbose: + print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') + print(f'For the chosen value of eta, which is {eta}, ' + f'this results in the following sigma_t schedule for ddim sampler {sigmas}') + return sigmas, alphas, alphas_prev + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + + +def extract_into_tensor(a, t, x_shape): + b, *_ = t.shape + out = a.gather(-1, t) + return out.reshape(b, *((1,) * (len(x_shape) - 1))) + + +def checkpoint(func, inputs, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. + :param func: the function to evaluate. + :param inputs: the argument sequence to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + args = tuple(inputs) + tuple(params) + return CheckpointFunction.apply(func, len(inputs), *args) + else: + return func(*inputs) + + +class CheckpointFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, run_function, length, *args): + ctx.run_function = run_function + ctx.input_tensors = list(args[:length]) + ctx.input_params = list(args[length:]) + ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(), + "dtype": torch.get_autocast_gpu_dtype(), + "cache_enabled": torch.is_autocast_cache_enabled()} + with torch.no_grad(): + output_tensors = ctx.run_function(*ctx.input_tensors) + return output_tensors + + @staticmethod + def backward(ctx, *output_grads): + ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors] + with torch.enable_grad(), \ + torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs): + # Fixes a bug where the first op in run_function modifies the + # Tensor storage in place, which is not allowed for detach()'d + # Tensors. + shallow_copies = [x.view_as(x) for x in ctx.input_tensors] + output_tensors = ctx.run_function(*shallow_copies) + input_grads = torch.autograd.grad( + output_tensors, + ctx.input_tensors + ctx.input_params, + output_grads, + allow_unused=True, + ) + del ctx.input_tensors + del ctx.input_params + del output_tensors + return (None, None) + input_grads + + +def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): + """ + Create sinusoidal timestep embeddings. + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an [N x dim] Tensor of positional embeddings. + """ + if not repeat_only: + half = dim // 2 + freqs = torch.exp( + -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half + ).to(device=timesteps.device) + args = timesteps[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + else: + embedding = repeat(timesteps, 'b -> b d', d=dim) + return embedding + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def normalization(channels): + """ + Make a standard normalization layer. + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNorm32(32, channels) + + +# PyTorch 1.7 has SiLU, but we support PyTorch 1.5. +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + + +class GroupNorm32(nn.GroupNorm): + def forward(self, x): + return super().forward(x.float()).type(x.dtype) + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +class HybridConditioner(nn.Module): + + def __init__(self, c_concat_config, c_crossattn_config): + super().__init__() + self.concat_conditioner = instantiate_from_config(c_concat_config) + self.crossattn_conditioner = instantiate_from_config(c_crossattn_config) + + def forward(self, c_concat, c_crossattn): + c_concat = self.concat_conditioner(c_concat) + c_crossattn = self.crossattn_conditioner(c_crossattn) + return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]} + + +def noise_like(shape, device, repeat=False): + repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1))) + noise = lambda: torch.randn(shape, device=device) + return repeat_noise() if repeat else noise() \ No newline at end of file diff --git a/flagai/model/mm/modules/distributions/__init__.py b/flagai/model/mm/modules/distributions/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/flagai/model/mm/modules/distributions/distributions.py b/flagai/model/mm/modules/distributions/distributions.py new file mode 100644 index 00000000..f2b8ef90 --- /dev/null +++ b/flagai/model/mm/modules/distributions/distributions.py @@ -0,0 +1,92 @@ +import torch +import numpy as np + + +class AbstractDistribution: + def sample(self): + raise NotImplementedError() + + def mode(self): + raise NotImplementedError() + + +class DiracDistribution(AbstractDistribution): + def __init__(self, value): + self.value = value + + def sample(self): + return self.value + + def mode(self): + return self.value + + +class DiagonalGaussianDistribution(object): + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) + + def sample(self): + x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device) + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.]) + else: + if other is None: + return 0.5 * torch.sum(torch.pow(self.mean, 2) + + self.var - 1.0 - self.logvar, + dim=[1, 2, 3]) + else: + return 0.5 * torch.sum( + torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var - 1.0 - self.logvar + other.logvar, + dim=[1, 2, 3]) + + def nll(self, sample, dims=[1,2,3]): + if self.deterministic: + return torch.Tensor([0.]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum( + logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims) + + def mode(self): + return self.mean + + +def normal_kl(mean1, logvar1, mean2, logvar2): + """ + source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12 + Compute the KL divergence between two gaussians. + Shapes are automatically broadcasted, so batches can be compared to + scalars, among other use cases. + """ + tensor = None + for obj in (mean1, logvar1, mean2, logvar2): + if isinstance(obj, torch.Tensor): + tensor = obj + break + assert tensor is not None, "at least one argument must be a Tensor" + + # Force variances to be Tensors. Broadcasting helps convert scalars to + # Tensors, but it does not work for torch.exp(). + logvar1, logvar2 = [ + x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) + for x in (logvar1, logvar2) + ] + + return 0.5 * ( + -1.0 + + logvar2 + - logvar1 + + torch.exp(logvar1 - logvar2) + + ((mean1 - mean2) ** 2) * torch.exp(-logvar2) + ) diff --git a/flagai/model/mm/modules/ema.py b/flagai/model/mm/modules/ema.py new file mode 100644 index 00000000..5b789ebb --- /dev/null +++ b/flagai/model/mm/modules/ema.py @@ -0,0 +1,81 @@ +import torch +from torch import nn + + +class LitEma(nn.Module): + def __init__(self, model, decay=0.9999, use_num_upates=True): + super().__init__() + if decay < 0.0 or decay > 1.0: + raise ValueError('Decay must be between 0 and 1') + + self.m_name2s_name = {} + self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates + else torch.tensor(-1, dtype=torch.int)) + + for name, p in model.named_parameters(): + if p.requires_grad: + # remove as '.'-character is not allowed in buffers + s_name = name.replace('.', '') + self.m_name2s_name.update({name: s_name}) + self.register_buffer(s_name, p.clone().detach().data) + + self.collected_params = [] + + def reset_num_updates(self): + del self.num_updates + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int)) + + def forward(self, model): + decay = self.decay + + if self.num_updates >= 0: + self.num_updates += 1 + # 这个地方是动态调整decay的数值，使得在训练的一开始参数的更新还是比较快的 + decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates)) + + one_minus_decay = 1.0 - decay + + with torch.no_grad(): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + + for key in m_param: + if m_param[key].requires_grad: + sname = self.m_name2s_name[key] + shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) + shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) + else: + assert not key in self.m_name2s_name + + def copy_to(self, model): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + for key in m_param: + if m_param[key].requires_grad: + m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) + else: + assert not key in self.m_name2s_name + + def store(self, parameters): + """ + Save the current parameters for restoring later. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + temporarily stored. + """ + self.collected_params = [param.clone() for param in parameters] + + def restore(self, parameters): + """ + Restore the parameters stored with the `store` method. + Useful to validate the model with EMA parameters without affecting the + original optimization process. Store the parameters before the + `copy_to` method. After validation (or model saving), use this to + restore the former parameters. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. + """ + for c_param, param in zip(self.collected_params, parameters): + param.data.copy_(c_param.data) diff --git a/flagai/model/mm/modules/encoders/__init__.py b/flagai/model/mm/modules/encoders/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/flagai/model/mm/modules/encoders/modules.py b/flagai/model/mm/modules/encoders/modules.py new file mode 100644 index 00000000..9942075e --- /dev/null +++ b/flagai/model/mm/modules/encoders/modules.py @@ -0,0 +1,313 @@ +import torch +import torch.nn as nn +from torch.utils.checkpoint import checkpoint + +from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel + +import open_clip +from ldm.util import default, count_params + +##### 为了克服网速十分慢，手动load了一下open_clip的模型权重########## + +import math +import collections.abc +from itertools import repeat +import torch.nn.functional as F + + +def _ntuple(n): + def parse(x): + if isinstance(x, collections.abc.Iterable): + return x + return tuple(repeat(x, n)) + return parse + +to_2tuple = _ntuple(2) + +def resize_pos_embed(state_dict, model, interpolation: str = 'bicubic', antialias: bool = True): + # Rescale the grid of position embeddings when loading from state_dict + old_pos_embed = state_dict.get('visual.positional_embedding', None) + if old_pos_embed is None or not hasattr(model.visual, 'grid_size'): + return + grid_size = to_2tuple(model.visual.grid_size) + extra_tokens = 1 # FIXME detect different token configs (ie no class token, or more) + new_seq_len = grid_size[0] * grid_size[1] + extra_tokens + if new_seq_len == old_pos_embed.shape[0]: + return + + if extra_tokens: + pos_emb_tok, pos_emb_img = old_pos_embed[:extra_tokens], old_pos_embed[extra_tokens:] + else: + pos_emb_tok, pos_emb_img = None, old_pos_embed + old_grid_size = to_2tuple(int(math.sqrt(len(pos_emb_img)))) + + # logging.info('Resizing position embedding grid-size from %s to %s', old_grid_size, grid_size) + pos_emb_img = pos_emb_img.reshape(1, old_grid_size[0], old_grid_size[1], -1).permute(0, 3, 1, 2) + pos_emb_img = F.interpolate( + pos_emb_img, + size=grid_size, + mode=interpolation, + antialias=antialias, + align_corners=False, + ) + pos_emb_img = pos_emb_img.permute(0, 2, 3, 1).reshape(1, grid_size[0] * grid_size[1], -1)[0] + if pos_emb_tok is not None: + new_pos_embed = torch.cat([pos_emb_tok, pos_emb_img], dim=0) + else: + new_pos_embed = pos_emb_img + state_dict['visual.positional_embedding'] = new_pos_embed + +def convert_to_custom_text_state_dict(state_dict: dict): + if 'text_projection' in state_dict: + # old format state_dict, move text tower -> .text + new_state_dict = {} + for k, v in state_dict.items(): + if any(k.startswith(p) for p in ( + 'text_projection', + 'positional_embedding', + 'token_embedding', + 'transformer', + 'ln_final', + )): + k = 'text.' + k + new_state_dict[k] = v + return new_state_dict + return state_dict + + +def load_state_dict(checkpoint_path: str, map_location='cpu'): + checkpoint = torch.load(checkpoint_path, map_location=map_location) + if isinstance(checkpoint, dict) and 'state_dict' in checkpoint: + state_dict = checkpoint['state_dict'] + else: + state_dict = checkpoint + if next(iter(state_dict.items()))[0].startswith('module'): + state_dict = {k[7:]: v for k, v in state_dict.items()} + return state_dict + + +def load_checkpoint(model, checkpoint_path, strict=True): + state_dict = load_state_dict(checkpoint_path) + # detect old format and make compatible with new format + if 'positional_embedding' in state_dict and not hasattr(model, 'positional_embedding'): + state_dict = convert_to_custom_text_state_dict(state_dict) + resize_pos_embed(state_dict, model) + incompatible_keys = model.load_state_dict(state_dict, strict=strict) + return incompatible_keys + +###### 这部代码就先强行插入在这里####### + + + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class IdentityEncoder(AbstractEncoder): + + def encode(self, x): + return x + + +class ClassEmbedder(nn.Module): + def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1): + super().__init__() + self.key = key + self.embedding = nn.Embedding(n_classes, embed_dim) + self.n_classes = n_classes + self.ucg_rate = ucg_rate + + def forward(self, batch, key=None, disable_dropout=False): + if key is None: + key = self.key + # this is for use in crossattn + c = batch[key][:, None] + if self.ucg_rate > 0. and not disable_dropout: + mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate) + c = mask * c + (1-mask) * torch.ones_like(c)*(self.n_classes-1) + c = c.long() + c = self.embedding(c) + return c + + def get_unconditional_conditioning(self, bs, device="cuda"): + uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) + uc = torch.ones((bs,), device=device) * uc_class + uc = {self.key: uc} + return uc + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +class FrozenT5Embedder(AbstractEncoder): + """Uses the T5 transformer encoder for text""" + def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = T5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length # TODO: typical value? + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + #self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens) + + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenCLIPEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for text (from huggingface)""" + LAYERS = [ + "last", + "pooled", + "hidden" + ] + def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, + freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32 + super().__init__() + assert layer in self.LAYERS + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.transformer = CLIPTextModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + self.layer_idx = layer_idx + if layer == "hidden": + assert layer_idx is not None + assert 0 <= abs(layer_idx) <= 12 + + def freeze(self): + self.transformer = self.transformer.eval() + #self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer=="hidden") + if self.layer == "last": + z = outputs.last_hidden_state + elif self.layer == "pooled": + z = outputs.pooler_output[:, None, :] + else: + z = outputs.hidden_states[self.layer_idx] + return z + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP transformer encoder for text + """ + LAYERS = [ + #"pooled", + "last", + "penultimate" + ] + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="penultimate"): + super().__init__() + assert layer in self.LAYERS + path = '/share/project/yfl/database/ckpt/stable_diffusion_2.0/models--laion--CLIP-ViT-H-14-laion2B-s32B-b79K/snapshots/58a1e03a7acfacbe6b95ebc24ae0394eda6a14fc/open_clip_pytorch_model.bin' + # model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version, cache_dir="/share/project/yfl/database/hub/") + # model, _, _ = open_clip.create_model_from_pretrained(arch, "/share/project/liuguang/CLIP-ViT-H-14/", device=torch.device('cpu')) + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu')) + tmp = load_checkpoint(model=model, checkpoint_path=path) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + tokens = open_clip.tokenize(text) + z = self.encode_with_transformer(tokens.to(self.device)) + return z + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.model.ln_final(x) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask = None): + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - self.layer_idx: + break + if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting(): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + return x + + def encode(self, text): + return self(text) + +class FrozenCLIPT5Encoder(AbstractEncoder): + def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda", + clip_max_length=77, t5_max_length=77): + super().__init__() + self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length) + self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) + print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder)*1.e-6:.2f} M parameters, " + f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder)*1.e-6:.2f} M params.") + + def encode(self, text): + return self(text) + + def forward(self, text): + clip_z = self.clip_encoder.encode(text) + t5_z = self.t5_encoder.encode(text) + return [clip_z, t5_z] + + + + + + + + diff --git a/flagai/model/mm/openaimodel.py b/flagai/model/mm/openaimodel.py new file mode 100644 index 00000000..b344eaa8 --- /dev/null +++ b/flagai/model/mm/openaimodel.py @@ -0,0 +1,904 @@ +from abc import abstractmethod +import math +import inspect + +import numpy as np +import torch as th +import torch.nn as nn +import torch.nn.functional as F +from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_ + +from flagai.model.mm.modules.diffusionmodules.util import ( + checkpoint, + conv_nd, + linear, + avg_pool_nd, + zero_module, + normalization, + timestep_embedding, +) +from flagai.model.mm.modules.attention import SpatialTransformer +from flagai.model.mm.utils import exists + + +# dummy replace +def convert_module_to_f16(x): + pass + +def convert_module_to_f32(x): + pass + + +## go +class AttentionPool2d(nn.Module): + """ + Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py + """ + + def __init__( + self, + spacial_dim: int, + embed_dim: int, + num_heads_channels: int, + output_dim: int = None, + ): + super().__init__() + self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5) + self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) + self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) + self.num_heads = embed_dim // num_heads_channels + self.attention = QKVAttention(self.num_heads) + + def forward(self, x): + b, c, *_spatial = x.shape + x = x.reshape(b, c, -1) # NC(HW) + x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1) + x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1) + x = self.qkv_proj(x) + x = self.attention(x) + x = self.c_proj(x) + return x[:, :, 0] + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward(self, x, emb, context=None, heypernetwork=None): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate( + x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" + ) + else: + x = F.interpolate(x, scale_factor=2, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + +class TransposedUpsample(nn.Module): + 'Learned 2x upsampling without padding' + def __init__(self, channels, out_channels=None, ks=5): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + + self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) + + def forward(self,x): + return self.up(x) + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=padding + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + return checkpoint( + self._forward, (x, emb), self.parameters(), self.use_checkpoint + ) + + + def _forward(self, x, emb): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels, + num_heads=1, + num_head_channels=-1, + use_checkpoint=False, + use_new_attention_order=False, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels) + self.qkv = conv_nd(1, channels, channels * 3, 1) + if use_new_attention_order: + # split qkv before split heads + self.attention = QKVAttention(self.num_heads) + else: + # split heads before split qkv + self.attention = QKVAttentionLegacy(self.num_heads) + + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x): + return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! + #return pt_checkpoint(self._forward, x) # pytorch + + def _forward(self, x): + b, c, *spatial = x.shape + x = x.reshape(b, c, -1) + qkv = self.qkv(self.norm(x)) + h = self.attention(qkv) + h = self.proj_out(h) + return (x + h).reshape(b, c, *spatial) + + +def count_flops_attn(model, _x, y): + """ + A counter for the `thop` package to count the operations in an + attention operation. + Meant to be used like: + macs, params = thop.profile( + model, + inputs=(inputs, timestamps), + custom_ops={QKVAttention: QKVAttention.count_flops}, + ) + """ + b, c, *spatial = y[0].shape + num_spatial = int(np.prod(spatial)) + # We perform two matmuls with the same number of ops. + # The first computes the weight matrix, the second computes + # the combination of the value vectors. + matmul_ops = 2 * b * (num_spatial ** 2) * c + model.total_ops += th.DoubleTensor([matmul_ops]) + + +class QKVAttentionLegacy(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention and splits in a different order. + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.chunk(3, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", + (q * scale).view(bs * self.n_heads, ch, length), + (k * scale).view(bs * self.n_heads, ch, length), + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + +class HypernetworkModule(nn.Module): + multiplier = 1.0 + activation_dict = { + "linear": nn.Identity, + "relu": nn.ReLU, + "leakyrelu": nn.LeakyReLU, + "elu": nn.ELU, + "swish": nn.Hardswish, + "tanh": nn.Tanh, + "sigmoid": nn.Sigmoid, + } + activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'}) + + def __init__(self, dim=1024, state_dict=None, layer_structure=[1,2,2,1], activation_func="relu", weight_init='Normal', + add_layer_norm=False, use_dropout=False, activate_output=False, last_layer_dropout=True): + super().__init__() + + assert layer_structure is not None, "layer_structure must not be None" + assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!" + assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!" + + linears = [] + for i in range(len(layer_structure) - 1): + + # Add a fully-connected layer + linears.append(nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) + + # Add an activation func except last layer + if activation_func == "linear" or activation_func is None or (i >= len(layer_structure) - 2 and not activate_output): + pass + elif activation_func in self.activation_dict: + linears.append(self.activation_dict[activation_func]()) + else: + raise RuntimeError(f'hypernetwork uses an unsupported activation function: {activation_func}') + + # Add layer normalization + if add_layer_norm: + linears.append(nn.LayerNorm(int(dim * layer_structure[i+1]))) + + # Add dropout except last layer + if use_dropout and (i < len(layer_structure) - 3 or last_layer_dropout and i < len(layer_structure) - 2): + linears.append(nn.Dropout(p=0.3)) + + self.project_in = nn.Sequential(*[nn.Linear(768, 1024), nn.LayerNorm(1024)]) + self.linear = nn.Sequential(*linears) + + if state_dict is not None: + self.fix_old_state_dict(state_dict) + self.load_state_dict(state_dict) + else: + for layer in self.linear: + if type(layer) == nn.Linear or type(layer) == nn.LayerNorm: + w, b = layer.weight.data, layer.bias.data + if weight_init == "Normal" or type(layer) == nn.LayerNorm: + normal_(w, mean=0.0, std=0.01) + normal_(b, mean=0.0, std=0) + elif weight_init == 'XavierUniform': + xavier_uniform_(w) + zeros_(b) + elif weight_init == 'XavierNormal': + xavier_normal_(w) + zeros_(b) + elif weight_init == 'KaimingUniform': + kaiming_uniform_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu') + zeros_(b) + elif weight_init == 'KaimingNormal': + kaiming_normal_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu') + zeros_(b) + else: + raise KeyError(f"Key {weight_init} is not defined as initialization!") + # self.to(devices.device) + + def fix_old_state_dict(self, state_dict): + changes = { + 'linear1.bias': 'linear.0.bias', + 'linear1.weight': 'linear.0.weight', + 'linear2.bias': 'linear.1.bias', + 'linear2.weight': 'linear.1.weight', + } + + for fr, to in changes.items(): + x = state_dict.get(fr, None) + if x is None: + continue + + del state_dict[fr] + state_dict[to] = x + + def forward(self, x): + x = self.project_in(x) + return x + self.linear(x) * self.multiplier + # return self.linear(x) * self.multiplier + + def trainables(self): + layer_structure = [] + for layer in self.linear: + if type(layer) == nn.Linear or type(layer) == nn.LayerNorm: + layer_structure += [layer.weight, layer.bias] + return layer_structure + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param use_new_attention_order: use a different attention pattern for potentially + increased efficiency. + """ + + def __init__( + self, + image_size, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + num_classes=None, + use_checkpoint=False, + use_fp16=False, + num_heads=-1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + resblock_updown=False, + use_new_attention_order=False, + use_spatial_transformer=False, # custom transformer support + transformer_depth=1, # custom transformer support + context_dim=None, # custom transformer support + n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model + legacy=True, + disable_self_attentions=None, + num_attention_blocks=None, + disable_middle_self_attn=False, + use_linear_in_transformer=False, + use_hypernetwork=False + ): + super().__init__() + if use_spatial_transformer: + assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' + + if context_dim is not None: + assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' + from omegaconf.listconfig import ListConfig + if type(context_dim) == ListConfig: + context_dim = list(context_dim) + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.image_size = image_size + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError("provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult") + self.num_res_blocks = num_res_blocks + if disable_self_attentions is not None: + # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) + print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set.") + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.dtype = th.float16 if use_fp16 else th.float32 + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.predict_codebook_ids = n_embed is not None + self.use_hypernetwork = use_hypernetwork + + if self.use_hypernetwork: + self.hypernetwork_layers = nn.Sequential(*[HypernetworkModule(), HypernetworkModule()]) + else: + self.hypernetwork = None + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + print("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + else: + raise ValueError() + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(self.num_res_blocks[level] + 1): + ich = input_block_chans.pop() + layers = [ + ResBlock( + ch + ich, + time_embed_dim, + dropout, + out_channels=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or i < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads_upsample, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + if level and i == self.num_res_blocks[level]: + out_ch = ch + layers.append( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + if self.predict_codebook_ids: + self.id_predictor = nn.Sequential( + normalization(ch), + conv_nd(dims, model_channels, n_embed, 1), + #nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits + ) + + # hypernetwork = Hypernetwork() + # self.input_blocks.Sp.crossattention.hypernetwork = self.h + + def convert_to_fp16(self): + """ + Convert the torso of the model to float16. + """ + self.input_blocks.apply(convert_module_to_f16) + self.middle_block.apply(convert_module_to_f16) + self.output_blocks.apply(convert_module_to_f16) + + def convert_to_fp32(self): + """ + Convert the torso of the model to float32. + """ + self.input_blocks.apply(convert_module_to_f32) + self.middle_block.apply(convert_module_to_f32) + self.output_blocks.apply(convert_module_to_f32) + + def forward(self, x, timesteps=None, context=None, y=None,**kwargs): + """ + Apply the model to an input batch. + :param x: an [N x C x ...] Tensor of inputs. + :param timesteps: a 1-D batch of timesteps. + :param context: conditioning plugged in via crossattn + :param y: an [N] Tensor of labels, if class-conditional. + :return: an [N x C x ...] Tensor of outputs. + """ + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + if self.use_hypernetwork: + context_k = self.hypernetwork_layers[0](context) + context_v = self.hypernetwork_layers[1](context) + context = [context_k, context_v] + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) + + h = x.type(self.dtype) + for module in self.input_blocks: + # h:(4,4,64,64) emb:(4,1280) context:(4,77,1024) + # import pdb + # pdb.set_trace() + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, context) + h = h.type(x.dtype) + if self.predict_codebook_ids: + return self.id_predictor(h) + else: + return self.out(h) diff --git a/flagai/model/mm/utils.py b/flagai/model/mm/utils.py index de5b6618..e9e6cee5 100644 --- a/flagai/model/mm/utils.py +++ b/flagai/model/mm/utils.py @@ -7,7 +7,7 @@ from collections import abc from einops import rearrange from functools import partial - +from torch import optim import multiprocessing as mp from threading import Thread from queue import Queue @@ -515,3 +515,112 @@ def extract_into_tensor(a, t, x_shape): b, *_ = t.shape out = a.gather(-1, t) return out.reshape(b, *((1, ) * (len(x_shape) - 1))) + +class AdamWwithEMAandWings(optim.Optimizer): + # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298 + def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8, # TODO: check hyperparameters before using + weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999, # ema decay to match previous code + ema_power=1., param_names=()): + """AdamW that saves EMA versions of the parameters.""" + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0.0 <= weight_decay: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + if not 0.0 <= ema_decay <= 1.0: + raise ValueError("Invalid ema_decay value: {}".format(ema_decay)) + defaults = dict(lr=lr, betas=betas, eps=eps, + weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay, + ema_power=ema_power, param_names=param_names) + super().__init__(params, defaults) + + def __setstate__(self, state): + super().__setstate__(state) + for group in self.param_groups: + group.setdefault('amsgrad', False) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + params_with_grad = [] + grads = [] + exp_avgs = [] + exp_avg_sqs = [] + ema_params_with_grad = [] + state_sums = [] + max_exp_avg_sqs = [] + state_steps = [] + amsgrad = group['amsgrad'] + beta1, beta2 = group['betas'] + ema_decay = group['ema_decay'] + ema_power = group['ema_power'] + + for p in group['params']: + if p.grad is None: + continue + params_with_grad.append(p) + if p.grad.is_sparse: + raise RuntimeError('AdamW does not support sparse gradients') + grads.append(p.grad) + + state = self.state[p] + + # State initialization + if len(state) == 0: + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of squared gradient values + state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of parameter values + state['param_exp_avg'] = p.detach().float().clone() + + exp_avgs.append(state['exp_avg']) + exp_avg_sqs.append(state['exp_avg_sq']) + ema_params_with_grad.append(state['param_exp_avg']) + + if amsgrad: + max_exp_avg_sqs.append(state['max_exp_avg_sq']) + + # update the steps for each param group update + state['step'] += 1 + # record the step after step update + state_steps.append(state['step']) + + optim._functional.adamw(params_with_grad, + grads, + exp_avgs, + exp_avg_sqs, + max_exp_avg_sqs, + state_steps, + amsgrad=amsgrad, + beta1=beta1, + beta2=beta2, + lr=group['lr'], + weight_decay=group['weight_decay'], + eps=group['eps'], + maximize=False) + + cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power) + for param, ema_param in zip(params_with_grad, ema_params_with_grad): + ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay) + + return loss \ No newline at end of file diff --git a/flagai/model/predictor/predictor.py b/flagai/model/predictor/predictor.py index 03cf7ff4..a7005594 100644 --- a/flagai/model/predictor/predictor.py +++ b/flagai/model/predictor/predictor.py @@ -11,6 +11,7 @@ from flagai.model.predictor.gpt import gpt_random_sample_use_cache from flagai.model.predictor.simctg import contrastive_search from flagai.model.mm.Sampler import DDIMSampler, PLMSSampler +from flagai.model.mm.dpm_solver import DPMSolverSampler import os from PIL import Image from tqdm import trange, tqdm @@ -351,6 +352,116 @@ def predict_generate_randomsample(self, import os os._exit(0) + def predict_generate_images_m18(self, + prompt: str, + negative_prompt: str, + outpath: str = "AltDiffusionOutputs", + n_samples: int = 4, + n_rows: int = 0, + skip_grid: bool = False, + skip_save: bool = False, + steps: int = 50, + n_iter: int = 1, + plms: bool = False, + dpm: bool = False, + fixed_code: bool = False, + ddim_eta: float = 0.0, + H: int = 512, + W: int = 512, + C: int = 4, + f: int = 8, + scale: float = 7.5, + from_file: str = None, + seed: int = 34234, + fp16: bool = False): + from torchvision.utils import make_grid + from pytorch_lightning import seed_everything + from flagai.model.predictor.utils import chunk, check_safety, get_safety_checker + safety_checker, safety_feature_extractor = get_safety_checker() + """ + Args: + prompt: the prompt text + out_path: the output path + n_samples: how many images to be generated + skip_grid: not to grid images + skip_save: do not save images + ddim_step: number of steps in ddim model + n_iter: number of iterations + plms: use PLMSSampler + fixed_code: sampled from a initial start code + seed: Random seed + H: height of image + W: width of image + C: channels of images, 4 for colored images + """ + if plms: + sampler = PLMSSampler(self.model) + elif dpm: + sampler = DPMSolverSampler(self.model) + else: + sampler = DDIMSampler(self.model) + + seed_everything(seed) + assert "diffusion" in self.class_name.lower() + batch_size = n_samples + device = next(self.model.parameters()).device + n_rows = n_rows if n_rows > 0 else batch_size + + precision_scope = autocast if opt.precision == "autocast" else nullcontext + sample_path = os.path.join(outpath, "samples") + os.makedirs(sample_path, exist_ok=True) + base_count = len(os.listdir(sample_path)) + grid_count = len(os.listdir(outpath)) - 1 + + start_code = None + if fixed_code: + start_code = torch.randn([n_samples, C, H // f, W // f], + device=device) + + + with torch.no_grad(), \ + precision_scope(True), \ + self.model.ema_scope(): + all_samples = list() + prompts = [batch_size * [prompt]] + for prompts in tqdm(prompts, desc="data"): + uc = None + if scale != 1.0: + uc = self.model.get_learned_conditioning(batch_size * [negative_prompt]) + if isinstance(prompts, tuple): + prompts = list(prompts) + c = self.model.get_learned_conditioning(prompts) + shape = [C, H // f, W // f] + samples, _ = sampler.sample(S=steps, + conditioning=c, + batch_size=n_samples, + shape=shape, + verbose=False, + unconditional_guidance_scale=scale, + unconditional_conditioning=uc, + eta=ddim_eta, + x_T=start_code) + + x_samples = self.model.decode_first_stage(samples) + x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0) + + for x_sample in x_samples: + x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c') + img = Image.fromarray(x_sample.astype(np.uint8)) + + all_samples.append(x_samples) + + # additionally, save as grid + grid = torch.stack(all_samples, 0) + grid = rearrange(grid, 'n b c h w -> (n b) c h w') + grid = make_grid(grid, nrow=n_rows) + + # to image + grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy() + grid = Image.fromarray(grid.astype(np.uint8)) + grid.show() + grid.save('./output.png') + def predict_generate_images(self, prompt: str, outpath: str = "AltDiffusionOutputs", @@ -497,8 +608,6 @@ def predict_generate_images(self, os.path.join(outpath, f'grid-{grid_count:04}.png')) grid_count += 1 - toc = time.time() - print( f"Your samples are ready and waiting for you here: \n{outpath} \n" f" \nEnjoy.") From ee99cf02e839a47f088a3b79b94a8ba0cbbfc62a Mon Sep 17 00:00:00 2001 From: ftgreat Date: Wed, 29 Mar 2023 15:36:26 +0800 Subject: [PATCH 2/7] updated Signed-off-by: ftgreat --- examples/AltCLIP-m18/altclip_inference.py | 1 + flagai/model/mm/AltCLIP.py | 10 +++------- flagai/model/mm/AltDiffusion2.py | 1 + flagai/model/mm/lm/dcn_clip.py | 2 +- 4 files changed, 6 insertions(+), 8 deletions(-) diff --git a/examples/AltCLIP-m18/altclip_inference.py b/examples/AltCLIP-m18/altclip_inference.py index 7be92f1e..788eef40 100644 --- a/examples/AltCLIP-m18/altclip_inference.py +++ b/examples/AltCLIP-m18/altclip_inference.py @@ -1,4 +1,5 @@ import torch +import sys;sys.path.append("/home/yanzhaodong/FlagAI") from PIL import Image from flagai.auto_model.auto_loader import AutoLoader diff --git a/flagai/model/mm/AltCLIP.py b/flagai/model/mm/AltCLIP.py index 999648b1..5a33ca41 100644 --- a/flagai/model/mm/AltCLIP.py +++ b/flagai/model/mm/AltCLIP.py @@ -76,14 +76,11 @@ def __init__(self, num_layers=3, variant='invert', **kwargs): - super().__init__(text_config_dict, vision_config_dict, projection_dim, - logit_scale_init_value, **kwargs) + super().__init__(text_config_dict=None, vision_config_dict=None, projection_dim=None,logit_scale_init_value=None, **kwargs) if text_config_dict is None: text_config_dict = {} # when reload the config from local, we need name to select which class should be instanced. - import pdb;pdb.set_trace() - self.text_config = STUDENT_CONFIG_DICT[ - kwargs['text_config']['model_type']](**kwargs.pop('text_config')) + self.text_config = STUDENT_CONFIG_DICT[kwargs['text_config']['model_type']](**kwargs.pop('text_config')) self.num_layers = num_layers self.text_model_name = text_model_name self.vision_model_name = vision_model_name @@ -100,7 +97,6 @@ def __init__(self, config: AltCLIPConfig, clip_model=None): raise ValueError( "config.vision_config is expected to be of type CLIPVisionConfig but is of type" f" {type(config.vision_config)}.") - text_config = config.text_config vision_config = config.vision_config @@ -447,7 +443,7 @@ def from_pretrain(cls, only_download_config=False, device="cpu", **kwargs): - # super().download(download_path, model_name, only_download_config=only_download_config) + super().download(download_path, model_name, only_download_config=only_download_config) pretrained_model_name_or_path = os.path.join(download_path, model_name) print(pretrained_model_name_or_path) return CLIPHF.from_pretrained(pretrained_model_name_or_path) diff --git a/flagai/model/mm/AltDiffusion2.py b/flagai/model/mm/AltDiffusion2.py index 4c17b7e4..3a2517b1 100644 --- a/flagai/model/mm/AltDiffusion2.py +++ b/flagai/model/mm/AltDiffusion2.py @@ -564,6 +564,7 @@ def __init__(self, else: self.register_buffer('scale_factor', torch.tensor(scale_factor)) self.instantiate_first_stage(first_stage_config) + import pdb;pdb.set_trace() self.instantiate_cond_stage(cond_stage_config) self.cond_stage_forward = cond_stage_forward # if self.cond_stage_forward is None: diff --git a/flagai/model/mm/lm/dcn_clip.py b/flagai/model/mm/lm/dcn_clip.py index f91b2326..2a2e9748 100644 --- a/flagai/model/mm/lm/dcn_clip.py +++ b/flagai/model/mm/lm/dcn_clip.py @@ -32,7 +32,7 @@ def __init__(self, device="cuda", max_length=77, ckpt_path=None, download_path=N super().__init__() self.device = device self.max_length = max_length - + import pdb;pdb.set_trace() self.ch_clip_model = AltCLIPModel.from_pretrain(ckpt_path) print("^_^ Using the right lm model!!!") self.ch_clip_model = self.ch_clip_model.eval() From cc9bf39f54e4a25abeb571cec57dad23c6e84390 Mon Sep 17 00:00:00 2001 From: ftgreat Date: Thu, 30 Mar 2023 15:10:25 +0800 Subject: [PATCH 3/7] saved work 3.30 Signed-off-by: ftgreat --- examples/AltCLIP-m18/altclip_inference.py | 59 ++++++++++++----------- examples/AltDiffusion/generate_18m.py | 12 ++--- flagai/model/mm/AltCLIP.py | 6 ++- flagai/model/mm/AltDiffusion2.py | 1 - flagai/model/mm/lm/dcn_clip.py | 15 +++--- 5 files changed, 49 insertions(+), 44 deletions(-) diff --git a/examples/AltCLIP-m18/altclip_inference.py b/examples/AltCLIP-m18/altclip_inference.py index 788eef40..510cb37d 100644 --- a/examples/AltCLIP-m18/altclip_inference.py +++ b/examples/AltCLIP-m18/altclip_inference.py @@ -12,31 +12,34 @@ ) model = loader.get_model() -tokenizer = loader.get_tokenizer() -transform = loader.get_transform() - -model.eval() -model.to(device) -tokenizer = loader.get_tokenizer() - -def inference(): - image = Image.open("./dog.jpeg") - image = transform(image) - image = torch.tensor(image["pixel_values"]).to(device) - tokenizer_out = tokenizer(["a rat", "a dog", "a cat"], - padding=True, - truncation=True, - max_length=77, - return_tensors='pt') - - text = tokenizer_out["input_ids"].to(device) - attention_mask = tokenizer_out["attention_mask"].to(device) - with torch.no_grad(): - image_features = model.get_image_features(image) - text_features = model.get_text_features(text, attention_mask=attention_mask) - text_probs = (image_features @ text_features.T).softmax(dim=-1) - - print(text_probs.cpu().numpy()[0].tolist()) - -if __name__=="__main__": - inference() \ No newline at end of file +for name, param in model.named_parameters(): + print(name) + +# tokenizer = loader.get_tokenizer() +# transform = loader.get_transform() + +# model.eval() +# model.to(device) +# tokenizer = loader.get_tokenizer() + +# def inference(): +# image = Image.open("./examples/AltCLIP-m18//dog.jpeg") +# image = transform(image) +# image = torch.tensor(image["pixel_values"]).to(device) +# tokenizer_out = tokenizer(["a rat", "a dog", "a cat"], +# padding=True, +# truncation=True, +# max_length=77, +# return_tensors='pt') + +# text = tokenizer_out["input_ids"].to(device) +# attention_mask = tokenizer_out["attention_mask"].to(device) +# with torch.no_grad(): +# image_features = model.get_image_features(image) +# text_features = model.get_text_features(text, attention_mask=attention_mask) +# text_probs = (image_features @ text_features.T).softmax(dim=-1) + +# print(text_probs.cpu().numpy()[0].tolist()) + +# if __name__=="__main__": +# inference() \ No newline at end of file diff --git a/examples/AltDiffusion/generate_18m.py b/examples/AltDiffusion/generate_18m.py index 766a0149..ede0c67b 100644 --- a/examples/AltDiffusion/generate_18m.py +++ b/examples/AltDiffusion/generate_18m.py @@ -20,9 +20,9 @@ # if name.startswith("cond_stage_model"): # print(name) import pdb;pdb.set_trace() -# model.eval() -# model.to(device) -# predictor = Predictor(model) -# predictor.predict_generate_images_m18( -# "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" -# ) +model.eval() +model.to(device) +predictor = Predictor(model) +predictor.predict_generate_images_m18( + "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" +) diff --git a/flagai/model/mm/AltCLIP.py b/flagai/model/mm/AltCLIP.py index 5a33ca41..f477b378 100644 --- a/flagai/model/mm/AltCLIP.py +++ b/flagai/model/mm/AltCLIP.py @@ -76,7 +76,7 @@ def __init__(self, num_layers=3, variant='invert', **kwargs): - super().__init__(text_config_dict=None, vision_config_dict=None, projection_dim=None,logit_scale_init_value=None, **kwargs) + super().__init__(projection_dim=projection_dim, logit_scale_init_value=logit_scale_init_value, **kwargs) if text_config_dict is None: text_config_dict = {} # when reload the config from local, we need name to select which class should be instanced. @@ -434,7 +434,7 @@ def forward( class AltCLIP(BaseModel): def __init__(self, config, **kwargs): - super().__init__(config, **kwargs) + super().__init__(config=config, **kwargs) @classmethod def from_pretrain(cls, @@ -446,4 +446,6 @@ def from_pretrain(cls, super().download(download_path, model_name, only_download_config=only_download_config) pretrained_model_name_or_path = os.path.join(download_path, model_name) print(pretrained_model_name_or_path) + print("Downloading AltCLIP") + return CLIPHF.from_pretrained(pretrained_model_name_or_path) diff --git a/flagai/model/mm/AltDiffusion2.py b/flagai/model/mm/AltDiffusion2.py index 3a2517b1..4c17b7e4 100644 --- a/flagai/model/mm/AltDiffusion2.py +++ b/flagai/model/mm/AltDiffusion2.py @@ -564,7 +564,6 @@ def __init__(self, else: self.register_buffer('scale_factor', torch.tensor(scale_factor)) self.instantiate_first_stage(first_stage_config) - import pdb;pdb.set_trace() self.instantiate_cond_stage(cond_stage_config) self.cond_stage_forward = cond_stage_forward # if self.cond_stage_forward is None: diff --git a/flagai/model/mm/lm/dcn_clip.py b/flagai/model/mm/lm/dcn_clip.py index 2a2e9748..cbbbd2be 100644 --- a/flagai/model/mm/lm/dcn_clip.py +++ b/flagai/model/mm/lm/dcn_clip.py @@ -1,7 +1,8 @@ import torch +import os import torch.nn as nn from transformers import AltCLIPProcessor -from flagai.model.mm.AltCLIP import AltCLIP as AltCLIPModel +from flagai.model.mm.modeling_altclip import AltCLIPModel class AbstractEncoder(nn.Module): def __init__(self): @@ -28,20 +29,20 @@ def forward(self, x): class DCNCLIP30M1024(AbstractEncoder): - def __init__(self, device="cuda", max_length=77, ckpt_path=None, download_path=None): + def __init__(self, device="cuda", max_length=77, model_name=None, download_path=None): super().__init__() self.device = device self.max_length = max_length - import pdb;pdb.set_trace() - self.ch_clip_model = AltCLIPModel.from_pretrain(ckpt_path) - print("^_^ Using the right lm model!!!") + ckpt_path = os.path.join(download_path, model_name) + self.ch_clip_model = AltCLIPModel.from_pretrained(ckpt_path=ckpt_path) + # print("^_^ Using the right lm model!!!") self.ch_clip_model = self.ch_clip_model.eval() print("Language model Loaded!!!^_^") for param in self.ch_clip_model.parameters(): param.requires_grad = False - - self.processor = AltCLIPProcessor.from_pretrain(ckpt_path) + ckpt_path = os.path.join(download_path, model_name) + self.processor = AltCLIPProcessor.from_pretrained(ckpt_path) self.tokenizer = self.processor.tokenizer self.text_encoder = self.ch_clip_model.text_model From ea8b52df1683775ededa2a0b90766217358410bd Mon Sep 17 00:00:00 2001 From: ftgreat Date: Tue, 11 Apr 2023 15:34:10 +0800 Subject: [PATCH 4/7] work saved Signed-off-by: ftgreat --- examples/AltCLIP/altclip_inference.py | 1 + examples/AltDiffusion/generate.py | 20 +- examples/AltDiffusion/generate_18m.py | 17 +- flagai/auto_model/auto_loader.py | 2 +- flagai/model/base_model.py | 10 +- flagai/model/mm/AltDiffusion.py | 1 + flagai/model/mm/AltDiffusion2.py | 32 +- flagai/model/mm/modeling_altclip.py | 1259 +++++++++---------------- flagai/model/mm/modeling_berts.py | 38 +- flagai/model/predictor/predictor.py | 3 +- load_model.py | 8 + 11 files changed, 545 insertions(+), 846 deletions(-) create mode 100644 load_model.py diff --git a/examples/AltCLIP/altclip_inference.py b/examples/AltCLIP/altclip_inference.py index 6ba53e41..e46d6ce6 100644 --- a/examples/AltCLIP/altclip_inference.py +++ b/examples/AltCLIP/altclip_inference.py @@ -1,5 +1,6 @@ import torch from PIL import Image +import sys;sys.path.append("/home/yanzhaodong/FlagAI") from flagai.auto_model.auto_loader import AutoLoader device = torch.device("cuda" if torch.cuda.is_available() else "cpu") diff --git a/examples/AltDiffusion/generate.py b/examples/AltDiffusion/generate.py index f29524bd..2b682782 100644 --- a/examples/AltDiffusion/generate.py +++ b/examples/AltDiffusion/generate.py @@ -14,16 +14,22 @@ model_dir="./checkpoints", use_fp16=False) -loader2 = AutoLoader(task_name="text2img", model_name="AltDiffusion-m9") model = loader.get_model() import pdb;pdb.set_trace() for name, param in model.named_parameters(): if name.startswith("cond_stage_model"): print(name) -# model.eval() -# model.to(device) -# predictor = Predictor(model) -# predictor.predict_generate_images( -# "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" -# ) +import pdb;pdb.set_trace() + +# import pdb;pdb.set_trace() +# for name, param in model.named_parameters(): +# if name.startswith("cond_stage_model"): +# print(name) + +model.eval() +model.to(device) +predictor = Predictor(model) +predictor.predict_generate_images( + "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" +) diff --git a/examples/AltDiffusion/generate_18m.py b/examples/AltDiffusion/generate_18m.py index ede0c67b..fc776b65 100644 --- a/examples/AltDiffusion/generate_18m.py +++ b/examples/AltDiffusion/generate_18m.py @@ -12,17 +12,28 @@ loader = AutoLoader(task_name="text2img", #contrastive learning model_name="AltDiffusion-m18", model_dir="./checkpoints", - use_fp16=False) + use_fp16=False,) # loader = AutoLoader(task_name="text2img", #contrastive learning # model_name="AltDiffusion-m18") model = loader.get_model() # for name, param in model.named_parameters(): # if name.startswith("cond_stage_model"): # print(name) -import pdb;pdb.set_trace() +# import pdb;pdb.set_trace() +for name, param in model.named_parameters(): + if 'pre_LN' in name: + print(name,param) + + model.eval() model.to(device) predictor = Predictor(model) +prompt = "Daenerys Targaryen as a mermeid with a piercing gaze wearing an enchanted bikini in an underwater magical forest, highly detailed face, realistic face, beautiful detailed eyes, fantasy art, in the style of artgerm, illustration, epic, fantasy, intricate, hyper detailed, artstation, concept art, smooth, sharp focus, ray tracing, vibrant, photorealistic" +negative_prompt = "nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, extra head, extra legs,fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry" +seed = 553124 +W = 512 +H = 512 +scale=7.5 predictor.predict_generate_images_m18( - "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" + prompt=prompt,negative_prompt=negative_prompt,seed=seed, W=W, H=H, scale=scale ) diff --git a/flagai/auto_model/auto_loader.py b/flagai/auto_model/auto_loader.py index c167a0ae..237c16d7 100644 --- a/flagai/auto_model/auto_loader.py +++ b/flagai/auto_model/auto_loader.py @@ -4,6 +4,7 @@ import importlib import os import copy +import sys;sys.path.append("/home/yanzhaodong/FlagAI") from flagai.model.file_utils import _get_model_id @@ -256,7 +257,6 @@ def is_exist_finetuned_model(self, raw_model_name, task_name): return model_name_ else : return raw_model_name - except: print("Model hub is not reachable.") return raw_model_name diff --git a/flagai/model/base_model.py b/flagai/model/base_model.py index 2399ed3a..2a51b23f 100644 --- a/flagai/model/base_model.py +++ b/flagai/model/base_model.py @@ -130,11 +130,11 @@ def load_diffusion_local(yaml_path, only_download_config=False, **kwargs): model_config.params.cond_stage_config.params.download_path = raw_download_path kwargs.update(model_config.get("params", dict())) model = cls(**kwargs) - if not only_download_config: - model = cls._load_state_dict_into_model( - model, - checkpoint_path, - ) + # if not only_download_config: + # model = cls._load_state_dict_into_model( + # model, + # checkpoint_path, + # ) return model yaml_path = os.path.join(download_path, "config.yaml") diff --git a/flagai/model/mm/AltDiffusion.py b/flagai/model/mm/AltDiffusion.py index 9a6b12cf..cfab924b 100644 --- a/flagai/model/mm/AltDiffusion.py +++ b/flagai/model/mm/AltDiffusion.py @@ -663,6 +663,7 @@ def func(c): return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + import pdb;pdb.set_trace() text["input_ids"] = text["input_ids"].clone().detach().to(device) text["attention_mask"] = text['attention_mask'].clone().detach().to(device) features = self.cond_stage_model(**text) diff --git a/flagai/model/mm/AltDiffusion2.py b/flagai/model/mm/AltDiffusion2.py index 4c17b7e4..12055432 100644 --- a/flagai/model/mm/AltDiffusion2.py +++ b/flagai/model/mm/AltDiffusion2.py @@ -566,8 +566,8 @@ def __init__(self, self.instantiate_first_stage(first_stage_config) self.instantiate_cond_stage(cond_stage_config) self.cond_stage_forward = cond_stage_forward - # if self.cond_stage_forward is None: - # self.set_cond_stage_forward() + if self.cond_stage_forward is None: + self.set_cond_stage_forward() self.clip_denoised = False self.bbox_tokenizer = None @@ -672,17 +672,37 @@ def get_first_stage_encoding(self, encoder_posterior): raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") return self.scale_factor * z + def set_cond_stage_forward(self): + def func(c): + device = next(self.cond_stage_model.parameters()).device + text = self.tokenizer(c, + truncation=True, + max_length=77, + return_length=False, + return_overflowing_tokens=False, + padding="max_length", + return_tensors="pt") + text["input_ids"] = text["input_ids"].clone().detach().to(device) + text["attention_mask"] = text['attention_mask'].clone().detach().to(device) + features = self.cond_stage_model(**text) + return features['projection_state'] + self.cond_stage_forward = func + def get_learned_conditioning(self, c): + # C will be directly returned if it is None + if c is None: + return None if self.cond_stage_forward is None: - if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): - c = self.cond_stage_model.encode(c) + if hasattr(self.cond_stage_model, 'encode') and callable( + self.cond_stage_model.encode): + c = self.cond_stage_model.encode(c, self.tokenizer) if isinstance(c, DiagonalGaussianDistribution): c = c.mode() else: c = self.cond_stage_model(c) else: - assert hasattr(self.cond_stage_model, self.cond_stage_forward) - c = getattr(self.cond_stage_model, self.cond_stage_forward)(c) + c = self.cond_stage_forward(c) + import pdb;pdb.set_trace() return c def meshgrid(self, h, w): diff --git a/flagai/model/mm/modeling_altclip.py b/flagai/model/mm/modeling_altclip.py index ffc9d5ab..5ad4ba01 100644 --- a/flagai/model/mm/modeling_altclip.py +++ b/flagai/model/mm/modeling_altclip.py @@ -31,7 +31,122 @@ from transformers.modeling_utils import PreTrainedModel from transformers.pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer from transformers.utils import ModelOutput, add_start_docstrings_to_model_forward, logging, replace_return_docstrings -from .configuration_altclip import AltCLIPConfig, AltCLIPTextConfig, AltCLIPVisionConfig +from transformers import PretrainedConfig +# from transformers.models.xlm_roberta.configuration_xlm_roberta import XLMRobertaConfig +# from transformers.models.clip.configuration_clip import CLIPConfig + + +class AltCLIPTextConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`AltCLIPTextModel`] or a [`TFAltCLIPTextModel`]. + It is used to instantiate a AltCLIP text model according to the specified arguments, defining the model + architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the + XLMRoBERTa [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) architecture. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + + Args: + vocab_size (`int`, *optional*, defaults to 30522): + Vocabulary size of the XLM-RoBERTa model. Defines the number of different tokens that can be represented by + the `inputs_ids` passed when calling [`AltCLIPTextModel`] or [`TFAltCLIPTextModel`]. + hidden_size (`int`, *optional*, defaults to 768): + Dimensionality of the encoder layers and the pooler layer. + num_hidden_layers (`int`, *optional*, defaults to 12): + Number of hidden layers in the Transformer encoder. + num_attention_heads (`int`, *optional*, defaults to 12): + Number of attention heads for each attention layer in the Transformer encoder. + intermediate_size (`int`, *optional*, defaults to 3072): + Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder. + hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`): + The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`, + `"relu"`, `"silu"` and `"gelu_new"` are supported. + hidden_dropout_prob (`float`, *optional*, defaults to 0.1): + The dropout probability for all fully connected layers in the embeddings, encoder, and pooler. + attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1): + The dropout ratio for the attention probabilities. + max_position_embeddings (`int`, *optional*, defaults to 512): + The maximum sequence length that this model might ever be used with. Typically set this to something large + just in case (e.g., 512 or 1024 or 2048). + type_vocab_size (`int`, *optional*, defaults to 2): + The vocabulary size of the `token_type_ids` passed when calling [`AltCLIPTextModel`] or + [`TFAltCLIPTextModel`]. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + layer_norm_eps (`float`, *optional*, defaults to 1e-12): + The epsilon used by the layer normalization layers. + position_embedding_type (`str`, *optional*, defaults to `"absolute"`): + Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For + positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to + [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155). + For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models + with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658). + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). Only + relevant if `config.is_decoder=True`. + classifier_dropout (`float`, *optional*): + The dropout ratio for the classification head. + + Examples: + + ```python + >>> from transformers import XLMRobertaConfig, AltCLIPTextModel + + >>> # Initializing a XLM-RoBERTa xlm-roberta-base style configuration + >>> configuration = XLMRobertaConfig() + + >>> # Initializing a model (with random weights) from the xlm-roberta-base style configuration + >>> model = AltCLIPTextModel(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + model_type = "altclip_text_model" + + def __init__( + self, + vocab_size=30522, + hidden_size=768, + num_hidden_layers=12, + num_attention_heads=12, + intermediate_size=3072, + hidden_act="gelu", + hidden_dropout_prob=0.1, + attention_probs_dropout_prob=0.1, + max_position_embeddings=512, + type_vocab_size=2, + initializer_range=0.02, + layer_norm_eps=1e-12, + pad_token_id=1, + bos_token_id=0, + eos_token_id=2, + position_embedding_type="absolute", + use_cache=True, + classifier_dropout=None, + project_dim=512, + pooler_fn="cls", + **kwargs + ): + super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs) + + self.vocab_size = vocab_size + self.hidden_size = hidden_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.hidden_act = hidden_act + self.intermediate_size = intermediate_size + self.hidden_dropout_prob = hidden_dropout_prob + self.attention_probs_dropout_prob = attention_probs_dropout_prob + self.max_position_embeddings = max_position_embeddings + self.type_vocab_size = type_vocab_size + self.initializer_range = initializer_range + self.layer_norm_eps = layer_norm_eps + self.position_embedding_type = position_embedding_type + self.use_cache = use_cache + self.classifier_dropout = classifier_dropout + self.project_dim = project_dim + self.pooler_fn = pooler_fn @dataclass @@ -753,496 +868,16 @@ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: pooled_output = self.dense(first_token_tensor) pooled_output = self.activation(pooled_output) return pooled_output - - -# Copied from transformers.models.clip.modeling_clip.CLIPAttention with CLIP->AltCLIP -class AltCLIPAttention(nn.Module): - """Multi-headed attention from 'Attention Is All You Need' paper""" - - def __init__(self, config): - super().__init__() - self.config = config - self.embed_dim = config.hidden_size - self.num_heads = config.num_attention_heads - self.head_dim = self.embed_dim // self.num_heads - if self.head_dim * self.num_heads != self.embed_dim: - raise ValueError( - f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:" - f" {self.num_heads})." - ) - self.scale = self.head_dim**-0.5 - self.dropout = config.attention_dropout - - self.k_proj = nn.Linear(self.embed_dim, self.embed_dim) - self.v_proj = nn.Linear(self.embed_dim, self.embed_dim) - self.q_proj = nn.Linear(self.embed_dim, self.embed_dim) - self.out_proj = nn.Linear(self.embed_dim, self.embed_dim) - - def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): - return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() - - def forward( - self, - hidden_states: torch.Tensor, - attention_mask: Optional[torch.Tensor] = None, - causal_attention_mask: Optional[torch.Tensor] = None, - output_attentions: Optional[bool] = False, - ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: - """Input shape: Batch x Time x Channel""" - - bsz, tgt_len, embed_dim = hidden_states.size() - - # get query proj - query_states = self.q_proj(hidden_states) * self.scale - key_states = self._shape(self.k_proj(hidden_states), -1, bsz) - value_states = self._shape(self.v_proj(hidden_states), -1, bsz) - - proj_shape = (bsz * self.num_heads, -1, self.head_dim) - query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape) - key_states = key_states.view(*proj_shape) - value_states = value_states.view(*proj_shape) - - src_len = key_states.size(1) - attn_weights = torch.bmm(query_states, key_states.transpose(1, 2)) - - if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len): - raise ValueError( - f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is" - f" {attn_weights.size()}" - ) - - # apply the causal_attention_mask first - if causal_attention_mask is not None: - if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is" - f" {causal_attention_mask.size()}" - ) - attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask - attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) - - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, tgt_len, src_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" - ) - attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask - attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) - - attn_weights = nn.functional.softmax(attn_weights, dim=-1) - - if output_attentions: - # this operation is a bit akward, but it's required to - # make sure that attn_weights keeps its gradient. - # In order to do so, attn_weights have to reshaped - # twice and have to be reused in the following - attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) - attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len) - else: - attn_weights_reshaped = None - - attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training) - - attn_output = torch.bmm(attn_probs, value_states) - - if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim): - raise ValueError( - f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is" - f" {attn_output.size()}" - ) - - attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim) - attn_output = attn_output.transpose(1, 2) - attn_output = attn_output.reshape(bsz, tgt_len, embed_dim) - - attn_output = self.out_proj(attn_output) - - return attn_output, attn_weights_reshaped - - -# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->AltCLIP -class AltCLIPMLP(nn.Module): - def __init__(self, config): - super().__init__() - self.config = config - self.activation_fn = ACT2FN[config.hidden_act] - self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size) - self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size) - - def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: - hidden_states = self.fc1(hidden_states) - hidden_states = self.activation_fn(hidden_states) - hidden_states = self.fc2(hidden_states) - return hidden_states - - -# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->AltCLIP -class AltCLIPEncoderLayer(nn.Module): - def __init__(self, config: AltCLIPConfig): - super().__init__() - self.embed_dim = config.hidden_size - self.self_attn = AltCLIPAttention(config) - self.layer_norm1 = nn.LayerNorm(self.embed_dim) - self.mlp = AltCLIPMLP(config) - self.layer_norm2 = nn.LayerNorm(self.embed_dim) - - def forward( - self, - hidden_states: torch.Tensor, - attention_mask: torch.Tensor, - causal_attention_mask: torch.Tensor, - output_attentions: Optional[bool] = False, - ) -> Tuple[torch.FloatTensor]: - """ - Args: - hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` - attention_mask (`torch.FloatTensor`): attention mask of size - `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. - `(config.encoder_attention_heads,)`. - output_attentions (`bool`, *optional*): - Whether or not to return the attentions tensors of all attention layers. See `attentions` under - returned tensors for more detail. - """ - residual = hidden_states - - hidden_states = self.layer_norm1(hidden_states) - hidden_states, attn_weights = self.self_attn( - hidden_states=hidden_states, - attention_mask=attention_mask, - causal_attention_mask=causal_attention_mask, - output_attentions=output_attentions, - ) - hidden_states = residual + hidden_states - - residual = hidden_states - hidden_states = self.layer_norm2(hidden_states) - hidden_states = self.mlp(hidden_states) - hidden_states = residual + hidden_states - - outputs = (hidden_states,) - - if output_attentions: - outputs += (attn_weights,) - - return outputs - - -# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->AltCLIP -class AltCLIPEncoder(nn.Module): - """ - Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a - [`AltCLIPEncoderLayer`]. - - Args: - config: AltCLIPConfig - """ - - def __init__(self, config: AltCLIPConfig): - super().__init__() - self.config = config - self.layers = nn.ModuleList([AltCLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)]) - self.gradient_checkpointing = False - - def forward( - self, - inputs_embeds, - attention_mask: Optional[torch.Tensor] = None, - causal_attention_mask: Optional[torch.Tensor] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ) -> Union[Tuple, BaseModelOutput]: - r""" - Args: - inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): - Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. - This is useful if you want more control over how to convert `input_ids` indices into associated vectors - than the model's internal embedding lookup matrix. - attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): - Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - - - 1 for tokens that are **not masked**, - - 0 for tokens that are **masked**. - - [What are attention masks?](../glossary#attention-mask) - causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): - Causal mask for the text model. Mask values selected in `[0, 1]`: - - - 1 for tokens that are **not masked**, - - 0 for tokens that are **masked**. - - [What are attention masks?](../glossary#attention-mask) - output_attentions (`bool`, *optional*): - Whether or not to return the attentions tensors of all attention layers. See `attentions` under - returned tensors for more detail. - output_hidden_states (`bool`, *optional*): - Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors - for more detail. - return_dict (`bool`, *optional*): - Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. - """ - output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions - output_hidden_states = ( - output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states - ) - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - encoder_states = () if output_hidden_states else None - all_attentions = () if output_attentions else None - - hidden_states = inputs_embeds - for idx, encoder_layer in enumerate(self.layers): - if output_hidden_states: - encoder_states = encoder_states + (hidden_states,) - if self.gradient_checkpointing and self.training: - - def create_custom_forward(module): - def custom_forward(*inputs): - return module(*inputs, output_attentions) - - return custom_forward - - layer_outputs = torch.utils.checkpoint.checkpoint( - create_custom_forward(encoder_layer), - hidden_states, - attention_mask, - causal_attention_mask, - ) - else: - layer_outputs = encoder_layer( - hidden_states, - attention_mask, - causal_attention_mask, - output_attentions=output_attentions, - ) - - hidden_states = layer_outputs[0] - - if output_attentions: - all_attentions = all_attentions + (layer_outputs[1],) - - if output_hidden_states: - encoder_states = encoder_states + (hidden_states,) - - if not return_dict: - return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) - return BaseModelOutput( - last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions - ) - - -# Copied from transformers.models.clip.modeling_clip.CLIPVisionEmbeddings with CLIP->AltCLIP -class AltCLIPVisionEmbeddings(nn.Module): - def __init__(self, config: AltCLIPVisionConfig): - super().__init__() - self.config = config - self.embed_dim = config.hidden_size - self.image_size = config.image_size - self.patch_size = config.patch_size - - self.class_embedding = nn.Parameter(torch.randn(self.embed_dim)) - - self.patch_embedding = nn.Conv2d( - in_channels=config.num_channels, - out_channels=self.embed_dim, - kernel_size=self.patch_size, - stride=self.patch_size, - bias=False, - ) - - self.num_patches = (self.image_size // self.patch_size) ** 2 - self.num_positions = self.num_patches + 1 - self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim) - self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1))) - - def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor: - batch_size = pixel_values.shape[0] - patch_embeds = self.patch_embedding(pixel_values) # shape = [*, width, grid, grid] - patch_embeds = patch_embeds.flatten(2).transpose(1, 2) - - class_embeds = self.class_embedding.expand(batch_size, 1, -1) - embeddings = torch.cat([class_embeds, patch_embeds], dim=1) - embeddings = embeddings + self.position_embedding(self.position_ids) - return embeddings - - -class AltCLIPPreTrainedModel(PreTrainedModel): - """ - An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained - models. - """ - - config_class = AltCLIPConfig - base_model_prefix = "altclip" - supports_gradient_checkpointing = True - _keys_to_ignore_on_load_missing = [r"position_ids"] - - def _init_weights(self, module): - """Initialize the weights""" - factor = self.config.initializer_factor - if isinstance(module, AltCLIPVisionEmbeddings): - factor = self.config.initializer_factor - nn.init.normal_(module.class_embedding, mean=0.0, std=module.embed_dim**-0.5 * factor) - nn.init.normal_(module.patch_embedding.weight, std=module.config.initializer_range * factor) - nn.init.normal_(module.position_embedding.weight, std=module.config.initializer_range * factor) - elif isinstance(module, AltCLIPAttention): - factor = self.config.initializer_factor - in_proj_std = (module.embed_dim**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor - out_proj_std = (module.embed_dim**-0.5) * factor - nn.init.normal_(module.q_proj.weight, std=in_proj_std) - nn.init.normal_(module.k_proj.weight, std=in_proj_std) - nn.init.normal_(module.v_proj.weight, std=in_proj_std) - nn.init.normal_(module.out_proj.weight, std=out_proj_std) - elif isinstance(module, AltCLIPMLP): - factor = self.config.initializer_factor - in_proj_std = ( - (module.config.hidden_size**-0.5) * ((2 * module.config.num_hidden_layers) ** -0.5) * factor - ) - fc_std = (2 * module.config.hidden_size) ** -0.5 * factor - nn.init.normal_(module.fc1.weight, std=fc_std) - nn.init.normal_(module.fc2.weight, std=in_proj_std) - elif isinstance(module, AltCLIPModel): - nn.init.normal_( - module.text_projection.weight, - std=module.text_embed_dim**-0.5 * self.config.initializer_factor, - ) - nn.init.normal_( - module.visual_projection.weight, - std=module.vision_embed_dim**-0.5 * self.config.initializer_factor, - ) - elif isinstance(module, nn.LayerNorm): - module.bias.data.zero_() - module.weight.data.fill_(1.0) - elif isinstance(module, nn.Linear): - module.weight.data.normal_(mean=0.0, std=self.config.initializer_factor) - if module.bias is not None: - module.bias.data.zero_() - elif isinstance(module, nn.Embedding): - module.weight.data.normal_(mean=0.0, std=self.config.initializer_factor) - if module.padding_idx is not None: - module.weight.data[module.padding_idx].zero_() - - def _set_gradient_checkpointing(self, module, value=False): - if isinstance(module, AltCLIPEncoder): - module.gradient_checkpointing = value - if isinstance(module, AltRobertaEncoder): - module.gradient_checkpointing = value - - -# Copied from transformers.models.clip.modeling_clip.CLIPVisionTransformer with CLIPVisionTransformer->AltCLIPVisionTransformer,CLIPVisionConfig->AltCLIPVisionConfig,CLIPVisionEmbeddings->AltCLIPVisionEmbeddings,CLIPEncoder->AltCLIPEncoder,CLIP_VISION_INPUTS_DOCSTRING->ALTCLIP_VISION_INPUTS_DOCSTRING -class AltCLIPVisionTransformer(nn.Module): - def __init__(self, config: AltCLIPVisionConfig): - super().__init__() - self.config = config - embed_dim = config.hidden_size - - self.embeddings = AltCLIPVisionEmbeddings(config) - self.pre_layrnorm = nn.LayerNorm(embed_dim) - self.encoder = AltCLIPEncoder(config) - self.post_layernorm = nn.LayerNorm(embed_dim) - - @add_start_docstrings_to_model_forward(ALTCLIP_VISION_INPUTS_DOCSTRING) - @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=AltCLIPVisionConfig) - def forward( - self, - pixel_values: Optional[torch.FloatTensor] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ) -> Union[Tuple, BaseModelOutputWithPooling]: - r""" - Returns: - - """ - output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions - output_hidden_states = ( - output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states - ) - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - if pixel_values is None: - raise ValueError("You have to specify pixel_values") - - hidden_states = self.embeddings(pixel_values) - hidden_states = self.pre_layrnorm(hidden_states) - - encoder_outputs = self.encoder( - inputs_embeds=hidden_states, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - return_dict=return_dict, - ) - - last_hidden_state = encoder_outputs[0] - pooled_output = last_hidden_state[:, 0, :] - pooled_output = self.post_layernorm(pooled_output) - - if not return_dict: - return (last_hidden_state, pooled_output) + encoder_outputs[1:] - - return BaseModelOutputWithPooling( - last_hidden_state=last_hidden_state, - pooler_output=pooled_output, - hidden_states=encoder_outputs.hidden_states, - attentions=encoder_outputs.attentions, - ) - - -class AltCLIPVisionModel(AltCLIPPreTrainedModel): - config_class = AltCLIPVisionConfig - main_input_name = "pixel_values" - - def __init__(self, config: AltCLIPVisionConfig): - super().__init__(config) - self.vision_model = AltCLIPVisionTransformer(config) - # Initialize weights and apply final processing - self.post_init() - - def get_input_embeddings(self) -> nn.Module: - return self.vision_model.embeddings.patch_embedding - - @add_start_docstrings_to_model_forward(ALTCLIP_VISION_INPUTS_DOCSTRING) - @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=AltCLIPVisionConfig) - def forward( - self, - pixel_values: Optional[torch.FloatTensor] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ) -> Union[Tuple, BaseModelOutputWithPooling]: - r""" - Returns: - - Examples: - - ```python - >>> from PIL import Image - >>> import requests - >>> from transformers import AltCLIPProcessor, AltCLIPVisionModel - - >>> model = AltCLIPVisionModel.from_pretrained("BAAI/AltCLIP") - >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") - - >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" - >>> image = Image.open(requests.get(url, stream=True).raw) - - >>> inputs = processor(images=image, return_tensors="pt") - - >>> outputs = model(**inputs) - >>> last_hidden_state = outputs.last_hidden_state - >>> pooled_output = outputs.pooler_output # pooled CLS states - ```""" - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - return self.vision_model( - pixel_values=pixel_values, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - return_dict=return_dict, - ) - - -class AltRobertaModel(AltCLIPPreTrainedModel): +from transformers.modeling_utils import PreTrainedModel +from transformers import BertConfig + +class BertSeriesConfig(BertConfig): + def __init__(self, vocab_size=30522, hidden_size=1024, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, pad_token_id=0, position_embedding_type="absolute", use_cache=True, classifier_dropout=None,project_dim=512, pooler_fn="average",learn_encoder=False,model_type='bert',**kwargs): + super().__init__(vocab_size, hidden_size, num_hidden_layers, num_attention_heads, intermediate_size, hidden_act, hidden_dropout_prob, attention_probs_dropout_prob, max_position_embeddings, type_vocab_size, initializer_range, layer_norm_eps, pad_token_id, position_embedding_type, use_cache, classifier_dropout, **kwargs) + self.project_dim = project_dim + self.pooler_fn = pooler_fn + self.learn_encoder = learn_encoder +class AltRobertaModel(PreTrainedModel): """ The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of @@ -1258,7 +893,7 @@ class AltRobertaModel(AltCLIPPreTrainedModel): """ - config_class = AltCLIPTextConfig + config_class = BertSeriesConfig # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->AltRoberta def __init__(self, config, add_pooling_layer=True): @@ -1418,328 +1053,328 @@ def forward( ) -class AltCLIPTextModel(AltCLIPPreTrainedModel): - config_class = AltCLIPTextConfig - - def __init__(self, config): - super().__init__(config) - self.roberta = AltRobertaModel(config, add_pooling_layer=False) - self.transformation = nn.Linear(config.hidden_size, config.project_dim) - self.transformation_pre = nn.Linear(config.hidden_size, config.project_dim) - self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) - self.post_init() - - def get_input_embeddings(self) -> nn.Module: - return self.roberta.embeddings.word_embeddings - - def set_input_embeddings(self, value: nn.Embedding) -> None: - self.roberta.embeddings.word_embeddings = value - - def resize_token_embeddings(self, new_num_tokens: Optional[int] = None) -> nn.Embedding: - return super().resize_token_embeddings(new_num_tokens) - - @add_start_docstrings_to_model_forward(ALTCLIP_TEXT_INPUTS_DOCSTRING) - @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndProjection, config_class=AltCLIPTextConfig) - def forward( - self, - input_ids: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - token_type_ids: Optional[torch.Tensor] = None, - position_ids: Optional[torch.Tensor] = None, - head_mask: Optional[torch.Tensor] = None, - inputs_embeds: Optional[torch.Tensor] = None, - encoder_hidden_states: Optional[torch.Tensor] = None, - encoder_attention_mask: Optional[torch.Tensor] = None, - output_attentions: Optional[bool] = None, - return_dict: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - ): - r""" - Returns: - - Examples: - - ```python - >>> from transformers import AltCLIPProcessor, AltCLIPTextModel - - >>> model = AltCLIPTextModel.from_pretrained("BAAI/AltCLIP") - >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") - - >>> texts = ["it's a cat", "it's a dog"] - - >>> inputs = processor(text=texts, padding=True, return_tensors="pt") - - >>> outputs = model(**inputs) - >>> last_hidden_state = outputs.last_hidden_state - >>> pooled_output = outputs.pooler_output # pooled CLS states - ```""" - - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - outputs = self.roberta( - input_ids=input_ids, - attention_mask=attention_mask, - token_type_ids=token_type_ids, - position_ids=position_ids, - head_mask=head_mask, - inputs_embeds=inputs_embeds, - encoder_hidden_states=encoder_hidden_states, - encoder_attention_mask=encoder_attention_mask, - output_attentions=output_attentions, - output_hidden_states=True, - return_dict=return_dict, - ) - - # last module outputs - sequence_output = outputs[0] - - # project the last outputs - sequence_output = self.pre_LN(sequence_output) - - # pooler - projection_state = self.transformation(sequence_output) - pooler_output = projection_state[:, 0] - - sequence_output2 = outputs[1][-2] - - # project every module - sequence_output2 = self.pre_LN(sequence_output2) - - # pooler - projection_state2 = self.transformation_pre(sequence_output2) - pooler_output2 = projection_state2[:, 0] - if not return_dict: - return (projection_state, pooler_output) + outputs[2:4] - - return BaseModelOutputWithPoolingAndProjection( - last_hidden_state=projection_state, - penultimate_hidden_state=projection_state2, - pooler_output=pooler_output, - pooler_output2 = pooler_output2, - hidden_states=outputs.hidden_states, - attentions=outputs.attentions, - ) - - -class AltCLIPModel(AltCLIPPreTrainedModel): - config_class = AltCLIPConfig - - def __init__(self, config: AltCLIPConfig): - super().__init__(config) - - if not isinstance(config.vision_config, AltCLIPVisionConfig): - raise ValueError( - "config.vision_config is expected to be of type AltCLIPVisionConfig but is of type" - f" {type(config.vision_config)}." - ) - if not isinstance(config.text_config, AltCLIPTextConfig): - raise ValueError( - "config.text_config is expected to be of type AltCLIPTextConfig but is of type" - f" {type(config.text_config)}." - ) - - text_config = config.text_config - vision_config = config.vision_config - - self.projection_dim = config.projection_dim - self.text_embed_dim = text_config.project_dim - self.vision_embed_dim = vision_config.hidden_size - - self.text_model = AltCLIPTextModel(text_config) - self.vision_model = AltCLIPVisionTransformer(vision_config) - - self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False) - self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False) - self.logit_scale = nn.Parameter(torch.ones([]) * self.config.logit_scale_init_value) - - # Initialize weights and apply final processing - self.post_init() - - @add_start_docstrings_to_model_forward(ALTCLIP_TEXT_INPUTS_DOCSTRING) - def get_text_features( - self, - input_ids: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - position_ids: Optional[torch.Tensor] = None, - token_type_ids=None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ) -> torch.FloatTensor: - r""" - Returns: - text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by - applying the projection layer to the pooled output of [`AltCLIPTextModel`]. - - Examples: - - ```python - >>> from transformers import AltCLIPProcessor, AltCLIPModel - - >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") - >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") - >>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt") - >>> text_features = model.get_text_features(**inputs) - ```""" - # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components. - output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions - output_hidden_states = ( - output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states - ) - return_dict = return_dict if return_dict is not None else self.config.use_return_dict +# class AltCLIPTextModel(AltCLIPPreTrainedModel): +# config_class = AltCLIPTextConfig + +# def __init__(self, config): +# super().__init__(config) +# self.roberta = AltRobertaModel(config, add_pooling_layer=False) +# self.transformation = nn.Linear(config.hidden_size, config.project_dim) +# self.transformation_pre = nn.Linear(config.hidden_size, config.project_dim) +# self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) +# self.post_init() + +# def get_input_embeddings(self) -> nn.Module: +# return self.roberta.embeddings.word_embeddings + +# def set_input_embeddings(self, value: nn.Embedding) -> None: +# self.roberta.embeddings.word_embeddings = value + +# def resize_token_embeddings(self, new_num_tokens: Optional[int] = None) -> nn.Embedding: +# return super().resize_token_embeddings(new_num_tokens) + +# @add_start_docstrings_to_model_forward(ALTCLIP_TEXT_INPUTS_DOCSTRING) +# @replace_return_docstrings(output_type=BaseModelOutputWithPoolingAndProjection, config_class=AltCLIPTextConfig) +# def forward( +# self, +# input_ids: Optional[torch.Tensor] = None, +# attention_mask: Optional[torch.Tensor] = None, +# token_type_ids: Optional[torch.Tensor] = None, +# position_ids: Optional[torch.Tensor] = None, +# head_mask: Optional[torch.Tensor] = None, +# inputs_embeds: Optional[torch.Tensor] = None, +# encoder_hidden_states: Optional[torch.Tensor] = None, +# encoder_attention_mask: Optional[torch.Tensor] = None, +# output_attentions: Optional[bool] = None, +# return_dict: Optional[bool] = None, +# output_hidden_states: Optional[bool] = None, +# ): +# r""" +# Returns: + +# Examples: + +# ```python +# >>> from transformers import AltCLIPProcessor, AltCLIPTextModel + +# >>> model = AltCLIPTextModel.from_pretrained("BAAI/AltCLIP") +# >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") + +# >>> texts = ["it's a cat", "it's a dog"] + +# >>> inputs = processor(text=texts, padding=True, return_tensors="pt") + +# >>> outputs = model(**inputs) +# >>> last_hidden_state = outputs.last_hidden_state +# >>> pooled_output = outputs.pooler_output # pooled CLS states +# ```""" + +# return_dict = return_dict if return_dict is not None else self.config.use_return_dict + +# outputs = self.roberta( +# input_ids=input_ids, +# attention_mask=attention_mask, +# token_type_ids=token_type_ids, +# position_ids=position_ids, +# head_mask=head_mask, +# inputs_embeds=inputs_embeds, +# encoder_hidden_states=encoder_hidden_states, +# encoder_attention_mask=encoder_attention_mask, +# output_attentions=output_attentions, +# output_hidden_states=True, +# return_dict=return_dict, +# ) + +# # last module outputs +# sequence_output = outputs[0] + +# # project the last outputs +# sequence_output = self.pre_LN(sequence_output) + +# # pooler +# projection_state = self.transformation(sequence_output) +# pooler_output = projection_state[:, 0] + +# sequence_output2 = outputs[1][-2] + +# # project every module +# sequence_output2 = self.pre_LN(sequence_output2) + +# # pooler +# projection_state2 = self.transformation_pre(sequence_output2) +# pooler_output2 = projection_state2[:, 0] +# if not return_dict: +# return (projection_state, pooler_output) + outputs[2:4] + +# return BaseModelOutputWithPoolingAndProjection( +# last_hidden_state=projection_state, +# penultimate_hidden_state=projection_state2, +# pooler_output=pooler_output, +# pooler_output2 = pooler_output2, +# hidden_states=outputs.hidden_states, +# attentions=outputs.attentions, +# ) + + +# class AltCLIPModel(AltCLIPPreTrainedModel): +# config_class = AltCLIPConfig + +# def __init__(self, config: AltCLIPConfig): +# super().__init__(config) + +# if not isinstance(config.vision_config, AltCLIPVisionConfig): +# raise ValueError( +# "config.vision_config is expected to be of type AltCLIPVisionConfig but is of type" +# f" {type(config.vision_config)}." +# ) +# if not isinstance(config.text_config, AltCLIPTextConfig): +# raise ValueError( +# "config.text_config is expected to be of type AltCLIPTextConfig but is of type" +# f" {type(config.text_config)}." +# ) + +# text_config = config.text_config +# vision_config = config.vision_config + +# self.projection_dim = config.projection_dim +# self.text_embed_dim = text_config.project_dim +# self.vision_embed_dim = vision_config.hidden_size + +# self.text_model = AltCLIPTextModel(text_config) +# self.vision_model = AltCLIPVisionTransformer(vision_config) + +# self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False) +# self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False) +# self.logit_scale = nn.Parameter(torch.ones([]) * self.config.logit_scale_init_value) + +# # Initialize weights and apply final processing +# self.post_init() + +# @add_start_docstrings_to_model_forward(ALTCLIP_TEXT_INPUTS_DOCSTRING) +# def get_text_features( +# self, +# input_ids: Optional[torch.Tensor] = None, +# attention_mask: Optional[torch.Tensor] = None, +# position_ids: Optional[torch.Tensor] = None, +# token_type_ids=None, +# output_attentions: Optional[bool] = None, +# output_hidden_states: Optional[bool] = None, +# return_dict: Optional[bool] = None, +# ) -> torch.FloatTensor: +# r""" +# Returns: +# text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by +# applying the projection layer to the pooled output of [`AltCLIPTextModel`]. + +# Examples: + +# ```python +# >>> from transformers import AltCLIPProcessor, AltCLIPModel + +# >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") +# >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") +# >>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt") +# >>> text_features = model.get_text_features(**inputs) +# ```""" +# # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components. +# output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions +# output_hidden_states = ( +# output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states +# ) +# return_dict = return_dict if return_dict is not None else self.config.use_return_dict - text_outputs = self.text_model( - input_ids=input_ids, - attention_mask=attention_mask, - position_ids=position_ids, - token_type_ids=token_type_ids, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - return_dict=return_dict, - ) +# text_outputs = self.text_model( +# input_ids=input_ids, +# attention_mask=attention_mask, +# position_ids=position_ids, +# token_type_ids=token_type_ids, +# output_attentions=output_attentions, +# output_hidden_states=output_hidden_states, +# return_dict=return_dict, +# ) - pooled_output = text_outputs[1] - text_features = self.text_projection(pooled_output) - - return text_features - - @add_start_docstrings_to_model_forward(ALTCLIP_VISION_INPUTS_DOCSTRING) - def get_image_features( - self, - pixel_values: Optional[torch.FloatTensor] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ) -> torch.FloatTensor: - r""" - Returns: - image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by - applying the projection layer to the pooled output of [`AltCLIPVisionModel`]. - - Examples: - - ```python - >>> from PIL import Image - >>> import requests - >>> from transformers import AltCLIPProcessor, AltCLIPModel - - >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") - >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") - >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" - >>> image = Image.open(requests.get(url, stream=True).raw) - >>> inputs = processor(images=image, return_tensors="pt") - >>> image_features = model.get_image_features(**inputs) - ```""" - # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components. - output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions - output_hidden_states = ( - output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states - ) - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - vision_outputs = self.vision_model( - pixel_values=pixel_values, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - return_dict=return_dict, - ) - - pooled_output = vision_outputs[1] # pooled_output - image_features = self.visual_projection(pooled_output) - - return image_features - - @add_start_docstrings_to_model_forward(ALTCLIP_INPUTS_DOCSTRING) - @replace_return_docstrings(output_type=AltCLIPOutput, config_class=AltCLIPConfig) - def forward( - self, - input_ids: Optional[torch.LongTensor] = None, - pixel_values: Optional[torch.FloatTensor] = None, - attention_mask: Optional[torch.Tensor] = None, - position_ids: Optional[torch.LongTensor] = None, - token_type_ids=None, - return_loss: Optional[bool] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ) -> Union[Tuple, AltCLIPOutput]: - r""" - Returns: - - Examples: - - ```python - >>> from PIL import Image - >>> import requests - >>> from transformers import AltCLIPProcessor, AltCLIPModel - - >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") - >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") - >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" - >>> image = Image.open(requests.get(url, stream=True).raw) - >>> inputs = processor( - ... text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True - ... ) - >>> outputs = model(**inputs) - >>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score - >>> probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities - ```""" - # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components. - output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions - output_hidden_states = ( - output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states - ) - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - text_outputs = self.text_model( - input_ids=input_ids, - attention_mask=attention_mask, - token_type_ids=token_type_ids, - position_ids=position_ids, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - return_dict=return_dict, - ) - - vision_outputs = self.vision_model( - pixel_values=pixel_values, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - return_dict=return_dict, - ) - - image_embeds = vision_outputs[1] - image_embeds = self.visual_projection(image_embeds) - - text_embeds = text_outputs[1] - text_embeds = self.text_projection(text_embeds) - - # normalized features - image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True) - text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True) - - # cosine similarity as logits - logit_scale = self.logit_scale.exp() - logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale - logits_per_image = logits_per_text.T - - loss = None - if return_loss: - loss = clip_loss(logits_per_text) - - if not return_dict: - output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs) - return ((loss,) + output) if loss is not None else output - - return AltCLIPOutput( - loss=loss, - logits_per_image=logits_per_image, - logits_per_text=logits_per_text, - text_embeds=text_embeds, - image_embeds=image_embeds, - text_model_output=text_outputs, - vision_model_output=vision_outputs, - ) +# pooled_output = text_outputs[1] +# text_features = self.text_projection(pooled_output) + +# return text_features + +# @add_start_docstrings_to_model_forward(ALTCLIP_VISION_INPUTS_DOCSTRING) +# def get_image_features( +# self, +# pixel_values: Optional[torch.FloatTensor] = None, +# output_attentions: Optional[bool] = None, +# output_hidden_states: Optional[bool] = None, +# return_dict: Optional[bool] = None, +# ) -> torch.FloatTensor: +# r""" +# Returns: +# image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by +# applying the projection layer to the pooled output of [`AltCLIPVisionModel`]. + +# Examples: + +# ```python +# >>> from PIL import Image +# >>> import requests +# >>> from transformers import AltCLIPProcessor, AltCLIPModel + +# >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") +# >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") +# >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" +# >>> image = Image.open(requests.get(url, stream=True).raw) +# >>> inputs = processor(images=image, return_tensors="pt") +# >>> image_features = model.get_image_features(**inputs) +# ```""" +# # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components. +# output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions +# output_hidden_states = ( +# output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states +# ) +# return_dict = return_dict if return_dict is not None else self.config.use_return_dict + +# vision_outputs = self.vision_model( +# pixel_values=pixel_values, +# output_attentions=output_attentions, +# output_hidden_states=output_hidden_states, +# return_dict=return_dict, +# ) + +# pooled_output = vision_outputs[1] # pooled_output +# image_features = self.visual_projection(pooled_output) + +# return image_features + +# @add_start_docstrings_to_model_forward(ALTCLIP_INPUTS_DOCSTRING) +# @replace_return_docstrings(output_type=AltCLIPOutput, config_class=AltCLIPConfig) +# def forward( +# self, +# input_ids: Optional[torch.LongTensor] = None, +# pixel_values: Optional[torch.FloatTensor] = None, +# attention_mask: Optional[torch.Tensor] = None, +# position_ids: Optional[torch.LongTensor] = None, +# token_type_ids=None, +# return_loss: Optional[bool] = None, +# output_attentions: Optional[bool] = None, +# output_hidden_states: Optional[bool] = None, +# return_dict: Optional[bool] = None, +# ) -> Union[Tuple, AltCLIPOutput]: +# r""" +# Returns: + +# Examples: + +# ```python +# >>> from PIL import Image +# >>> import requests +# >>> from transformers import AltCLIPProcessor, AltCLIPModel + +# >>> model = AltCLIPModel.from_pretrained("BAAI/AltCLIP") +# >>> processor = AltCLIPProcessor.from_pretrained("BAAI/AltCLIP") +# >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" +# >>> image = Image.open(requests.get(url, stream=True).raw) +# >>> inputs = processor( +# ... text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True +# ... ) +# >>> outputs = model(**inputs) +# >>> logits_per_image = outputs.logits_per_image # this is the image-text similarity score +# >>> probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities +# ```""" +# # Use AltCLIP model's config for some fields (if specified) instead of those of vision & text components. +# output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions +# output_hidden_states = ( +# output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states +# ) +# return_dict = return_dict if return_dict is not None else self.config.use_return_dict + +# text_outputs = self.text_model( +# input_ids=input_ids, +# attention_mask=attention_mask, +# token_type_ids=token_type_ids, +# position_ids=position_ids, +# output_attentions=output_attentions, +# output_hidden_states=output_hidden_states, +# return_dict=return_dict, +# ) + +# vision_outputs = self.vision_model( +# pixel_values=pixel_values, +# output_attentions=output_attentions, +# output_hidden_states=output_hidden_states, +# return_dict=return_dict, +# ) + +# image_embeds = vision_outputs[1] +# image_embeds = self.visual_projection(image_embeds) + +# text_embeds = text_outputs[1] +# text_embeds = self.text_projection(text_embeds) + +# # normalized features +# image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True) +# text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True) + +# # cosine similarity as logits +# logit_scale = self.logit_scale.exp() +# logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale +# logits_per_image = logits_per_text.T + +# loss = None +# if return_loss: +# loss = clip_loss(logits_per_text) + +# if not return_dict: +# output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs) +# return ((loss,) + output) if loss is not None else output + +# return AltCLIPOutput( +# loss=loss, +# logits_per_image=logits_per_image, +# logits_per_text=logits_per_text, +# text_embeds=text_embeds, +# image_embeds=image_embeds, +# text_model_output=text_outputs, +# vision_model_output=vision_outputs, + # ) # Copied from transformers.models.roberta.modeling_roberta.create_position_ids_from_input_ids diff --git a/flagai/model/mm/modeling_berts.py b/flagai/model/mm/modeling_berts.py index 44bee2c7..a5bf2857 100644 --- a/flagai/model/mm/modeling_berts.py +++ b/flagai/model/mm/modeling_berts.py @@ -4,10 +4,10 @@ from transformers.models.xlm_roberta.configuration_xlm_roberta import XLMRobertaConfig from transformers import XLMRobertaModel from typing import Optional +from flagai.model.mm.modeling_altclip import AltRobertaModel class BertSeriesConfig(BertConfig): - def __init__(self, vocab_size=30522, hidden_size=768, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, pad_token_id=0, position_embedding_type="absolute", use_cache=True, classifier_dropout=None,project_dim=512, pooler_fn="average",learn_encoder=False,model_type='bert',**kwargs): - + def __init__(self, vocab_size=30522, hidden_size=1024, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, pad_token_id=0, position_embedding_type="absolute", use_cache=True, classifier_dropout=None,project_dim=512, pooler_fn="average",learn_encoder=False,model_type='bert',**kwargs): super().__init__(vocab_size, hidden_size, num_hidden_layers, num_attention_heads, intermediate_size, hidden_act, hidden_dropout_prob, attention_probs_dropout_prob, max_position_embeddings, type_vocab_size, initializer_range, layer_norm_eps, pad_token_id, position_embedding_type, use_cache, classifier_dropout, **kwargs) self.project_dim = project_dim self.pooler_fn = pooler_fn @@ -26,7 +26,7 @@ class BertSeriesModelWithTransformation(BertPreTrainedModel): _keys_to_ignore_on_load_unexpected = [r"pooler"] _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"] config_class = BertSeriesConfig - + def __init__(self, config=None, **kargs): # modify initialization for autoloading if config is None: @@ -51,14 +51,18 @@ def __init__(self, config=None, **kargs): config.type_vocab_size= 1 config.use_cache=True config.vocab_size= 250002 - config.project_dim = 768 + config.project_dim = 1024 # Changed from 768!!!!!!! config.learn_encoder = False super().__init__(config) if config.model_type == 'bert': self.bert = BertModel(config) elif config.model_type == 'xlm-roberta': - self.roberta = XLMRobertaModel(config) + # self.roberta = XLMRobertaModel(config) + self.roberta = AltRobertaModel(config, add_pooling_layer=False) self.transformation = nn.Linear(config.hidden_size,config.project_dim) + self.transformation_pre = nn.Linear(config.hidden_size, config.project_dim) + # self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.learn_encoder = config.learn_encoder if config.learn_encoder: self.encoder_tfm = nn.Linear(config.hidden_size,config.project_dim) @@ -106,17 +110,29 @@ def forward( sequence_output = outputs[0] + # project the last outputs + sequence_output = self.pre_LN(sequence_output) + + # pooler + projection_state = self.transformation(sequence_output) + pooler_output = projection_state[:, 0] + + sequence_output2 = outputs[1][-2] + # project every module - sequence_output_ln = self.pre_LN(sequence_output) - + sequence_output2 = self.pre_LN(sequence_output2) + import pdb;pdb.set_trace() # pooler - pooler_output = self.pooler(sequence_output_ln) - pooler_output = self.transformation(pooler_output) - projection_state = self.transformation(outputs.last_hidden_state) + projection_state2 = self.transformation_pre(sequence_output2) + pooler_output2 = projection_state2[:, 0] + # if not return_dict: + # return (projection_state, pooler_output) + outputs[2:4] + return { 'pooler_output':pooler_output, - 'last_hidden_state':outputs.last_hidden_state, + 'pooler_output2':pooler_output2, + 'last_hidden_state':projection_state, 'hidden_states':outputs.hidden_states, 'attentions':outputs.attentions, 'projection_state':projection_state, diff --git a/flagai/model/predictor/predictor.py b/flagai/model/predictor/predictor.py index a7005594..40b4f948 100644 --- a/flagai/model/predictor/predictor.py +++ b/flagai/model/predictor/predictor.py @@ -373,6 +373,7 @@ def predict_generate_images_m18(self, scale: float = 7.5, from_file: str = None, seed: int = 34234, + precision='autocast', fp16: bool = False): from torchvision.utils import make_grid from pytorch_lightning import seed_everything @@ -407,7 +408,7 @@ def predict_generate_images_m18(self, device = next(self.model.parameters()).device n_rows = n_rows if n_rows > 0 else batch_size - precision_scope = autocast if opt.precision == "autocast" else nullcontext + precision_scope = autocast if precision == "autocast" else nullcontext sample_path = os.path.join(outpath, "samples") os.makedirs(sample_path, exist_ok=True) base_count = len(os.listdir(sample_path)) diff --git a/load_model.py b/load_model.py new file mode 100644 index 00000000..5a667af6 --- /dev/null +++ b/load_model.py @@ -0,0 +1,8 @@ +import sys;sys.path.append("/home/yanzhaodong/FlagAI") +import torch +from flagai.model.mm.AltDiffusion2 import LatentDiffusion + + +device = torch.device("cuda" if torch.cuda.is_available() else "cpu") +model = LatentDiffusion.from_pretrain(download_path="./checkpoints", model_name="AltDiffusion-m18",device=device) +import pdb;pdb.set_trace() \ No newline at end of file From ee5253b757a73fb35a1a7a8a8488d5d5239d1f6b Mon Sep 17 00:00:00 2001 From: ftgreat Date: Wed, 12 Apr 2023 16:12:51 +0800 Subject: [PATCH 5/7] added AltDiffusion-m18 Signed-off-by: ftgreat --- examples/AltCLIP-m18/altclip_inference.py | 61 ++++---- examples/AltCLIP/altclip_inference.py | 1 - .../generate.py} | 19 +-- examples/AltDiffusion/generate.py | 15 +- flagai/auto_model/auto_loader.py | 3 +- flagai/model/base_model.py | 10 +- flagai/model/mm/AltDiffusion.py | 1 - .../{AltDiffusion2.py => AltDiffusionM18.py} | 1 - flagai/model/mm/modeling_berts.py | 40 ++--- flagai/model/mm/modeling_berts_m18.py | 142 ++++++++++++++++++ flagai/model/predictor/predictor.py | 117 +-------------- load_model.py | 8 - 12 files changed, 195 insertions(+), 223 deletions(-) rename examples/{AltDiffusion/generate_18m.py => AltDiffusion-m18/generate.py} (68%) rename flagai/model/mm/{AltDiffusion2.py => AltDiffusionM18.py} (99%) create mode 100644 flagai/model/mm/modeling_berts_m18.py delete mode 100644 load_model.py diff --git a/examples/AltCLIP-m18/altclip_inference.py b/examples/AltCLIP-m18/altclip_inference.py index 510cb37d..7b466654 100644 --- a/examples/AltCLIP-m18/altclip_inference.py +++ b/examples/AltCLIP-m18/altclip_inference.py @@ -1,5 +1,4 @@ import torch -import sys;sys.path.append("/home/yanzhaodong/FlagAI") from PIL import Image from flagai.auto_model.auto_loader import AutoLoader @@ -12,34 +11,32 @@ ) model = loader.get_model() -for name, param in model.named_parameters(): - print(name) - -# tokenizer = loader.get_tokenizer() -# transform = loader.get_transform() - -# model.eval() -# model.to(device) -# tokenizer = loader.get_tokenizer() - -# def inference(): -# image = Image.open("./examples/AltCLIP-m18//dog.jpeg") -# image = transform(image) -# image = torch.tensor(image["pixel_values"]).to(device) -# tokenizer_out = tokenizer(["a rat", "a dog", "a cat"], -# padding=True, -# truncation=True, -# max_length=77, -# return_tensors='pt') - -# text = tokenizer_out["input_ids"].to(device) -# attention_mask = tokenizer_out["attention_mask"].to(device) -# with torch.no_grad(): -# image_features = model.get_image_features(image) -# text_features = model.get_text_features(text, attention_mask=attention_mask) -# text_probs = (image_features @ text_features.T).softmax(dim=-1) - -# print(text_probs.cpu().numpy()[0].tolist()) - -# if __name__=="__main__": -# inference() \ No newline at end of file + +tokenizer = loader.get_tokenizer() +transform = loader.get_transform() + +model.eval() +model.to(device) +tokenizer = loader.get_tokenizer() + +def inference(): + image = Image.open("./examples/AltCLIP-m18//dog.jpeg") + image = transform(image) + image = torch.tensor(image["pixel_values"]).to(device) + tokenizer_out = tokenizer(["a rat", "a dog", "a cat"], + padding=True, + truncation=True, + max_length=77, + return_tensors='pt') + + text = tokenizer_out["input_ids"].to(device) + attention_mask = tokenizer_out["attention_mask"].to(device) + with torch.no_grad(): + image_features = model.get_image_features(image) + text_features = model.get_text_features(text, attention_mask=attention_mask) + text_probs = (image_features @ text_features.T).softmax(dim=-1) + + print(text_probs.cpu().numpy()[0].tolist()) + +if __name__=="__main__": + inference() \ No newline at end of file diff --git a/examples/AltCLIP/altclip_inference.py b/examples/AltCLIP/altclip_inference.py index e46d6ce6..6ba53e41 100644 --- a/examples/AltCLIP/altclip_inference.py +++ b/examples/AltCLIP/altclip_inference.py @@ -1,6 +1,5 @@ import torch from PIL import Image -import sys;sys.path.append("/home/yanzhaodong/FlagAI") from flagai.auto_model.auto_loader import AutoLoader device = torch.device("cuda" if torch.cuda.is_available() else "cpu") diff --git a/examples/AltDiffusion/generate_18m.py b/examples/AltDiffusion-m18/generate.py similarity index 68% rename from examples/AltDiffusion/generate_18m.py rename to examples/AltDiffusion-m18/generate.py index fc776b65..345a4fdb 100644 --- a/examples/AltDiffusion/generate_18m.py +++ b/examples/AltDiffusion-m18/generate.py @@ -1,7 +1,6 @@ # Copyright © 2022 BAAI. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License") -import sys;sys.path.append("/home/yanzhaodong/FlagAI") import torch from flagai.auto_model.auto_loader import AutoLoader from flagai.model.predictor.predictor import Predictor @@ -13,27 +12,13 @@ model_name="AltDiffusion-m18", model_dir="./checkpoints", use_fp16=False,) -# loader = AutoLoader(task_name="text2img", #contrastive learning -# model_name="AltDiffusion-m18") model = loader.get_model() -# for name, param in model.named_parameters(): -# if name.startswith("cond_stage_model"): -# print(name) -# import pdb;pdb.set_trace() -for name, param in model.named_parameters(): - if 'pre_LN' in name: - print(name,param) - - model.eval() model.to(device) predictor = Predictor(model) prompt = "Daenerys Targaryen as a mermeid with a piercing gaze wearing an enchanted bikini in an underwater magical forest, highly detailed face, realistic face, beautiful detailed eyes, fantasy art, in the style of artgerm, illustration, epic, fantasy, intricate, hyper detailed, artstation, concept art, smooth, sharp focus, ray tracing, vibrant, photorealistic" negative_prompt = "nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, extra head, extra legs,fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry" seed = 553124 -W = 512 -H = 512 -scale=7.5 -predictor.predict_generate_images_m18( - prompt=prompt,negative_prompt=negative_prompt,seed=seed, W=W, H=H, scale=scale +predictor.predict_generate_images( + prompt=prompt,negative_prompt=negative_prompt,seed=seed ) diff --git a/examples/AltDiffusion/generate.py b/examples/AltDiffusion/generate.py index 2b682782..0c2b7bd4 100644 --- a/examples/AltDiffusion/generate.py +++ b/examples/AltDiffusion/generate.py @@ -1,7 +1,6 @@ # Copyright © 2022 BAAI. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License") -import sys;sys.path.append("/home/yanzhaodong/FlagAI") import torch from flagai.auto_model.auto_loader import AutoLoader from flagai.model.predictor.predictor import Predictor @@ -14,22 +13,10 @@ model_dir="./checkpoints", use_fp16=False) - model = loader.get_model() -import pdb;pdb.set_trace() -for name, param in model.named_parameters(): - if name.startswith("cond_stage_model"): - print(name) -import pdb;pdb.set_trace() - -# import pdb;pdb.set_trace() -# for name, param in model.named_parameters(): -# if name.startswith("cond_stage_model"): -# print(name) - model.eval() model.to(device) predictor = Predictor(model) predictor.predict_generate_images( "Anime portrait of natalie portman as an anime girl by stanley artgerm lau, wlop, rossdraws, james jean, andrei riabovitchev, marc simonetti, and sakimichan, trending on artstation" -) +) \ No newline at end of file diff --git a/flagai/auto_model/auto_loader.py b/flagai/auto_model/auto_loader.py index 237c16d7..33ef839d 100644 --- a/flagai/auto_model/auto_loader.py +++ b/flagai/auto_model/auto_loader.py @@ -4,7 +4,6 @@ import importlib import os import copy -import sys;sys.path.append("/home/yanzhaodong/FlagAI") from flagai.model.file_utils import _get_model_id @@ -66,7 +65,7 @@ def __getattr__(self, name): "cpm3_lm": ("flagai.model.cpm3_model", "CPM3"), "cpm3_train": ("flagai.model.cpm3_train_model", "CPM3"), "diffusion_text2img": ("flagai.model.mm.AltDiffusion", "LatentDiffusion"), - "diffusion2_text2img": ("flagai.model.mm.AltDiffusion2", "LatentDiffusion"), + "diffusion2_text2img": ("flagai.model.mm.AltDiffusionM18", "LatentDiffusion"), "altclip_txt_img_matching": ("flagai.model.mm.AltCLIP", "AltCLIP"), "evaclip_txt_img_matching": ("flagai.model.mm.eva_clip_model", "EVA_CLIP"), } diff --git a/flagai/model/base_model.py b/flagai/model/base_model.py index 2a51b23f..2399ed3a 100644 --- a/flagai/model/base_model.py +++ b/flagai/model/base_model.py @@ -130,11 +130,11 @@ def load_diffusion_local(yaml_path, only_download_config=False, **kwargs): model_config.params.cond_stage_config.params.download_path = raw_download_path kwargs.update(model_config.get("params", dict())) model = cls(**kwargs) - # if not only_download_config: - # model = cls._load_state_dict_into_model( - # model, - # checkpoint_path, - # ) + if not only_download_config: + model = cls._load_state_dict_into_model( + model, + checkpoint_path, + ) return model yaml_path = os.path.join(download_path, "config.yaml") diff --git a/flagai/model/mm/AltDiffusion.py b/flagai/model/mm/AltDiffusion.py index cfab924b..9a6b12cf 100644 --- a/flagai/model/mm/AltDiffusion.py +++ b/flagai/model/mm/AltDiffusion.py @@ -663,7 +663,6 @@ def func(c): return_overflowing_tokens=False, padding="max_length", return_tensors="pt") - import pdb;pdb.set_trace() text["input_ids"] = text["input_ids"].clone().detach().to(device) text["attention_mask"] = text['attention_mask'].clone().detach().to(device) features = self.cond_stage_model(**text) diff --git a/flagai/model/mm/AltDiffusion2.py b/flagai/model/mm/AltDiffusionM18.py similarity index 99% rename from flagai/model/mm/AltDiffusion2.py rename to flagai/model/mm/AltDiffusionM18.py index 12055432..541e5504 100644 --- a/flagai/model/mm/AltDiffusion2.py +++ b/flagai/model/mm/AltDiffusionM18.py @@ -702,7 +702,6 @@ def get_learned_conditioning(self, c): c = self.cond_stage_model(c) else: c = self.cond_stage_forward(c) - import pdb;pdb.set_trace() return c def meshgrid(self, h, w): diff --git a/flagai/model/mm/modeling_berts.py b/flagai/model/mm/modeling_berts.py index a5bf2857..dcf41462 100644 --- a/flagai/model/mm/modeling_berts.py +++ b/flagai/model/mm/modeling_berts.py @@ -4,10 +4,10 @@ from transformers.models.xlm_roberta.configuration_xlm_roberta import XLMRobertaConfig from transformers import XLMRobertaModel from typing import Optional -from flagai.model.mm.modeling_altclip import AltRobertaModel class BertSeriesConfig(BertConfig): - def __init__(self, vocab_size=30522, hidden_size=1024, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, pad_token_id=0, position_embedding_type="absolute", use_cache=True, classifier_dropout=None,project_dim=512, pooler_fn="average",learn_encoder=False,model_type='bert',**kwargs): + def __init__(self, vocab_size=30522, hidden_size=768, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, pad_token_id=0, position_embedding_type="absolute", use_cache=True, classifier_dropout=None,project_dim=512, pooler_fn="average",learn_encoder=False,model_type='bert',**kwargs): + super().__init__(vocab_size, hidden_size, num_hidden_layers, num_attention_heads, intermediate_size, hidden_act, hidden_dropout_prob, attention_probs_dropout_prob, max_position_embeddings, type_vocab_size, initializer_range, layer_norm_eps, pad_token_id, position_embedding_type, use_cache, classifier_dropout, **kwargs) self.project_dim = project_dim self.pooler_fn = pooler_fn @@ -26,7 +26,7 @@ class BertSeriesModelWithTransformation(BertPreTrainedModel): _keys_to_ignore_on_load_unexpected = [r"pooler"] _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"] config_class = BertSeriesConfig - + def __init__(self, config=None, **kargs): # modify initialization for autoloading if config is None: @@ -51,18 +51,14 @@ def __init__(self, config=None, **kargs): config.type_vocab_size= 1 config.use_cache=True config.vocab_size= 250002 - config.project_dim = 1024 # Changed from 768!!!!!!! + config.project_dim = 768 config.learn_encoder = False super().__init__(config) if config.model_type == 'bert': self.bert = BertModel(config) elif config.model_type == 'xlm-roberta': - # self.roberta = XLMRobertaModel(config) - self.roberta = AltRobertaModel(config, add_pooling_layer=False) + self.roberta = XLMRobertaModel(config) self.transformation = nn.Linear(config.hidden_size,config.project_dim) - self.transformation_pre = nn.Linear(config.hidden_size, config.project_dim) - # self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) - self.learn_encoder = config.learn_encoder if config.learn_encoder: self.encoder_tfm = nn.Linear(config.hidden_size,config.project_dim) @@ -110,29 +106,17 @@ def forward( sequence_output = outputs[0] - # project the last outputs - sequence_output = self.pre_LN(sequence_output) - - # pooler - projection_state = self.transformation(sequence_output) - pooler_output = projection_state[:, 0] - - sequence_output2 = outputs[1][-2] - # project every module - sequence_output2 = self.pre_LN(sequence_output2) - import pdb;pdb.set_trace() + sequence_output_ln = self.pre_LN(sequence_output) + # pooler - projection_state2 = self.transformation_pre(sequence_output2) - pooler_output2 = projection_state2[:, 0] - # if not return_dict: - # return (projection_state, pooler_output) + outputs[2:4] - + pooler_output = self.pooler(sequence_output_ln) + pooler_output = self.transformation(pooler_output) + projection_state = self.transformation(outputs.last_hidden_state) return { 'pooler_output':pooler_output, - 'pooler_output2':pooler_output2, - 'last_hidden_state':projection_state, + 'last_hidden_state':outputs.last_hidden_state, 'hidden_states':outputs.hidden_states, 'attentions':outputs.attentions, 'projection_state':projection_state, @@ -142,4 +126,4 @@ def forward( class RobertaSeriesModelWithTransformation(BertSeriesModelWithTransformation): base_model_prefix = 'roberta' - config_class= RobertaSeriesConfig + config_class= RobertaSeriesConfig \ No newline at end of file diff --git a/flagai/model/mm/modeling_berts_m18.py b/flagai/model/mm/modeling_berts_m18.py new file mode 100644 index 00000000..79a40edc --- /dev/null +++ b/flagai/model/mm/modeling_berts_m18.py @@ -0,0 +1,142 @@ +from transformers import BertPreTrainedModel,BertModel,BertConfig +import torch.nn as nn +import torch +from transformers.models.xlm_roberta.configuration_xlm_roberta import XLMRobertaConfig +from transformers import XLMRobertaModel +from typing import Optional +from flagai.model.mm.modeling_altclip import AltRobertaModel + +class BertSeriesConfig(BertConfig): + def __init__(self, vocab_size=30522, hidden_size=1024, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, hidden_act="gelu", hidden_dropout_prob=0.1, attention_probs_dropout_prob=0.1, max_position_embeddings=512, type_vocab_size=2, initializer_range=0.02, layer_norm_eps=1e-12, pad_token_id=0, position_embedding_type="absolute", use_cache=True, classifier_dropout=None,project_dim=512, pooler_fn="average",learn_encoder=False,model_type='bert',**kwargs): + super().__init__(vocab_size, hidden_size, num_hidden_layers, num_attention_heads, intermediate_size, hidden_act, hidden_dropout_prob, attention_probs_dropout_prob, max_position_embeddings, type_vocab_size, initializer_range, layer_norm_eps, pad_token_id, position_embedding_type, use_cache, classifier_dropout, **kwargs) + self.project_dim = project_dim + self.pooler_fn = pooler_fn + self.learn_encoder = learn_encoder + +class RobertaSeriesConfig(XLMRobertaConfig): + def __init__(self, pad_token_id=1, bos_token_id=0, eos_token_id=2,project_dim=512,pooler_fn='cls',learn_encoder=False, **kwargs): + super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs) + self.project_dim = project_dim + self.pooler_fn = pooler_fn + self.learn_encoder = learn_encoder + + +class BertSeriesModelWithTransformation(BertPreTrainedModel): + + _keys_to_ignore_on_load_unexpected = [r"pooler"] + _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"] + config_class = BertSeriesConfig + + def __init__(self, config=None, **kargs): + # modify initialization for autoloading + if config is None: + config = XLMRobertaConfig() + config.attention_probs_dropout_prob= 0.1 + config.bos_token_id=0 + config.eos_token_id=2 + config.hidden_act='gelu' + config.hidden_dropout_prob=0.1 + config.hidden_size=1024 + config.initializer_range=0.02 + config.intermediate_size=4096 + config.layer_norm_eps=1e-05 + config.max_position_embeddings=514 + + config.num_attention_heads=16 + config.num_hidden_layers=24 + config.output_past=True + config.pad_token_id=1 + config.position_embedding_type= "absolute" + + config.type_vocab_size= 1 + config.use_cache=True + config.vocab_size= 250002 + config.project_dim = 1024 # Changed from 768!!!!!!! + config.learn_encoder = False + super().__init__(config) + if config.model_type == 'bert': + self.bert = BertModel(config) + elif config.model_type == 'xlm-roberta': + # self.roberta = XLMRobertaModel(config) + self.roberta = AltRobertaModel(config, add_pooling_layer=False) + self.transformation = nn.Linear(config.hidden_size,config.project_dim) + self.transformation_pre = nn.Linear(config.hidden_size, config.project_dim) + # self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + self.learn_encoder = config.learn_encoder + if config.learn_encoder: + self.encoder_tfm = nn.Linear(config.hidden_size,config.project_dim) + self.pre_LN=nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + + self.pooler = lambda x: x[:,0] + self.post_init() + + + def forward( + self, + input_ids: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + token_type_ids: Optional[torch.Tensor] = None, + position_ids: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + inputs_embeds: Optional[torch.Tensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + output_attentions: Optional[bool] = None, + return_dict: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + ) : + r""" + """ + + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + + outputs = self.base_model( + input_ids=input_ids, + attention_mask=attention_mask, + token_type_ids=token_type_ids, + position_ids=position_ids, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + output_attentions=output_attentions, + output_hidden_states=True, + return_dict=return_dict, + ) + + # last module outputs + sequence_output = outputs[0] + + + # project the last outputs + sequence_output = self.pre_LN(sequence_output) + + # pooler + projection_state = self.transformation(sequence_output) + pooler_output = projection_state[:, 0] + + sequence_output2 = outputs[1][-2] + + # project every module + sequence_output2 = self.pre_LN(sequence_output2) + # pooler + projection_state2 = self.transformation_pre(sequence_output2) + pooler_output2 = projection_state2[:, 0] + # if not return_dict: + # return (projection_state, pooler_output) + outputs[2:4] + return { + 'pooler_output':pooler_output, + 'pooler_output2':pooler_output2, + 'last_hidden_state':projection_state, + 'hidden_states':outputs.hidden_states, + 'attentions':outputs.attentions, + 'projection_state':projection_state2, + 'sequence_out': sequence_output + } + + +class RobertaSeriesModelWithTransformation(BertSeriesModelWithTransformation): + base_model_prefix = 'roberta' + config_class= RobertaSeriesConfig diff --git a/flagai/model/predictor/predictor.py b/flagai/model/predictor/predictor.py index 40b4f948..45a3fa61 100644 --- a/flagai/model/predictor/predictor.py +++ b/flagai/model/predictor/predictor.py @@ -352,116 +352,6 @@ def predict_generate_randomsample(self, import os os._exit(0) - def predict_generate_images_m18(self, - prompt: str, - negative_prompt: str, - outpath: str = "AltDiffusionOutputs", - n_samples: int = 4, - n_rows: int = 0, - skip_grid: bool = False, - skip_save: bool = False, - steps: int = 50, - n_iter: int = 1, - plms: bool = False, - dpm: bool = False, - fixed_code: bool = False, - ddim_eta: float = 0.0, - H: int = 512, - W: int = 512, - C: int = 4, - f: int = 8, - scale: float = 7.5, - from_file: str = None, - seed: int = 34234, - precision='autocast', - fp16: bool = False): - from torchvision.utils import make_grid - from pytorch_lightning import seed_everything - from flagai.model.predictor.utils import chunk, check_safety, get_safety_checker - safety_checker, safety_feature_extractor = get_safety_checker() - """ - Args: - prompt: the prompt text - out_path: the output path - n_samples: how many images to be generated - skip_grid: not to grid images - skip_save: do not save images - ddim_step: number of steps in ddim model - n_iter: number of iterations - plms: use PLMSSampler - fixed_code: sampled from a initial start code - seed: Random seed - H: height of image - W: width of image - C: channels of images, 4 for colored images - """ - if plms: - sampler = PLMSSampler(self.model) - elif dpm: - sampler = DPMSolverSampler(self.model) - else: - sampler = DDIMSampler(self.model) - - seed_everything(seed) - assert "diffusion" in self.class_name.lower() - batch_size = n_samples - device = next(self.model.parameters()).device - n_rows = n_rows if n_rows > 0 else batch_size - - precision_scope = autocast if precision == "autocast" else nullcontext - sample_path = os.path.join(outpath, "samples") - os.makedirs(sample_path, exist_ok=True) - base_count = len(os.listdir(sample_path)) - grid_count = len(os.listdir(outpath)) - 1 - - start_code = None - if fixed_code: - start_code = torch.randn([n_samples, C, H // f, W // f], - device=device) - - - with torch.no_grad(), \ - precision_scope(True), \ - self.model.ema_scope(): - all_samples = list() - prompts = [batch_size * [prompt]] - for prompts in tqdm(prompts, desc="data"): - uc = None - if scale != 1.0: - uc = self.model.get_learned_conditioning(batch_size * [negative_prompt]) - if isinstance(prompts, tuple): - prompts = list(prompts) - c = self.model.get_learned_conditioning(prompts) - shape = [C, H // f, W // f] - samples, _ = sampler.sample(S=steps, - conditioning=c, - batch_size=n_samples, - shape=shape, - verbose=False, - unconditional_guidance_scale=scale, - unconditional_conditioning=uc, - eta=ddim_eta, - x_T=start_code) - - x_samples = self.model.decode_first_stage(samples) - x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0) - - for x_sample in x_samples: - x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c') - img = Image.fromarray(x_sample.astype(np.uint8)) - - all_samples.append(x_samples) - - # additionally, save as grid - grid = torch.stack(all_samples, 0) - grid = rearrange(grid, 'n b c h w -> (n b) c h w') - grid = make_grid(grid, nrow=n_rows) - - # to image - grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy() - grid = Image.fromarray(grid.astype(np.uint8)) - grid.show() - grid.save('./output.png') def predict_generate_images(self, prompt: str, @@ -482,7 +372,8 @@ def predict_generate_images(self, scale: float = 7.5, from_file: str = None, seed: int = 34234, - fp16: bool = False): + fp16: bool = False, + negative_prompt=""): from torchvision.utils import make_grid from pytorch_lightning import seed_everything from flagai.model.predictor.utils import chunk, check_safety, get_safety_checker @@ -545,8 +436,7 @@ def predict_generate_images(self, for prompts in tqdm(data, desc="data"): uc = None if scale != 1.0: - uc = self.model.get_learned_conditioning( - batch_size * [""]) + uc = self.model.get_learned_conditioning(batch_size * [negative_prompt]) if isinstance(prompts, tuple): prompts = list(prompts) c = self.model.get_learned_conditioning(prompts) @@ -573,7 +463,6 @@ def predict_generate_images(self, x_checked_image_torch = torch.from_numpy( x_checked_image).permute(0, 3, 1, 2) - prompt_count = 0 if not skip_save: for x_sample in x_checked_image_torch: diff --git a/load_model.py b/load_model.py deleted file mode 100644 index 5a667af6..00000000 --- a/load_model.py +++ /dev/null @@ -1,8 +0,0 @@ -import sys;sys.path.append("/home/yanzhaodong/FlagAI") -import torch -from flagai.model.mm.AltDiffusion2 import LatentDiffusion - - -device = torch.device("cuda" if torch.cuda.is_available() else "cpu") -model = LatentDiffusion.from_pretrain(download_path="./checkpoints", model_name="AltDiffusion-m18",device=device) -import pdb;pdb.set_trace() \ No newline at end of file From 8253a8e48b803bdaf8b290083804abfe5f40ed4e Mon Sep 17 00:00:00 2001 From: ftgreat Date: Wed, 12 Apr 2023 16:38:47 +0800 Subject: [PATCH 6/7] name changed Signed-off-by: ftgreat --- examples/AltCLIP/altclip_inference.py | 2 +- flagai/auto_model/auto_loader.py | 6 +- flagai/model/base_model.py | 2 - .../{autoencoder2.py => autoencoder_m18.py} | 0 .../modules/diffusionmodules/openaimodel.py | 7 +- .../mm/modules/diffusionmodules/upscaling.py | 81 -- flagai/model/mm/modules/encoders/__init__.py | 0 flagai/model/mm/modules/encoders/modules.py | 313 ------ flagai/model/mm/openaimodel.py | 904 ------------------ 9 files changed, 8 insertions(+), 1307 deletions(-) rename flagai/model/mm/{autoencoder2.py => autoencoder_m18.py} (100%) delete mode 100644 flagai/model/mm/modules/diffusionmodules/upscaling.py delete mode 100644 flagai/model/mm/modules/encoders/__init__.py delete mode 100644 flagai/model/mm/modules/encoders/modules.py delete mode 100644 flagai/model/mm/openaimodel.py diff --git a/examples/AltCLIP/altclip_inference.py b/examples/AltCLIP/altclip_inference.py index 6ba53e41..2967a93c 100644 --- a/examples/AltCLIP/altclip_inference.py +++ b/examples/AltCLIP/altclip_inference.py @@ -19,7 +19,7 @@ tokenizer = loader.get_tokenizer() def inference(): - image = Image.open("./dog.jpeg") + image = Image.open("./examples/AltCLIP/dog.jpeg") image = transform(image) image = torch.tensor(image["pixel_values"]).to(device) tokenizer_out = tokenizer(["a rat", "a dog", "a cat"], diff --git a/flagai/auto_model/auto_loader.py b/flagai/auto_model/auto_loader.py index 33ef839d..2dbeeb97 100644 --- a/flagai/auto_model/auto_loader.py +++ b/flagai/auto_model/auto_loader.py @@ -65,7 +65,7 @@ def __getattr__(self, name): "cpm3_lm": ("flagai.model.cpm3_model", "CPM3"), "cpm3_train": ("flagai.model.cpm3_train_model", "CPM3"), "diffusion_text2img": ("flagai.model.mm.AltDiffusion", "LatentDiffusion"), - "diffusion2_text2img": ("flagai.model.mm.AltDiffusionM18", "LatentDiffusion"), + "diffusion_m18_text2img": ("flagai.model.mm.AltDiffusionM18", "LatentDiffusion"), "altclip_txt_img_matching": ("flagai.model.mm.AltCLIP", "AltCLIP"), "evaclip_txt_img_matching": ("flagai.model.mm.eva_clip_model", "EVA_CLIP"), } @@ -122,9 +122,9 @@ def __getattr__(self, name): "altdiffusion": ["flagai.model.mm.diffusion", "LatentDiffusion", "diffusion", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], "altdiffusion-m9": - ["flagai.model.mm.diffusion2", "LatentDiffusion", "diffusion", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], + ["flagai.model.mm.diffusionM18", "LatentDiffusion", "diffusion", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], "altdiffusion-m18": - ["flagai.model.mm.Altdiffusion2", "LatentDiffusion", "diffusion2", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], + ["flagai.model.mm.AltdiffusionM18", "LatentDiffusion", "diffusion_m18", "mm","flagai.model.mm.AltCLIP", "AltCLIPProcess"], "swinv1-base-patch4-window7-224": ["flagai.model.vision.swinv1", "SwinTransformer", "swinv1", "vision"], "swinv2-base-patch4-window8-256": diff --git a/flagai/model/base_model.py b/flagai/model/base_model.py index 2399ed3a..09db9834 100644 --- a/flagai/model/base_model.py +++ b/flagai/model/base_model.py @@ -228,6 +228,4 @@ def download(cls, if not file_name.endswith("bin"): _get_vocab_path(os.path.join(download_path, model_name), file_name, model_id) else : - if only_download_config: - continue _get_checkpoint_path(os.path.join(download_path, model_name), file_name, model_id) \ No newline at end of file diff --git a/flagai/model/mm/autoencoder2.py b/flagai/model/mm/autoencoder_m18.py similarity index 100% rename from flagai/model/mm/autoencoder2.py rename to flagai/model/mm/autoencoder_m18.py diff --git a/flagai/model/mm/modules/diffusionmodules/openaimodel.py b/flagai/model/mm/modules/diffusionmodules/openaimodel.py index cee042d8..b344eaa8 100644 --- a/flagai/model/mm/modules/diffusionmodules/openaimodel.py +++ b/flagai/model/mm/modules/diffusionmodules/openaimodel.py @@ -1,13 +1,14 @@ from abc import abstractmethod import math import inspect + import numpy as np import torch as th import torch.nn as nn import torch.nn.functional as F from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_ -from ldm.modules.diffusionmodules.util import ( +from flagai.model.mm.modules.diffusionmodules.util import ( checkpoint, conv_nd, linear, @@ -16,8 +17,8 @@ normalization, timestep_embedding, ) -from ldm.modules.attention import SpatialTransformer -from ldm.util import exists +from flagai.model.mm.modules.attention import SpatialTransformer +from flagai.model.mm.utils import exists # dummy replace diff --git a/flagai/model/mm/modules/diffusionmodules/upscaling.py b/flagai/model/mm/modules/diffusionmodules/upscaling.py deleted file mode 100644 index 03816662..00000000 --- a/flagai/model/mm/modules/diffusionmodules/upscaling.py +++ /dev/null @@ -1,81 +0,0 @@ -import torch -import torch.nn as nn -import numpy as np -from functools import partial - -from ldm.modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule -from ldm.util import default - - -class AbstractLowScaleModel(nn.Module): - # for concatenating a downsampled image to the latent representation - def __init__(self, noise_schedule_config=None): - super(AbstractLowScaleModel, self).__init__() - if noise_schedule_config is not None: - self.register_schedule(**noise_schedule_config) - - def register_schedule(self, beta_schedule="linear", timesteps=1000, - linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): - betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, - cosine_s=cosine_s) - alphas = 1. - betas - alphas_cumprod = np.cumprod(alphas, axis=0) - alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) - - timesteps, = betas.shape - self.num_timesteps = int(timesteps) - self.linear_start = linear_start - self.linear_end = linear_end - assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' - - to_torch = partial(torch.tensor, dtype=torch.float32) - - self.register_buffer('betas', to_torch(betas)) - self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) - self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) - - # calculations for diffusion q(x_t | x_{t-1}) and others - self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) - self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) - self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) - self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) - self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) - - def q_sample(self, x_start, t, noise=None): - noise = default(noise, lambda: torch.randn_like(x_start)) - return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + - extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) - - def forward(self, x): - return x, None - - def decode(self, x): - return x - - -class SimpleImageConcat(AbstractLowScaleModel): - # no noise level conditioning - def __init__(self): - super(SimpleImageConcat, self).__init__(noise_schedule_config=None) - self.max_noise_level = 0 - - def forward(self, x): - # fix to constant noise level - return x, torch.zeros(x.shape[0], device=x.device).long() - - -class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel): - def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False): - super().__init__(noise_schedule_config=noise_schedule_config) - self.max_noise_level = max_noise_level - - def forward(self, x, noise_level=None): - if noise_level is None: - noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long() - else: - assert isinstance(noise_level, torch.Tensor) - z = self.q_sample(x, noise_level) - return z, noise_level - - - diff --git a/flagai/model/mm/modules/encoders/__init__.py b/flagai/model/mm/modules/encoders/__init__.py deleted file mode 100644 index e69de29b..00000000 diff --git a/flagai/model/mm/modules/encoders/modules.py b/flagai/model/mm/modules/encoders/modules.py deleted file mode 100644 index 9942075e..00000000 --- a/flagai/model/mm/modules/encoders/modules.py +++ /dev/null @@ -1,313 +0,0 @@ -import torch -import torch.nn as nn -from torch.utils.checkpoint import checkpoint - -from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel - -import open_clip -from ldm.util import default, count_params - -##### 为了克服网速十分慢，手动load了一下open_clip的模型权重########## - -import math -import collections.abc -from itertools import repeat -import torch.nn.functional as F - - -def _ntuple(n): - def parse(x): - if isinstance(x, collections.abc.Iterable): - return x - return tuple(repeat(x, n)) - return parse - -to_2tuple = _ntuple(2) - -def resize_pos_embed(state_dict, model, interpolation: str = 'bicubic', antialias: bool = True): - # Rescale the grid of position embeddings when loading from state_dict - old_pos_embed = state_dict.get('visual.positional_embedding', None) - if old_pos_embed is None or not hasattr(model.visual, 'grid_size'): - return - grid_size = to_2tuple(model.visual.grid_size) - extra_tokens = 1 # FIXME detect different token configs (ie no class token, or more) - new_seq_len = grid_size[0] * grid_size[1] + extra_tokens - if new_seq_len == old_pos_embed.shape[0]: - return - - if extra_tokens: - pos_emb_tok, pos_emb_img = old_pos_embed[:extra_tokens], old_pos_embed[extra_tokens:] - else: - pos_emb_tok, pos_emb_img = None, old_pos_embed - old_grid_size = to_2tuple(int(math.sqrt(len(pos_emb_img)))) - - # logging.info('Resizing position embedding grid-size from %s to %s', old_grid_size, grid_size) - pos_emb_img = pos_emb_img.reshape(1, old_grid_size[0], old_grid_size[1], -1).permute(0, 3, 1, 2) - pos_emb_img = F.interpolate( - pos_emb_img, - size=grid_size, - mode=interpolation, - antialias=antialias, - align_corners=False, - ) - pos_emb_img = pos_emb_img.permute(0, 2, 3, 1).reshape(1, grid_size[0] * grid_size[1], -1)[0] - if pos_emb_tok is not None: - new_pos_embed = torch.cat([pos_emb_tok, pos_emb_img], dim=0) - else: - new_pos_embed = pos_emb_img - state_dict['visual.positional_embedding'] = new_pos_embed - -def convert_to_custom_text_state_dict(state_dict: dict): - if 'text_projection' in state_dict: - # old format state_dict, move text tower -> .text - new_state_dict = {} - for k, v in state_dict.items(): - if any(k.startswith(p) for p in ( - 'text_projection', - 'positional_embedding', - 'token_embedding', - 'transformer', - 'ln_final', - )): - k = 'text.' + k - new_state_dict[k] = v - return new_state_dict - return state_dict - - -def load_state_dict(checkpoint_path: str, map_location='cpu'): - checkpoint = torch.load(checkpoint_path, map_location=map_location) - if isinstance(checkpoint, dict) and 'state_dict' in checkpoint: - state_dict = checkpoint['state_dict'] - else: - state_dict = checkpoint - if next(iter(state_dict.items()))[0].startswith('module'): - state_dict = {k[7:]: v for k, v in state_dict.items()} - return state_dict - - -def load_checkpoint(model, checkpoint_path, strict=True): - state_dict = load_state_dict(checkpoint_path) - # detect old format and make compatible with new format - if 'positional_embedding' in state_dict and not hasattr(model, 'positional_embedding'): - state_dict = convert_to_custom_text_state_dict(state_dict) - resize_pos_embed(state_dict, model) - incompatible_keys = model.load_state_dict(state_dict, strict=strict) - return incompatible_keys - -###### 这部代码就先强行插入在这里####### - - - -class AbstractEncoder(nn.Module): - def __init__(self): - super().__init__() - - def encode(self, *args, **kwargs): - raise NotImplementedError - - -class IdentityEncoder(AbstractEncoder): - - def encode(self, x): - return x - - -class ClassEmbedder(nn.Module): - def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1): - super().__init__() - self.key = key - self.embedding = nn.Embedding(n_classes, embed_dim) - self.n_classes = n_classes - self.ucg_rate = ucg_rate - - def forward(self, batch, key=None, disable_dropout=False): - if key is None: - key = self.key - # this is for use in crossattn - c = batch[key][:, None] - if self.ucg_rate > 0. and not disable_dropout: - mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate) - c = mask * c + (1-mask) * torch.ones_like(c)*(self.n_classes-1) - c = c.long() - c = self.embedding(c) - return c - - def get_unconditional_conditioning(self, bs, device="cuda"): - uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) - uc = torch.ones((bs,), device=device) * uc_class - uc = {self.key: uc} - return uc - - -def disabled_train(self, mode=True): - """Overwrite model.train with this function to make sure train/eval mode - does not change anymore.""" - return self - - -class FrozenT5Embedder(AbstractEncoder): - """Uses the T5 transformer encoder for text""" - def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl - super().__init__() - self.tokenizer = T5Tokenizer.from_pretrained(version) - self.transformer = T5EncoderModel.from_pretrained(version) - self.device = device - self.max_length = max_length # TODO: typical value? - if freeze: - self.freeze() - - def freeze(self): - self.transformer = self.transformer.eval() - #self.train = disabled_train - for param in self.parameters(): - param.requires_grad = False - - def forward(self, text): - batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, - return_overflowing_tokens=False, padding="max_length", return_tensors="pt") - tokens = batch_encoding["input_ids"].to(self.device) - outputs = self.transformer(input_ids=tokens) - - z = outputs.last_hidden_state - return z - - def encode(self, text): - return self(text) - - -class FrozenCLIPEmbedder(AbstractEncoder): - """Uses the CLIP transformer encoder for text (from huggingface)""" - LAYERS = [ - "last", - "pooled", - "hidden" - ] - def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, - freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32 - super().__init__() - assert layer in self.LAYERS - self.tokenizer = CLIPTokenizer.from_pretrained(version) - self.transformer = CLIPTextModel.from_pretrained(version) - self.device = device - self.max_length = max_length - if freeze: - self.freeze() - self.layer = layer - self.layer_idx = layer_idx - if layer == "hidden": - assert layer_idx is not None - assert 0 <= abs(layer_idx) <= 12 - - def freeze(self): - self.transformer = self.transformer.eval() - #self.train = disabled_train - for param in self.parameters(): - param.requires_grad = False - - def forward(self, text): - batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, - return_overflowing_tokens=False, padding="max_length", return_tensors="pt") - tokens = batch_encoding["input_ids"].to(self.device) - outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer=="hidden") - if self.layer == "last": - z = outputs.last_hidden_state - elif self.layer == "pooled": - z = outputs.pooler_output[:, None, :] - else: - z = outputs.hidden_states[self.layer_idx] - return z - - def encode(self, text): - return self(text) - - -class FrozenOpenCLIPEmbedder(AbstractEncoder): - """ - Uses the OpenCLIP transformer encoder for text - """ - LAYERS = [ - #"pooled", - "last", - "penultimate" - ] - def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, - freeze=True, layer="penultimate"): - super().__init__() - assert layer in self.LAYERS - path = '/share/project/yfl/database/ckpt/stable_diffusion_2.0/models--laion--CLIP-ViT-H-14-laion2B-s32B-b79K/snapshots/58a1e03a7acfacbe6b95ebc24ae0394eda6a14fc/open_clip_pytorch_model.bin' - # model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version, cache_dir="/share/project/yfl/database/hub/") - # model, _, _ = open_clip.create_model_from_pretrained(arch, "/share/project/liuguang/CLIP-ViT-H-14/", device=torch.device('cpu')) - model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu')) - tmp = load_checkpoint(model=model, checkpoint_path=path) - del model.visual - self.model = model - - self.device = device - self.max_length = max_length - if freeze: - self.freeze() - self.layer = layer - if self.layer == "last": - self.layer_idx = 0 - elif self.layer == "penultimate": - self.layer_idx = 1 - else: - raise NotImplementedError() - - def freeze(self): - self.model = self.model.eval() - for param in self.parameters(): - param.requires_grad = False - - def forward(self, text): - tokens = open_clip.tokenize(text) - z = self.encode_with_transformer(tokens.to(self.device)) - return z - - def encode_with_transformer(self, text): - x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] - x = x + self.model.positional_embedding - x = x.permute(1, 0, 2) # NLD -> LND - x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) - x = x.permute(1, 0, 2) # LND -> NLD - x = self.model.ln_final(x) - return x - - def text_transformer_forward(self, x: torch.Tensor, attn_mask = None): - for i, r in enumerate(self.model.transformer.resblocks): - if i == len(self.model.transformer.resblocks) - self.layer_idx: - break - if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting(): - x = checkpoint(r, x, attn_mask) - else: - x = r(x, attn_mask=attn_mask) - return x - - def encode(self, text): - return self(text) - -class FrozenCLIPT5Encoder(AbstractEncoder): - def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda", - clip_max_length=77, t5_max_length=77): - super().__init__() - self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length) - self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) - print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder)*1.e-6:.2f} M parameters, " - f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder)*1.e-6:.2f} M params.") - - def encode(self, text): - return self(text) - - def forward(self, text): - clip_z = self.clip_encoder.encode(text) - t5_z = self.t5_encoder.encode(text) - return [clip_z, t5_z] - - - - - - - - diff --git a/flagai/model/mm/openaimodel.py b/flagai/model/mm/openaimodel.py deleted file mode 100644 index b344eaa8..00000000 --- a/flagai/model/mm/openaimodel.py +++ /dev/null @@ -1,904 +0,0 @@ -from abc import abstractmethod -import math -import inspect - -import numpy as np -import torch as th -import torch.nn as nn -import torch.nn.functional as F -from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_ - -from flagai.model.mm.modules.diffusionmodules.util import ( - checkpoint, - conv_nd, - linear, - avg_pool_nd, - zero_module, - normalization, - timestep_embedding, -) -from flagai.model.mm.modules.attention import SpatialTransformer -from flagai.model.mm.utils import exists - - -# dummy replace -def convert_module_to_f16(x): - pass - -def convert_module_to_f32(x): - pass - - -## go -class AttentionPool2d(nn.Module): - """ - Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py - """ - - def __init__( - self, - spacial_dim: int, - embed_dim: int, - num_heads_channels: int, - output_dim: int = None, - ): - super().__init__() - self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5) - self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) - self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) - self.num_heads = embed_dim // num_heads_channels - self.attention = QKVAttention(self.num_heads) - - def forward(self, x): - b, c, *_spatial = x.shape - x = x.reshape(b, c, -1) # NC(HW) - x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1) - x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1) - x = self.qkv_proj(x) - x = self.attention(x) - x = self.c_proj(x) - return x[:, :, 0] - - -class TimestepBlock(nn.Module): - """ - Any module where forward() takes timestep embeddings as a second argument. - """ - - @abstractmethod - def forward(self, x, emb): - """ - Apply the module to `x` given `emb` timestep embeddings. - """ - - -class TimestepEmbedSequential(nn.Sequential, TimestepBlock): - """ - A sequential module that passes timestep embeddings to the children that - support it as an extra input. - """ - - def forward(self, x, emb, context=None, heypernetwork=None): - for layer in self: - if isinstance(layer, TimestepBlock): - x = layer(x, emb) - elif isinstance(layer, SpatialTransformer): - x = layer(x, context) - else: - x = layer(x) - return x - - -class Upsample(nn.Module): - """ - An upsampling layer with an optional convolution. - :param channels: channels in the inputs and outputs. - :param use_conv: a bool determining if a convolution is applied. - :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then - upsampling occurs in the inner-two dimensions. - """ - - def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): - super().__init__() - self.channels = channels - self.out_channels = out_channels or channels - self.use_conv = use_conv - self.dims = dims - if use_conv: - self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) - - def forward(self, x): - assert x.shape[1] == self.channels - if self.dims == 3: - x = F.interpolate( - x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" - ) - else: - x = F.interpolate(x, scale_factor=2, mode="nearest") - if self.use_conv: - x = self.conv(x) - return x - -class TransposedUpsample(nn.Module): - 'Learned 2x upsampling without padding' - def __init__(self, channels, out_channels=None, ks=5): - super().__init__() - self.channels = channels - self.out_channels = out_channels or channels - - self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) - - def forward(self,x): - return self.up(x) - - -class Downsample(nn.Module): - """ - A downsampling layer with an optional convolution. - :param channels: channels in the inputs and outputs. - :param use_conv: a bool determining if a convolution is applied. - :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then - downsampling occurs in the inner-two dimensions. - """ - - def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1): - super().__init__() - self.channels = channels - self.out_channels = out_channels or channels - self.use_conv = use_conv - self.dims = dims - stride = 2 if dims != 3 else (1, 2, 2) - if use_conv: - self.op = conv_nd( - dims, self.channels, self.out_channels, 3, stride=stride, padding=padding - ) - else: - assert self.channels == self.out_channels - self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) - - def forward(self, x): - assert x.shape[1] == self.channels - return self.op(x) - - -class ResBlock(TimestepBlock): - """ - A residual block that can optionally change the number of channels. - :param channels: the number of input channels. - :param emb_channels: the number of timestep embedding channels. - :param dropout: the rate of dropout. - :param out_channels: if specified, the number of out channels. - :param use_conv: if True and out_channels is specified, use a spatial - convolution instead of a smaller 1x1 convolution to change the - channels in the skip connection. - :param dims: determines if the signal is 1D, 2D, or 3D. - :param use_checkpoint: if True, use gradient checkpointing on this module. - :param up: if True, use this block for upsampling. - :param down: if True, use this block for downsampling. - """ - - def __init__( - self, - channels, - emb_channels, - dropout, - out_channels=None, - use_conv=False, - use_scale_shift_norm=False, - dims=2, - use_checkpoint=False, - up=False, - down=False, - ): - super().__init__() - self.channels = channels - self.emb_channels = emb_channels - self.dropout = dropout - self.out_channels = out_channels or channels - self.use_conv = use_conv - self.use_checkpoint = use_checkpoint - self.use_scale_shift_norm = use_scale_shift_norm - - self.in_layers = nn.Sequential( - normalization(channels), - nn.SiLU(), - conv_nd(dims, channels, self.out_channels, 3, padding=1), - ) - - self.updown = up or down - - if up: - self.h_upd = Upsample(channels, False, dims) - self.x_upd = Upsample(channels, False, dims) - elif down: - self.h_upd = Downsample(channels, False, dims) - self.x_upd = Downsample(channels, False, dims) - else: - self.h_upd = self.x_upd = nn.Identity() - - self.emb_layers = nn.Sequential( - nn.SiLU(), - linear( - emb_channels, - 2 * self.out_channels if use_scale_shift_norm else self.out_channels, - ), - ) - self.out_layers = nn.Sequential( - normalization(self.out_channels), - nn.SiLU(), - nn.Dropout(p=dropout), - zero_module( - conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) - ), - ) - - if self.out_channels == channels: - self.skip_connection = nn.Identity() - elif use_conv: - self.skip_connection = conv_nd( - dims, channels, self.out_channels, 3, padding=1 - ) - else: - self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) - - def forward(self, x, emb): - """ - Apply the block to a Tensor, conditioned on a timestep embedding. - :param x: an [N x C x ...] Tensor of features. - :param emb: an [N x emb_channels] Tensor of timestep embeddings. - :return: an [N x C x ...] Tensor of outputs. - """ - return checkpoint( - self._forward, (x, emb), self.parameters(), self.use_checkpoint - ) - - - def _forward(self, x, emb): - if self.updown: - in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] - h = in_rest(x) - h = self.h_upd(h) - x = self.x_upd(x) - h = in_conv(h) - else: - h = self.in_layers(x) - emb_out = self.emb_layers(emb).type(h.dtype) - while len(emb_out.shape) < len(h.shape): - emb_out = emb_out[..., None] - if self.use_scale_shift_norm: - out_norm, out_rest = self.out_layers[0], self.out_layers[1:] - scale, shift = th.chunk(emb_out, 2, dim=1) - h = out_norm(h) * (1 + scale) + shift - h = out_rest(h) - else: - h = h + emb_out - h = self.out_layers(h) - return self.skip_connection(x) + h - - -class AttentionBlock(nn.Module): - """ - An attention block that allows spatial positions to attend to each other. - Originally ported from here, but adapted to the N-d case. - https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. - """ - - def __init__( - self, - channels, - num_heads=1, - num_head_channels=-1, - use_checkpoint=False, - use_new_attention_order=False, - ): - super().__init__() - self.channels = channels - if num_head_channels == -1: - self.num_heads = num_heads - else: - assert ( - channels % num_head_channels == 0 - ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" - self.num_heads = channels // num_head_channels - self.use_checkpoint = use_checkpoint - self.norm = normalization(channels) - self.qkv = conv_nd(1, channels, channels * 3, 1) - if use_new_attention_order: - # split qkv before split heads - self.attention = QKVAttention(self.num_heads) - else: - # split heads before split qkv - self.attention = QKVAttentionLegacy(self.num_heads) - - self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) - - def forward(self, x): - return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! - #return pt_checkpoint(self._forward, x) # pytorch - - def _forward(self, x): - b, c, *spatial = x.shape - x = x.reshape(b, c, -1) - qkv = self.qkv(self.norm(x)) - h = self.attention(qkv) - h = self.proj_out(h) - return (x + h).reshape(b, c, *spatial) - - -def count_flops_attn(model, _x, y): - """ - A counter for the `thop` package to count the operations in an - attention operation. - Meant to be used like: - macs, params = thop.profile( - model, - inputs=(inputs, timestamps), - custom_ops={QKVAttention: QKVAttention.count_flops}, - ) - """ - b, c, *spatial = y[0].shape - num_spatial = int(np.prod(spatial)) - # We perform two matmuls with the same number of ops. - # The first computes the weight matrix, the second computes - # the combination of the value vectors. - matmul_ops = 2 * b * (num_spatial ** 2) * c - model.total_ops += th.DoubleTensor([matmul_ops]) - - -class QKVAttentionLegacy(nn.Module): - """ - A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping - """ - - def __init__(self, n_heads): - super().__init__() - self.n_heads = n_heads - - def forward(self, qkv): - """ - Apply QKV attention. - :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. - :return: an [N x (H * C) x T] tensor after attention. - """ - bs, width, length = qkv.shape - assert width % (3 * self.n_heads) == 0 - ch = width // (3 * self.n_heads) - q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) - scale = 1 / math.sqrt(math.sqrt(ch)) - weight = th.einsum( - "bct,bcs->bts", q * scale, k * scale - ) # More stable with f16 than dividing afterwards - weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) - a = th.einsum("bts,bcs->bct", weight, v) - return a.reshape(bs, -1, length) - - @staticmethod - def count_flops(model, _x, y): - return count_flops_attn(model, _x, y) - - -class QKVAttention(nn.Module): - """ - A module which performs QKV attention and splits in a different order. - """ - - def __init__(self, n_heads): - super().__init__() - self.n_heads = n_heads - - def forward(self, qkv): - """ - Apply QKV attention. - :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. - :return: an [N x (H * C) x T] tensor after attention. - """ - bs, width, length = qkv.shape - assert width % (3 * self.n_heads) == 0 - ch = width // (3 * self.n_heads) - q, k, v = qkv.chunk(3, dim=1) - scale = 1 / math.sqrt(math.sqrt(ch)) - weight = th.einsum( - "bct,bcs->bts", - (q * scale).view(bs * self.n_heads, ch, length), - (k * scale).view(bs * self.n_heads, ch, length), - ) # More stable with f16 than dividing afterwards - weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) - a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) - return a.reshape(bs, -1, length) - - @staticmethod - def count_flops(model, _x, y): - return count_flops_attn(model, _x, y) - -class HypernetworkModule(nn.Module): - multiplier = 1.0 - activation_dict = { - "linear": nn.Identity, - "relu": nn.ReLU, - "leakyrelu": nn.LeakyReLU, - "elu": nn.ELU, - "swish": nn.Hardswish, - "tanh": nn.Tanh, - "sigmoid": nn.Sigmoid, - } - activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'}) - - def __init__(self, dim=1024, state_dict=None, layer_structure=[1,2,2,1], activation_func="relu", weight_init='Normal', - add_layer_norm=False, use_dropout=False, activate_output=False, last_layer_dropout=True): - super().__init__() - - assert layer_structure is not None, "layer_structure must not be None" - assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!" - assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!" - - linears = [] - for i in range(len(layer_structure) - 1): - - # Add a fully-connected layer - linears.append(nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) - - # Add an activation func except last layer - if activation_func == "linear" or activation_func is None or (i >= len(layer_structure) - 2 and not activate_output): - pass - elif activation_func in self.activation_dict: - linears.append(self.activation_dict[activation_func]()) - else: - raise RuntimeError(f'hypernetwork uses an unsupported activation function: {activation_func}') - - # Add layer normalization - if add_layer_norm: - linears.append(nn.LayerNorm(int(dim * layer_structure[i+1]))) - - # Add dropout except last layer - if use_dropout and (i < len(layer_structure) - 3 or last_layer_dropout and i < len(layer_structure) - 2): - linears.append(nn.Dropout(p=0.3)) - - self.project_in = nn.Sequential(*[nn.Linear(768, 1024), nn.LayerNorm(1024)]) - self.linear = nn.Sequential(*linears) - - if state_dict is not None: - self.fix_old_state_dict(state_dict) - self.load_state_dict(state_dict) - else: - for layer in self.linear: - if type(layer) == nn.Linear or type(layer) == nn.LayerNorm: - w, b = layer.weight.data, layer.bias.data - if weight_init == "Normal" or type(layer) == nn.LayerNorm: - normal_(w, mean=0.0, std=0.01) - normal_(b, mean=0.0, std=0) - elif weight_init == 'XavierUniform': - xavier_uniform_(w) - zeros_(b) - elif weight_init == 'XavierNormal': - xavier_normal_(w) - zeros_(b) - elif weight_init == 'KaimingUniform': - kaiming_uniform_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu') - zeros_(b) - elif weight_init == 'KaimingNormal': - kaiming_normal_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu') - zeros_(b) - else: - raise KeyError(f"Key {weight_init} is not defined as initialization!") - # self.to(devices.device) - - def fix_old_state_dict(self, state_dict): - changes = { - 'linear1.bias': 'linear.0.bias', - 'linear1.weight': 'linear.0.weight', - 'linear2.bias': 'linear.1.bias', - 'linear2.weight': 'linear.1.weight', - } - - for fr, to in changes.items(): - x = state_dict.get(fr, None) - if x is None: - continue - - del state_dict[fr] - state_dict[to] = x - - def forward(self, x): - x = self.project_in(x) - return x + self.linear(x) * self.multiplier - # return self.linear(x) * self.multiplier - - def trainables(self): - layer_structure = [] - for layer in self.linear: - if type(layer) == nn.Linear or type(layer) == nn.LayerNorm: - layer_structure += [layer.weight, layer.bias] - return layer_structure - -class UNetModel(nn.Module): - """ - The full UNet model with attention and timestep embedding. - :param in_channels: channels in the input Tensor. - :param model_channels: base channel count for the model. - :param out_channels: channels in the output Tensor. - :param num_res_blocks: number of residual blocks per downsample. - :param attention_resolutions: a collection of downsample rates at which - attention will take place. May be a set, list, or tuple. - For example, if this contains 4, then at 4x downsampling, attention - will be used. - :param dropout: the dropout probability. - :param channel_mult: channel multiplier for each level of the UNet. - :param conv_resample: if True, use learned convolutions for upsampling and - downsampling. - :param dims: determines if the signal is 1D, 2D, or 3D. - :param num_classes: if specified (as an int), then this model will be - class-conditional with `num_classes` classes. - :param use_checkpoint: use gradient checkpointing to reduce memory usage. - :param num_heads: the number of attention heads in each attention layer. - :param num_heads_channels: if specified, ignore num_heads and instead use - a fixed channel width per attention head. - :param num_heads_upsample: works with num_heads to set a different number - of heads for upsampling. Deprecated. - :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. - :param resblock_updown: use residual blocks for up/downsampling. - :param use_new_attention_order: use a different attention pattern for potentially - increased efficiency. - """ - - def __init__( - self, - image_size, - in_channels, - model_channels, - out_channels, - num_res_blocks, - attention_resolutions, - dropout=0, - channel_mult=(1, 2, 4, 8), - conv_resample=True, - dims=2, - num_classes=None, - use_checkpoint=False, - use_fp16=False, - num_heads=-1, - num_head_channels=-1, - num_heads_upsample=-1, - use_scale_shift_norm=False, - resblock_updown=False, - use_new_attention_order=False, - use_spatial_transformer=False, # custom transformer support - transformer_depth=1, # custom transformer support - context_dim=None, # custom transformer support - n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model - legacy=True, - disable_self_attentions=None, - num_attention_blocks=None, - disable_middle_self_attn=False, - use_linear_in_transformer=False, - use_hypernetwork=False - ): - super().__init__() - if use_spatial_transformer: - assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' - - if context_dim is not None: - assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' - from omegaconf.listconfig import ListConfig - if type(context_dim) == ListConfig: - context_dim = list(context_dim) - - if num_heads_upsample == -1: - num_heads_upsample = num_heads - - if num_heads == -1: - assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' - - if num_head_channels == -1: - assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' - - self.image_size = image_size - self.in_channels = in_channels - self.model_channels = model_channels - self.out_channels = out_channels - if isinstance(num_res_blocks, int): - self.num_res_blocks = len(channel_mult) * [num_res_blocks] - else: - if len(num_res_blocks) != len(channel_mult): - raise ValueError("provide num_res_blocks either as an int (globally constant) or " - "as a list/tuple (per-level) with the same length as channel_mult") - self.num_res_blocks = num_res_blocks - if disable_self_attentions is not None: - # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not - assert len(disable_self_attentions) == len(channel_mult) - if num_attention_blocks is not None: - assert len(num_attention_blocks) == len(self.num_res_blocks) - assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) - print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " - f"This option has LESS priority than attention_resolutions {attention_resolutions}, " - f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " - f"attention will still not be set.") - - self.attention_resolutions = attention_resolutions - self.dropout = dropout - self.channel_mult = channel_mult - self.conv_resample = conv_resample - self.num_classes = num_classes - self.use_checkpoint = use_checkpoint - self.dtype = th.float16 if use_fp16 else th.float32 - self.num_heads = num_heads - self.num_head_channels = num_head_channels - self.num_heads_upsample = num_heads_upsample - self.predict_codebook_ids = n_embed is not None - self.use_hypernetwork = use_hypernetwork - - if self.use_hypernetwork: - self.hypernetwork_layers = nn.Sequential(*[HypernetworkModule(), HypernetworkModule()]) - else: - self.hypernetwork = None - - time_embed_dim = model_channels * 4 - self.time_embed = nn.Sequential( - linear(model_channels, time_embed_dim), - nn.SiLU(), - linear(time_embed_dim, time_embed_dim), - ) - - if self.num_classes is not None: - if isinstance(self.num_classes, int): - self.label_emb = nn.Embedding(num_classes, time_embed_dim) - elif self.num_classes == "continuous": - print("setting up linear c_adm embedding layer") - self.label_emb = nn.Linear(1, time_embed_dim) - else: - raise ValueError() - - self.input_blocks = nn.ModuleList( - [ - TimestepEmbedSequential( - conv_nd(dims, in_channels, model_channels, 3, padding=1) - ) - ] - ) - self._feature_size = model_channels - input_block_chans = [model_channels] - ch = model_channels - ds = 1 - for level, mult in enumerate(channel_mult): - for nr in range(self.num_res_blocks[level]): - layers = [ - ResBlock( - ch, - time_embed_dim, - dropout, - out_channels=mult * model_channels, - dims=dims, - use_checkpoint=use_checkpoint, - use_scale_shift_norm=use_scale_shift_norm, - ) - ] - ch = mult * model_channels - if ds in attention_resolutions: - if num_head_channels == -1: - dim_head = ch // num_heads - else: - num_heads = ch // num_head_channels - dim_head = num_head_channels - if legacy: - #num_heads = 1 - dim_head = ch // num_heads if use_spatial_transformer else num_head_channels - if exists(disable_self_attentions): - disabled_sa = disable_self_attentions[level] - else: - disabled_sa = False - - if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: - layers.append( - AttentionBlock( - ch, - use_checkpoint=use_checkpoint, - num_heads=num_heads, - num_head_channels=dim_head, - use_new_attention_order=use_new_attention_order, - ) if not use_spatial_transformer else SpatialTransformer( - ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, - disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, - use_checkpoint=use_checkpoint - ) - ) - self.input_blocks.append(TimestepEmbedSequential(*layers)) - self._feature_size += ch - input_block_chans.append(ch) - if level != len(channel_mult) - 1: - out_ch = ch - self.input_blocks.append( - TimestepEmbedSequential( - ResBlock( - ch, - time_embed_dim, - dropout, - out_channels=out_ch, - dims=dims, - use_checkpoint=use_checkpoint, - use_scale_shift_norm=use_scale_shift_norm, - down=True, - ) - if resblock_updown - else Downsample( - ch, conv_resample, dims=dims, out_channels=out_ch - ) - ) - ) - ch = out_ch - input_block_chans.append(ch) - ds *= 2 - self._feature_size += ch - - if num_head_channels == -1: - dim_head = ch // num_heads - else: - num_heads = ch // num_head_channels - dim_head = num_head_channels - if legacy: - #num_heads = 1 - dim_head = ch // num_heads if use_spatial_transformer else num_head_channels - self.middle_block = TimestepEmbedSequential( - ResBlock( - ch, - time_embed_dim, - dropout, - dims=dims, - use_checkpoint=use_checkpoint, - use_scale_shift_norm=use_scale_shift_norm, - ), - AttentionBlock( - ch, - use_checkpoint=use_checkpoint, - num_heads=num_heads, - num_head_channels=dim_head, - use_new_attention_order=use_new_attention_order, - ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn - ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, - disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, - use_checkpoint=use_checkpoint - ), - ResBlock( - ch, - time_embed_dim, - dropout, - dims=dims, - use_checkpoint=use_checkpoint, - use_scale_shift_norm=use_scale_shift_norm, - ), - ) - self._feature_size += ch - - self.output_blocks = nn.ModuleList([]) - for level, mult in list(enumerate(channel_mult))[::-1]: - for i in range(self.num_res_blocks[level] + 1): - ich = input_block_chans.pop() - layers = [ - ResBlock( - ch + ich, - time_embed_dim, - dropout, - out_channels=model_channels * mult, - dims=dims, - use_checkpoint=use_checkpoint, - use_scale_shift_norm=use_scale_shift_norm, - ) - ] - ch = model_channels * mult - if ds in attention_resolutions: - if num_head_channels == -1: - dim_head = ch // num_heads - else: - num_heads = ch // num_head_channels - dim_head = num_head_channels - if legacy: - #num_heads = 1 - dim_head = ch // num_heads if use_spatial_transformer else num_head_channels - if exists(disable_self_attentions): - disabled_sa = disable_self_attentions[level] - else: - disabled_sa = False - - if not exists(num_attention_blocks) or i < num_attention_blocks[level]: - layers.append( - AttentionBlock( - ch, - use_checkpoint=use_checkpoint, - num_heads=num_heads_upsample, - num_head_channels=dim_head, - use_new_attention_order=use_new_attention_order, - ) if not use_spatial_transformer else SpatialTransformer( - ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, - disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, - use_checkpoint=use_checkpoint - ) - ) - if level and i == self.num_res_blocks[level]: - out_ch = ch - layers.append( - ResBlock( - ch, - time_embed_dim, - dropout, - out_channels=out_ch, - dims=dims, - use_checkpoint=use_checkpoint, - use_scale_shift_norm=use_scale_shift_norm, - up=True, - ) - if resblock_updown - else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) - ) - ds //= 2 - self.output_blocks.append(TimestepEmbedSequential(*layers)) - self._feature_size += ch - - self.out = nn.Sequential( - normalization(ch), - nn.SiLU(), - zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), - ) - if self.predict_codebook_ids: - self.id_predictor = nn.Sequential( - normalization(ch), - conv_nd(dims, model_channels, n_embed, 1), - #nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits - ) - - # hypernetwork = Hypernetwork() - # self.input_blocks.Sp.crossattention.hypernetwork = self.h - - def convert_to_fp16(self): - """ - Convert the torso of the model to float16. - """ - self.input_blocks.apply(convert_module_to_f16) - self.middle_block.apply(convert_module_to_f16) - self.output_blocks.apply(convert_module_to_f16) - - def convert_to_fp32(self): - """ - Convert the torso of the model to float32. - """ - self.input_blocks.apply(convert_module_to_f32) - self.middle_block.apply(convert_module_to_f32) - self.output_blocks.apply(convert_module_to_f32) - - def forward(self, x, timesteps=None, context=None, y=None,**kwargs): - """ - Apply the model to an input batch. - :param x: an [N x C x ...] Tensor of inputs. - :param timesteps: a 1-D batch of timesteps. - :param context: conditioning plugged in via crossattn - :param y: an [N] Tensor of labels, if class-conditional. - :return: an [N x C x ...] Tensor of outputs. - """ - assert (y is not None) == ( - self.num_classes is not None - ), "must specify y if and only if the model is class-conditional" - hs = [] - t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) - emb = self.time_embed(t_emb) - - if self.use_hypernetwork: - context_k = self.hypernetwork_layers[0](context) - context_v = self.hypernetwork_layers[1](context) - context = [context_k, context_v] - if self.num_classes is not None: - assert y.shape[0] == x.shape[0] - emb = emb + self.label_emb(y) - - h = x.type(self.dtype) - for module in self.input_blocks: - # h:(4,4,64,64) emb:(4,1280) context:(4,77,1024) - # import pdb - # pdb.set_trace() - h = module(h, emb, context) - hs.append(h) - h = self.middle_block(h, emb, context) - for module in self.output_blocks: - h = th.cat([h, hs.pop()], dim=1) - h = module(h, emb, context) - h = h.type(x.dtype) - if self.predict_codebook_ids: - return self.id_predictor(h) - else: - return self.out(h) From add30031c9200b2fa6af7aaef59097d90549a04c Mon Sep 17 00:00:00 2001 From: ftgreat Date: Thu, 13 Apr 2023 10:56:17 +0800 Subject: [PATCH 7/7] added fp16 for m18 Signed-off-by: ftgreat --- examples/AltDiffusion-m18/generate.py | 2 +- flagai/model/mm/AltDiffusionM18.py | 6 ++++-- .../model/mm/modules/diffusionmodules/openaimodel.py | 12 ++++++++---- flagai/model/predictor/predictor.py | 1 - 4 files changed, 13 insertions(+), 8 deletions(-) diff --git a/examples/AltDiffusion-m18/generate.py b/examples/AltDiffusion-m18/generate.py index 345a4fdb..af7230bc 100644 --- a/examples/AltDiffusion-m18/generate.py +++ b/examples/AltDiffusion-m18/generate.py @@ -11,7 +11,7 @@ loader = AutoLoader(task_name="text2img", #contrastive learning model_name="AltDiffusion-m18", model_dir="./checkpoints", - use_fp16=False,) + use_fp16=False) model = loader.get_model() model.eval() model.to(device) diff --git a/flagai/model/mm/AltDiffusionM18.py b/flagai/model/mm/AltDiffusionM18.py index 541e5504..acb30e95 100644 --- a/flagai/model/mm/AltDiffusionM18.py +++ b/flagai/model/mm/AltDiffusionM18.py @@ -16,6 +16,8 @@ from flagai.model.base_model import BaseModel from torch.cuda.amp import autocast as autocast import pytorch_lightning as pl +from torch.cuda.amp import autocast as autocast + __conditioning_keys__ = { 'concat': 'c_concat', @@ -889,8 +891,8 @@ def apply_model(self, x_noisy, t, cond, return_ids=False): cond = [cond] key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn' cond = {key: cond} - - x_recon = self.model(x_noisy, t, **cond) + with autocast(): + x_recon = self.model(x_noisy, t, **cond) if isinstance(x_recon, tuple) and not return_ids: return x_recon[0] diff --git a/flagai/model/mm/modules/diffusionmodules/openaimodel.py b/flagai/model/mm/modules/diffusionmodules/openaimodel.py index b344eaa8..4c85019b 100644 --- a/flagai/model/mm/modules/diffusionmodules/openaimodel.py +++ b/flagai/model/mm/modules/diffusionmodules/openaimodel.py @@ -7,6 +7,7 @@ import torch.nn as nn import torch.nn.functional as F from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_ +from torch.cuda.amp import autocast as autocast from flagai.model.mm.modules.diffusionmodules.util import ( checkpoint, @@ -81,7 +82,8 @@ class TimestepEmbedSequential(nn.Sequential, TimestepBlock): def forward(self, x, emb, context=None, heypernetwork=None): for layer in self: if isinstance(layer, TimestepBlock): - x = layer(x, emb) + with autocast(): + x = layer(x, emb) elif isinstance(layer, SpatialTransformer): x = layer(x, context) else: @@ -625,6 +627,7 @@ def __init__( self.num_heads_upsample = num_heads_upsample self.predict_codebook_ids = n_embed is not None self.use_hypernetwork = use_hypernetwork + self.use_fp16 = use_fp16 if self.use_hypernetwork: self.hypernetwork_layers = nn.Sequential(*[HypernetworkModule(), HypernetworkModule()]) @@ -876,7 +879,10 @@ def forward(self, x, timesteps=None, context=None, y=None,**kwargs): ), "must specify y if and only if the model is class-conditional" hs = [] t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) - emb = self.time_embed(t_emb) + if self.use_fp16: + emb = self.time_embed(t_emb.half()) + else: + emb = self.time_embed(t_emb) if self.use_hypernetwork: context_k = self.hypernetwork_layers[0](context) @@ -889,8 +895,6 @@ def forward(self, x, timesteps=None, context=None, y=None,**kwargs): h = x.type(self.dtype) for module in self.input_blocks: # h:(4,4,64,64) emb:(4,1280) context:(4,77,1024) - # import pdb - # pdb.set_trace() h = module(h, emb, context) hs.append(h) h = self.middle_block(h, emb, context) diff --git a/flagai/model/predictor/predictor.py b/flagai/model/predictor/predictor.py index 45a3fa61..43c1256f 100644 --- a/flagai/model/predictor/predictor.py +++ b/flagai/model/predictor/predictor.py @@ -372,7 +372,6 @@ def predict_generate_images(self, scale: float = 7.5, from_file: str = None, seed: int = 34234, - fp16: bool = False, negative_prompt=""): from torchvision.utils import make_grid from pytorch_lightning import seed_everything