main_vpo_mono.py

import os
import random

import numpy
import torch
import torch.multiprocessing as mp
import torch.nn as nn
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import wandb
from easydict import EasyDict
from loguru import logger
from torch.nn.parallel import DistributedDataParallel as DDP

from config.flags import add_tag, load_args_and_config
from engine.engine import Engine
from engine.lr_policy import WarmUpPolyLR
from engine.utils import group_weight
from models.audio.audio_network import AudioModel
from models.visual.visual_network import VisualModel
from utils import ddp_utils


def add_tag(tags, key):
    if len(tags) != 0:
        tags.append(key)
    else:
        tags = [key]
    return tags


def seed_it(seed):
    random.seed(seed)
    os.environ["PYTHONSEED"] = str(seed)
    numpy.random.seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = True
    torch.backends.cudnn.enabled = True
    torch.manual_seed(seed)


def set_group_lr(model_v, hyp_param_):
    param_lists_v = []
    # bkb: 0, 1 | cross_attn 2, 3 | seg 4:
    param_lists_v = group_weight(
        param_lists_v,
        model_v.backbone,
        norm_layer=torch.nn.BatchNorm2d,
        lr=hyp_param_.lr,
    )
    if not hyp_param_.use_baseline:
        param_lists_v.append(
            {"params": model_v.visual_projector.parameters(), "lr": hyp_param_.lr * 1}
        )
        param_lists_v.append(
            {"params": model_v.cross_att.parameters(), "lr": hyp_param_.lr * 1}
        )
    for module in model_v.segment.business_layer:
        param_lists_v = group_weight(
            param_lists_v, module, torch.nn.BatchNorm2d, hyp_param_.lr * 10.0
        )
    return param_lists_v


def main(local_rank, ngpus_per_node, hyp_param_):
    hyp_param_.local_rank = local_rank
    engine = Engine(custom_arg=hyp_param_, logger=logger)
    ddp_utils.supress_printer(hyp_param_.ddp, local_rank)

    if hyp_param_.local_rank <= 0:
        from utils.tensor_board import Tensorboard

        wandb_ = Tensorboard(hyp_param_)
    else:
        wandb_ = None

    hyp_param_.num_classes = (
        hyp_param_.vpo_num_classes if hyp_param_.use_vpo else hyp_param_.num_classes
    )

    if hyp_param_.use_baseline:
        model_v = VisualModel(
            hyp_param_.visual_backbone,
            hyp_param_.visual_backbone_pretrain_path,
            num_classes=hyp_param_.num_classes,
            seg_model=hyp_param_.seg_model,
            last_three_dilation_stride=hyp_param_.last_three_dilation_stride,
        )
        model_a = AudioModel(
            hyp_param_.audio_backbone,
            hyp_param_.audio_backbone_pretrain_path,
            out_plane=2048 if hyp_param_.visual_backbone == 50 else 512,
        )
    else:
        from models.cavp_model import CAVP

        model_v = CAVP(
            hyp_param_.visual_backbone,
            hyp_param_.visual_backbone_pretrain_path,
            num_classes=hyp_param_.num_classes,
            audio_backbone_pretrain_path=hyp_param_.audio_backbone_pretrain_path,
            visual_backbone=hyp_param_.visual_backbone,
            args=hyp_param_,
        )
        model_a = model_v.audio_backbone

    num_param = sum(p.numel() for p in model_v.parameters() if p.requires_grad)
    MODEL_PARAMS = numpy.round(num_param / 1e6, 4)
    logger.warning("Number of trainable parameters: {}M".format(MODEL_PARAMS))
    if local_rank <= 0:
        wandb_.tensor_board.config.update({"MODEL_PARAMS": MODEL_PARAMS})

    param_lists_v = set_group_lr(model_v, hyp_param_)

    optimizer_v = torch.optim.SGD(
        param_lists_v,
        lr=hyp_param_.lr,
        momentum=hyp_param_.momentum,
        weight_decay=hyp_param_.weight_decay,
    )

    optimizer_a = torch.optim.Adam(params=model_a.parameters(), lr=hyp_param_.lr)

    if hyp_param_.ddp:
        torch.cuda.set_device(hyp_param_.local_rank)
        model_v.cuda(hyp_param_.local_rank)
        visual_model = nn.SyncBatchNorm.convert_sync_batchnorm(model_v)
        model_v = DDP(
            visual_model,
            device_ids=[hyp_param_.local_rank],
            find_unused_parameters=True,
        )

        model_a.cuda(hyp_param_.local_rank)
        model_a = nn.SyncBatchNorm.convert_sync_batchnorm(model_a)
        model_a = DDP(
            model_a, device_ids=[hyp_param_.local_rank], find_unused_parameters=True
        )
    else:
        model_v = nn.DataParallel(model_v, device_ids=["cuda:0"])
        model_a = nn.DataParallel(model_a, device_ids=["cuda:0"])

    import pandas

    if hyp_param_.use_vpo:
        if hyp_param_.setup == "vpo_ss":
            df_name_ = "vpo_ss_data_stereo.csv"
        elif hyp_param_.setup == "vpo_ms":
            df_name_ = "vpo_ms_data_stereo.csv"
        elif hyp_param_.setup == "vpo_msmi":
            df_name_ = "vpo_msmi_data_stereo.csv"
        else:
            raise ValueError
        csv_path = os.path.join(hyp_param_.vpo_data_path, df_name_)
    else:
        raise ValueError

    logger.warning(f"Using <<{df_name_}>>")
    csv_ = pandas.read_csv(csv_path)
    if hyp_param_.setup == "vpo_ms" or hyp_param_.setup == "vpo_msmi":
        from dataset.vpo_mono.multi_source.av_datasets import AudioVisualDataset
        from dataset.vpo_mono.multi_source.visual.visual_dataset import \
            prepare_train_data
    else:
        from dataset.vpo_mono.single_source.av_datasets import AudioVisualDataset
        from dataset.vpo_mono.single_source.visual.visual_dataset import \
            prepare_train_data
    if hyp_param_.use_vpo:
        csv_ = prepare_train_data(csv_.copy(), hyp_param_)

    train_dataset = AudioVisualDataset(
        args=hyp_param_, mode="train", dataframe=csv_[csv_["split"] == "train"]
    )
    test_dataset = AudioVisualDataset(
        args=hyp_param_, mode="test", dataframe=csv_[csv_["split"] == "test"]
    )
    final_batch_size = hyp_param_.batch_size * hyp_param_.gpus
    lr_policy = WarmUpPolyLR(
        hyp_param_.lr,
        hyp_param_.lr_power,
        int(len(train_dataset) / final_batch_size) * hyp_param_.epochs,
        len(train_dataset) // final_batch_size * hyp_param_.warm_up_epoch,
    )
    train_sampler = (
        torch.utils.data.distributed.DistributedSampler(train_dataset)
        if hyp_param_.ddp
        else None
    )

    train_loader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=hyp_param_.batch_size,
        shuffle=(train_sampler is None),
        num_workers=hyp_param_.num_workers,
        pin_memory=True,
        sampler=train_sampler,
        drop_last=True,
    )
    test_loader = torch.utils.data.DataLoader(
        test_dataset,
        batch_size=1,
        shuffle=False,
        num_workers=hyp_param_.num_workers,
        pin_memory=True,
        drop_last=False,
    )

    from trainer.trainer_cavp_vpo_mono import CAVP_TRAINER

    trainer = CAVP_TRAINER(
        hyp_param_,
        train_loader,
        engine=engine,
        visual_tool=wandb_,
        lr_scheduler=lr_policy,
    )

    for epoch in range(0, hyp_param_.epochs):
        engine.register_state(
            dataloader=train_loader,
            model_v=model_v,
            optimizer_v=optimizer_v,
            model_a=model_a,
            optimizer_a=optimizer_a,
        )
        if hyp_param_.ddp:
            train_loader.sampler.set_epoch(epoch)
        trainer.train(model_v, model_a, optimizer_v, optimizer_a, epoch, train_loader)

        if local_rank <= 0:
            if epoch % 5 == 0 or epoch >= 50:
                trainer.validation(model_v, model_a, epoch, test_loader)
        ddp_utils.barrier(hyp_param_.ddp)

    if hyp_param_.local_rank <= 0:
        wandb_.finish()


if __name__ == "__main__":
    logger.warning("RUNNING ON MONO AUDIO")
    args, config = load_args_and_config()
    args.tags = add_tag(args.tags, "mono")
    hyp_param = EasyDict(config)
    hyp_param.update(**vars(args))
    hyp_param.lr *= hyp_param.gpus
    hyp_param.ddp = True if hyp_param.gpus > 1 else False
    hyp_param.world_size = hyp_param.gpus * hyp_param.nodes

    """ Pretrains """
    if hyp_param.seg_model == "HRNet":
        hyp_param.visual_backbone = "HRNet-W48"
    elif hyp_param.seg_model == "OCR":
        hyp_param.visual_backbone = "HRNet-W48"

    seed_it(hyp_param.seed + hyp_param.local_rank)

    if args.debug:
        logger.critical("DEBUG MODE ACTIVATED")
        hyp_param.wandb_mode = "disabled"
        hyp_param.experiment_name = "dummpy_test"
        # hyp_param.image_width = 128
        # hyp_param.image_height = 128

    logger.critical(f"SETUP: {hyp_param.setup}")
    logger.critical(f"EPOCH: {hyp_param.epochs}")
    logger.critical(f"BACKBONE: {hyp_param.visual_backbone}")
    logger.critical(f"BATCH SIZE: {hyp_param.batch_size}")
    logger.critical(f"LR: {hyp_param.lr}")
    logger.critical(f"WEIGHT DECAY: {hyp_param.weight_decay}")

    if hyp_param.ddp:
        mp.spawn(main, nprocs=hyp_param.gpus, args=(hyp_param.gpus, hyp_param))
    else:
        main(0, hyp_param.gpus, hyp_param)