ddim_solver.py

import numpy as np
import torch
from common_utils import extract_into_tensor


class DDIMSolver:
    def __init__(self, alpha_cumprods, timesteps=1000, ddim_timesteps=50):
        # DDIM sampling parameters
        step_ratio = timesteps // ddim_timesteps
        self.ddim_timesteps = (
            np.arange(1, ddim_timesteps + 1) * step_ratio
        ).round().astype(np.int64) - 1
        self.ddim_alpha_cumprods = alpha_cumprods[self.ddim_timesteps]
        self.ddim_alpha_cumprods_prev = np.asarray(
            [alpha_cumprods[0]] + alpha_cumprods[self.ddim_timesteps[:-1]].tolist()
        )
        # convert to torch tensors
        self.ddim_timesteps = torch.from_numpy(self.ddim_timesteps).long()
        self.ddim_alpha_cumprods = torch.from_numpy(self.ddim_alpha_cumprods)
        self.ddim_alpha_cumprods_prev = torch.from_numpy(self.ddim_alpha_cumprods_prev)

    def to(self, device):
        self.ddim_timesteps = self.ddim_timesteps.to(device)
        self.ddim_alpha_cumprods = self.ddim_alpha_cumprods.to(device)
        self.ddim_alpha_cumprods_prev = self.ddim_alpha_cumprods_prev.to(device)
        return self

    def ddim_step(self, pred_x0, pred_noise, timestep_index):
        alpha_cumprod_prev = extract_into_tensor(
            self.ddim_alpha_cumprods_prev, timestep_index, pred_x0.shape
        )
        dir_xt = (1.0 - alpha_cumprod_prev).sqrt() * pred_noise
        x_prev = alpha_cumprod_prev.sqrt() * pred_x0 + dir_xt
        return x_prev