Method p_sample_loop_trajectory

train_generation.py:267–292

Generate samples, returning intermediate images Useful for visualizing how denoised images evolve over time Args: repeat_noise_steps (int): Number of denoising timesteps in which the same noise is used across the batch. If >= 0, the initial noise is the

(self, denoise_fn, shape, device, freq,
                                 noise_fn=torch.randn,clip_denoised=True, keep_running=False)

        return img_t

    def p_sample_loop_trajectory(self, denoise_fn, shape, device, freq,
                                 noise_fn=torch.randn,clip_denoised=True, keep_running=False):
        """
        Generate samples, returning intermediate images
        Useful for visualizing how denoised images evolve over time
        Args:
          repeat_noise_steps (int): Number of denoising timesteps in which the same noise
            is used across the batch. If >= 0, the initial noise is the same for all batch elemements.
        """
        assert isinstance(shape, (tuple, list))

        total_steps =  self.num_timesteps if not keep_running else len(self.betas)

        img_t = noise_fn(size=shape, dtype=torch.float, device=device)
        imgs = [img_t]
        for t in reversed(range(0,total_steps)):

            t_ = torch.empty(shape[0], dtype=torch.int64, device=device).fill_(t)
            img_t = self.p_sample(denoise_fn=denoise_fn, data=img_t, t=t_, noise_fn=noise_fn,
                                  clip_denoised=clip_denoised,
                                  return_pred_xstart=False)
            if t % freq == 0 or t == total_steps-1:
                imgs.append(img_t)

        assert imgs[-1].shape == shape
        return imgs

    '''losses'''