OrdinalDiffusionModels/sample.py at main · GustavKS/OrdinalDiffusionModels · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
import os
from omegaconf import OmegaConf

import torch

from ori.models import get_model
from ori.models.autoencoder import VAE
from ori.training import NoiseScheduler


def denormalize(img,  mean=[0.3704248070716858, 0.2282254546880722, 0.13915641605854034],
                      std=[0.23381589353084564, 0.1512117236852646, 0.09653093665838242]):
    mean = torch.tensor(mean, device=img.device).view(1, -1, 1, 1)
    std = torch.tensor(std, device=img.device).view(1, -1, 1, 1)
    return img * std + mean


class Sampler:
  def __init__(self, exp_path, guide_w):
      self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

      self.cfg = OmegaConf.load(exp_path + "/hparams.yaml").cfg

      self.model = get_model(self.cfg)
      checkpoint_dir = os.path.join(exp_path, "checkpoints")
      checkpoint_files = [f for f in os.listdir(checkpoint_dir) if f.endswith(".pt")]

      self.vae = VAE(in_channels=3, latent_channels=4, ch_mult=[1, 2, 4])
      self.vae.load_state_dict(torch.load(self.cfg.vae.checkpoint, map_location=self.device))
      self.vae.to(self.device)
      self.vae.eval()

      if not checkpoint_files:
          raise FileNotFoundError(f"No .pt checkpoint found in {checkpoint_dir}")
      checkpoint_path = os.path.join(checkpoint_dir, checkpoint_files[0])
      self.model.load_state_dict(torch.load(checkpoint_path, map_location=self.device), strict=False)
      self.model.to(self.device)
      self.model.eval()

      self.output_dir = exp_path + f"/samples_gw{guide_w}"
      if not os.path.exists(self.output_dir):
          os.makedirs(self.output_dir)

  def sample(self, labels, iqs, guide_w, sample_method, batch_size):
    all_samples = []

    for i in range(0, len(labels), batch_size):
        batch_labels = labels[i:i+batch_size]
        batch_iqs = iqs[i:i+batch_size]
        batch_samples = self._sample_batch(batch_labels, batch_iqs, guide_w, sample_method)
        all_samples.append(batch_samples)

        torch.cuda.empty_cache()

    return torch.cat(all_samples, dim=0)

  def _sample_batch(self, labels, iqs, guide_w, sample_method="ddpm"):
    x_i = torch.randn((len(labels), 4, self.cfg.data.image_resolution // 4, self.cfg.data.image_resolution // 4), device=self.device)
    labels = labels.to(self.device)
    labels_in = labels.repeat(2)

    context_mask = torch.zeros_like(labels, dtype=torch.float).to(self.device)
    context_mask = context_mask.repeat(2)
    context_mask[labels.size(0):] = 1

    structure = torch.zeros((len(labels_in), 1024), device=self.device)
    structure_mask = torch.ones_like(context_mask, dtype=torch.float).to(self.device)

    NS = NoiseScheduler(self.cfg)
    NS.to(self.device)
    T = NS.timesteps-1

    start_w = 0.5

    with torch.no_grad():
      if sample_method == "ddpm":
        for t in reversed(range(T)):
          t_tensor = torch.full((len(labels),), t, device=self.device)
          t_in = t_tensor.repeat(2)

          x_in = x_i.repeat(2, 1, 1, 1)

          predicted_noise = self.model(x_in, labels_in, t_in, context_mask, structure, structure_mask)

          if self.cfg.model.name == "DiT":
             predicted_noise = predicted_noise[:, :4, :, :]
          eps_cond, eps_uncond = predicted_noise.chunk(2, dim=0)

          eps_guided = eps_uncond + guide_w * (eps_cond - eps_uncond)

          alpha_t = NS.alphas[t]
          alpha_hat_t = NS.alpha_hats[t]
          beta_t = NS.betas[t]

          x_i = (1/alpha_t.sqrt()) * (x_i - beta_t/(1-alpha_hat_t).sqrt() * eps_guided)

          if t > 0:
            temp = 1
            noise = torch.randn_like(x_i) * temp
            x_i += beta_t.sqrt() * noise
      elif sample_method == "ddim":
        ddim_steps = 100
        ddim_timesteps = torch.linspace(0, T, ddim_steps, device=self.device).long()
        for i in reversed(range(len(ddim_timesteps))):
          t = ddim_timesteps[i]
          t_prev = ddim_timesteps[i - 1] if i > 0 else 0

          t_tensor = torch.full((len(labels),), t, device=self.device)
          t_in = t_tensor.repeat(2)

          x_in = x_i.repeat(2, 1, 1, 1)

          predicted_noise = self.model(x_in, labels_in, t_in, context_mask, structure, structure_mask)
          if self.cfg.model.name == "DiT":
             predicted_noise = predicted_noise[:, :4, :, :]

          eps_cond, eps_uncond = predicted_noise.chunk(2, dim=0)

          progress = t/T

          guide_w_t = start_w + progress * (guide_w - start_w)
          eps_guided = eps_uncond + guide_w_t * (eps_cond - eps_uncond)

          alpha_hat_t = NS.alpha_hats[t]
          alpha_hat_t_prev = NS.alpha_hats[t_prev] if t_prev > 0 else torch.tensor(1.0).to(self.device)

          x_0_hat = (x_i - (1 - alpha_hat_t).sqrt() * eps_guided) / alpha_hat_t.sqrt()
          x_i = alpha_hat_t_prev.sqrt() * x_0_hat + (1 - alpha_hat_t_prev).sqrt() * eps_guided

    with torch.no_grad():
      x_i = self.vae.decoder(x_i)

    x_i = denormalize(x_i)
    return x_i