Text-to-Image
Diffusers
English
File size: 10,390 Bytes
eaefa93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ee90412
eaefa93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
'''
THis gile is to contain the DDPM implementation modularized for loading, prediciton and training.
'''

from torch import nn
import math
import torch
from utils import forward_diffusion_sample, sample_timestep, sample_plot_image
import torch.nn.functional as F
from attn_utils import SelfAttention, CBAM, Block_CBAM

class Block(nn.Module):
  def __init__(self, in_ch, out_ch, time_emb_dim, up=False):
    super().__init__()
    self.time_mlp = nn.Linear(time_emb_dim, out_ch)
    if up:
      ## up channel - gobig big big bigg from smol smol smol with 3x3 kernel
      self.conv1 = nn.Conv2d(2*in_ch, out_ch, 3, padding=1)
      self.transform = nn.ConvTranspose2d(out_ch, out_ch, 4, 2, 1)
    else:
      self.conv1 = nn.Conv2d(in_ch, out_ch, 3, padding=1)
      self.transform = nn.Conv2d(out_ch, out_ch, 4,2,1)
    self.conv2 = nn.Conv2d(out_ch, out_ch, 3, padding=1)
    self.relu = nn.ReLU()
    self.batch_norm1 = nn.BatchNorm2d(out_ch)
    self.batch_norm2 = nn.BatchNorm2d(out_ch)

  def forward(self, x, t, ):
    h = self.batch_norm1(self.relu(self.conv1(x)))
    time_emb = self.relu(self.time_mlp(t))
    time_emb = time_emb[(..., ) + (None, ) * 2]
    h = h + time_emb
    h = self.batch_norm2(self.relu(self.conv2(h)))
    return self.transform(h)
  
class PositionEmbeddings(nn.Module):
  def __init__(self,dim):
    super().__init__()
    self.dim = dim

  def forward(self, time):
    device = time.device
    half_dim = self.dim // 2
    embeddings = math.log(10000) / (half_dim - 1)
    embeddings = torch.exp(torch.arange(half_dim, device=device) * -embeddings)
    embeddings = time[:, None] * embeddings[None, :]
    embeddings = torch.cat((embeddings.sin(), embeddings.cos()), dim=-1)
    return embeddings
  


class SimpleUnet(nn.Module):
  def __init__(self):
    super().__init__()
    image_channels = 3
    down_channels = (64, 128, 256, 512, 1024)
    up_channels = (1024, 512, 256, 128, 64)
    self.device = "cuda" if torch.cuda.is_available() else "cpu"

    out_dim = 3
    time_emb_dim = 32

    ## timestep stored as positional encoding in terms of sine
    self.time_mlp = nn.Sequential(
        PositionEmbeddings(time_emb_dim),
        nn.Linear(time_emb_dim, time_emb_dim),
        nn.ReLU()
    )

    
    self.conv0 = nn.Conv2d(image_channels, down_channels[0], 3, padding=1)
    self.down_blocks = nn.ModuleList([
        Block(down_channels[i], down_channels[i+1], time_emb_dim)
        for i in range(len(down_channels)-1)
    ])
    self.up_blocks = nn.ModuleList([
        Block(up_channels[i], up_channels[i+1], time_emb_dim, up=True)
        for i in range(len(up_channels)-1)
    ])

    ## readout layer
    self.output = nn.Conv2d(up_channels[-1], out_dim, 1)

  def forward(self, x, timestep):
    t = self.time_mlp(timestep)
    x = self.conv0(x)
    residual_inputs = []
    for down in self.down_blocks:
      x = down(x, t)
      residual_inputs.append(x)
    for up in self.up_blocks:
      residual_x = residual_inputs.pop()
      x = torch.cat((x, residual_x), dim=1)
      x = up(x, t)
    return self.output(x)
  
  @torch.no_grad()
  def sample(self, noise):
      """
      Generate an image by denoising a given noise tensor using the reverse diffusion process.

      Args:
          noise (torch.Tensor): Initial noise tensor (e.g., sampled from a Gaussian distribution).
      
      Returns:
          torch.Tensor: Denoised image.
      """
      img = noise  # Start with the provided noise tensor
      T = self.num_timesteps  # Total timesteps for diffusion
      stepsize = 1  # You can adjust if needed

      # Iterate through the timesteps in reverse order
      for i in range(0, T)[::-1]:
          t = torch.full((noise.size(0),), i, device=noise.device, dtype=torch.long)  # Current timestep
          img = sample_timestep(self, img, t)  # Perform one reverse diffusion step
          img = torch.clamp(img, -1.0, 1.0)  # Clamp the image to ensure values stay in [-1, 1]

      return img

  def get_loss(self, x_0, t):
      x_noisy, noise = forward_diffusion_sample(x_0, t, self.device)
      noise_pred = self(x_noisy, t)
      return F.l1_loss(noise, noise_pred)

  def train(self, dataloader, BATCH_SIZE=64,T=300, EPOCHS=50, verbose=True):
      from torch.optim import Adam

      device = "cuda" if torch.cuda.is_available() else "cpu"
      self.to(device)
      optimizer = Adam(self.parameters(), lr=0.001)
      epochs = EPOCHS

      for epoch in range(epochs):
          for step, batch in enumerate(dataloader):
              optimizer.zero_grad()

              t = torch.randint(0, T, (BATCH_SIZE,), device=device).long()
              loss = self.get_loss(self, batch[0], t)
              loss.backward()
              optimizer.step()

              if verbose:
                if epoch % 5 == 0 and step % 150 == 0:
                    print(f"Epoch {epoch} | step {step:03d} Loss: {loss.item()} ")
                    sample_plot_image(self)

  def test():
    ## TODO: add the testing loop here
    pass




################################################################################################
####################### ATTENTION LAYERS ADDEDD TO THE MODEL ###################################
################################################################################################

class SimpleUnetWSelfAttn(nn.Module):
  def __init__(self):
    super().__init__()
    image_channels = 3
    down_channels = (64, 128, 256, 512, 1024)
    up_channels = (1024, 512, 256, 128, 64)

    out_dim = 3
    time_emb_dim = 32

    ## timestep stored as positional encoding in terms of sine
    self.time_mlp = nn.Sequential(
        PositionEmbeddings(time_emb_dim),
        nn.Linear(time_emb_dim, time_emb_dim),
        nn.ReLU()
    )
    self.num_timesteps = 300


    self.conv0 = nn.Conv2d(image_channels, down_channels[0], 3, padding=1)
    self.down_blocks = nn.ModuleList([
        Block(down_channels[i], down_channels[i+1], time_emb_dim)
        for i in range(len(down_channels)-1)
    ])
    self.up_blocks = nn.ModuleList([
        Block(up_channels[i], up_channels[i+1], time_emb_dim, up=True)
        for i in range(len(up_channels)-1)
    ])

    self.self_attention = SelfAttention(down_channels[-1])


    ## readout layer
    self.output = nn.Conv2d(up_channels[-1], out_dim, 1)


  # def settimestep()

  def forward(self, x, timestep):
    self.num_timesteps = timestep
    t = self.time_mlp(timestep)
    x = self.conv0(x)
    residual_inputs = []
    for down in self.down_blocks:
      x = down(x, t)
      residual_inputs.append(x)

    x = self.self_attention(x)
    
    for up in self.up_blocks:
      residual_x = residual_inputs.pop()
      x = torch.cat((x, residual_x), dim=1)
      x = up(x, t)
    return self.output(x)

  @torch.no_grad()
  def sample(self, noise):
      """
      Generate an image by denoising a given noise tensor using the reverse diffusion process.

      Args:
          noise (torch.Tensor): Initial noise tensor (e.g., sampled from a Gaussian distribution).
      
      Returns:
          torch.Tensor: Denoised image.
      """
      img = noise  # Start with the provided noise tensor
      T = self.num_timesteps  # Total timesteps for diffusion
      stepsize = 1  # You can adjust if needed
      print(noise.device)

      # Iterate through the timesteps in reverse order
      for i in range(T - 1, -1, -1):
          t = torch.full((noise.size(0),), i, device=noise.device, dtype=torch.long)  # Current timestep
          img = sample_timestep(self, img, t)  # Perform one reverse diffusion step
          img = torch.clamp(img, -1.0, 1.0)  # Clamp the image to ensure values stay in [-1, 1]

      return img
  


################################################################################################
#################### Convolutional Block Attention Module ADDED TO THE MODEL ###################
################################################################################################

class SimpleUnetWCBAM(nn.Module):
  def __init__(self):
    super().__init__()
    image_channels = 3
    down_channels = (64, 128, 256, 512, 1024)
    up_channels = (1024, 512, 256, 128, 64)

    out_dim = 3
    time_emb_dim = 32

    ## timestep stored as positional encoding in terms of sine
    self.time_mlp = nn.Sequential(
        PositionEmbeddings(time_emb_dim),
        nn.Linear(time_emb_dim, time_emb_dim),
        nn.ReLU()
    )
    self.num_timesteps = 300


    self.conv0 = nn.Conv2d(image_channels, down_channels[0], 3, padding=1)
    self.down_blocks = nn.ModuleList([
        Block_CBAM(down_channels[i], down_channels[i+1], time_emb_dim)
        for i in range(len(down_channels)-1)
    ])
    self.up_blocks = nn.ModuleList([
        Block_CBAM(up_channels[i], up_channels[i+1], time_emb_dim, up=True)
        for i in range(len(up_channels)-1)
    ])

    self.self_attention = SelfAttention(down_channels[-1])


    ## readout layer
    self.output = nn.Conv2d(up_channels[-1], out_dim, 1)


  # def settimestep()

  def forward(self, x, timestep):
    self.num_timesteps = timestep
    t = self.time_mlp(timestep)
    x = self.conv0(x)
    residual_inputs = []
    for down in self.down_blocks:
      x = down(x, t)
      residual_inputs.append(x)

    x = self.self_attention(x)
    
    for up in self.up_blocks:
      residual_x = residual_inputs.pop()
      x = torch.cat((x, residual_x), dim=1)
      x = up(x, t)
    return self.output(x)

  @torch.no_grad()
  def sample(self, noise):
      """
      Generate an image by denoising a given noise tensor using the reverse diffusion process.

      Args:
          noise (torch.Tensor): Initial noise tensor (e.g., sampled from a Gaussian distribution).
      
      Returns:
          torch.Tensor: Denoised image.
      """
      img = noise  # Start with the provided noise tensor
      T = self.num_timesteps  # Total timesteps for diffusion
      stepsize = 1  # You can adjust if needed
      print(noise.device)

      # Iterate through the timesteps in reverse order
      for i in range(T - 1, -1, -1):
          t = torch.full((noise.size(0),), i, device=noise.device, dtype=torch.long)  # Current timestep
          img = sample_timestep(self, img, t)  # Perform one reverse diffusion step
          img = torch.clamp(img, -1.0, 1.0)  # Clamp the image to ensure values stay in [-1, 1]

      return img