-
Notifications
You must be signed in to change notification settings - Fork 0
/
VAE-ResNET.py
278 lines (247 loc) · 10.1 KB
/
VAE-ResNET.py
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
import loader #loads the MNIST dataset
import torch
import torchvision
from torchvision import transforms,datasets,models
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
import torch.optim as optim
from tqdm import tqdm
torch.seed = 41
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
batch_size = 400
epochs = 10
alpha = 0.005
model_file='VAE.pth'
# creates a convolution layer with 3x3 kernel size
def conv3x3(in_channels, out_channels, stride=1):
return nn.Conv2d(in_channels, out_channels, kernel_size=3,stride=stride, padding=1).to(device)
#creates a transpose convolution layer with 3x3 kernel size
def convT3x3(in_channels, out_channels, stride=1,padding=1):
return nn.ConvTranspose2d(in_channels, out_channels, kernel_size=3,stride=stride, padding=padding).to(device)
# We are using a ResNet to train the data
# creating the residual blocks of the resnet model
# this is used as encoder residual block of the model
class ResidualBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride=1, downsample=None):
super(ResidualBlock, self).__init__()
self.subblock_1=nn.Sequential(
conv3x3(in_channels, out_channels, stride).to(device),
nn.BatchNorm2d(out_channels).to(device),
nn.ReLU()
)
self.subblock_2=nn.Sequential(
conv3x3(out_channels, out_channels).to(device),
nn.BatchNorm2d(out_channels).to(device),
nn.ReLU(),
)
self.downsample = downsample
def forward(self, x):
residual = x
x = self.subblock_1(x).to(device)
x = self.subblock_2(x).to(device)
if self.downsample:
residual = self.downsample(residual)
x += residual
x = F.relu(x)
return x
block=ResidualBlock
#decoder residual block of the model
class DecoderResidualBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride=1,padding=1,upsample=None):
super(DecoderResidualBlock, self).__init__()
self.subblock_1=nn.Sequential(
convT3x3(in_channels, out_channels, stride,padding).to(device),
nn.BatchNorm2d(out_channels).to(device),
nn.ReLU()
)
self.subblock_2=nn.Sequential(
convT3x3(out_channels, out_channels).to(device),
nn.BatchNorm2d(out_channels).to(device),
nn.ReLU(),
)
self.upsample = upsample
def forward(self, x):
residual = x
x = self.subblock_1(x).to(device)
x = self.subblock_2(x).to(device)
if self.upsample:
residual = self.upsample(residual)
x += residual
x = F.relu(x)
return x
decoder_block=DecoderResidualBlock
# Creating the ResNet and inverse ResNet layers
class ResNet(nn.Module):
def __init__(self,num_classes=10):
super(ResNet, self).__init__()
self.in_channels=16
self.in_channels_decoder=32
self.layer=nn.Sequential(
conv3x3(1, 16).to(device),
nn.BatchNorm2d(16).to(device),
nn.ReLU()
)
self.layer1 = self.make_layer(block,16,2).to(device)
self.layer2 = self.make_layer(block, 32,2,2).to(device)
self.layer3 = self.make_layer(block, 64,2,2).to(device)
self.max_pool = nn.MaxPool2d(7,1).to(device)
self.fc = nn.Flatten()
self.trans_layer1 = self.make_trans_layer(decoder_block,32,2).to(device)
self.trans_layer2 = self.make_trans_layer(decoder_block, 16,2,2).to(device)
self.trans_layer3 = self.make_trans_layer(decoder_block,3,2,2).to(device)
self.trans_layer4 = nn.ConvTranspose2d(3,1,4,2,2).to(device)
#making resnet layers
def make_layer(self, block, out_channels, blocks, stride=1):
downsample = None
if (stride != 1) or (self.in_channels != out_channels):
downsample = nn.Sequential(
conv3x3(self.in_channels, out_channels, stride=stride).to(device),
nn.BatchNorm2d(out_channels).to(device)
)
layers = []
layers.append(block(self.in_channels, out_channels, stride, downsample))
self.in_channels = out_channels
for i in range(1, blocks):
layers.append(block(out_channels, out_channels))
return nn.Sequential(*layers)
# making inverse resnet layer
def make_trans_layer(self, decoder_block, out_channels, blocks, stride=1,padding=0):
upsample = None
if (stride != 1) or (self.in_channels_decoder != out_channels):
upsample = nn.Sequential(
convT3x3(self.in_channels_decoder, out_channels, stride=stride,padding=padding).to(device),
nn.BatchNorm2d(out_channels).to(device)
)
layers_dec = []
layers_dec.append(decoder_block(self.in_channels_decoder, out_channels, stride,padding, upsample))
self.in_channels_decoder = out_channels
for i in range(1, blocks):
layers_dec.append(decoder_block(out_channels, out_channels))
return nn.Sequential(*layers_dec)
def forward(self, x):
# encoder part
x = self.layer(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.max_pool(x)
# latent vector generation
latent = self.fc(x)
mean,std=torch.chunk(latent,2,dim=1)
#sampling of latent vecotr
sample = mean + torch.randn_like(std)*std
x=sample.view(sample.shape[0],32,1,1)
x = self.trans_layer1(x)
x = self.trans_layer2(x)
x = self.trans_layer3(x)
x = self.trans_layer4(x)
return x,mean,std
def decode(self, mean, std):
sample = mean + torch.randn_like(std)*std
x=sample.view(sample.shape[0],32,1,1)
x = self.trans_layer1(x)
x = self.trans_layer2(x)
x = self.trans_layer3(x)
x = self.trans_layer4(x)
return x
AutoEncoder=ResNet()
#optimizer function
optimizer = torch.optim.Adam(AutoEncoder.parameters(), lr=alpha)
# calculation of variational loss
def variational_loss(output,X_in,mean,std):
loss_function = nn.MSELoss()
loss_by_function = loss_function(output,X_in)
kl_loss = -(0.5/batch_size)*torch.sum(1+torch.log(torch.pow(std,2)+1e-10)-torch.pow(std,2)-torch.pow(mean,2))
total_loss = loss_by_function+kl_loss
return total_loss
#training function
def train(X):
loss_list = []
im_list = []
iters=0
j=0
for epoch in tqdm(range(0,epochs)):
cost = 0
batch=torch.randperm(X.shape[0]).to(device)
for i in tqdm(range(0, X.shape[0],batch_size)):
output,mean,std = AutoEncoder(X[batch[i:i+batch_size]].to(device))
optimizer.zero_grad()
loss = variational_loss(output,X[batch[i:i+batch_size]],mean,std)
cost = cost+loss.item()
loss.backward()
optimizer.step()
# to generate random image
if (iters % 50 == 0) or ((epoch == epochs-1) and (j == len(X)-1)):
with torch.no_grad():
test = AutoEncoder.decode(mean,std).detach().cpu()
im_list.append(np.squeeze(test[0].permute(1,2,0)))
iters+=1
j+=1
loss_avg = cost / X.shape[0]
loss_list.append(loss_avg)
print("For iteration: ", epoch+1, " the loss is :", loss_avg)
return loss_list,im_list
def test(X):
with torch.no_grad():
cost = 0
batch = torch.randperm(X.shape[0])
for i in tqdm(range(0, X.shape[0],batch_size)):
output,mean,std = AutoEncoder(X[batch[i:i+batch_size]])
loss = variational_loss(output,X[batch[i:i+batch_size]],mean,std)
cost = cost+loss.item()
print("Test set loss:",cost/X.shape[0])
def main():
train_need = input("Press l to load model, t to train model, tl to load and train model: ").lower()
# Asks user whether to load saved model or train from scratch, or train the saved loss
if train_need == 't':
#loading train set images as tensors
train_images = loader.train_loader_fn()
loss_list,im_list = train(train_images)
elif train_need == 'l':
AutoEncoder.load_state_dict(torch.load(model_file))
elif train_need == 'tl':
AutoEncoder.load_state_dict(torch.load(model_file))
#loading train set images as tensors
train_images = loader.train_loader_fn()
loss_list,im_list = train(train_images)
#to save randomly generated images
i=0
try:
for l in im_list:
i+=1
plt.savefig(str(i))
plt.imshow(l,cmap="gray")
# plotting the cost function
plt.plot(loss_list)
plt.title("Loss curve")
plt.ylabel('cost')
plt.xlabel('epoch number')
plt.show()
except:
pass
# loading the test set of images
test_images=loader.test_loader_fn()
test(test_images)
n = 10 # number of images that are to be displayed
# n test images passed through variational autoencoder
output = AutoEncoder(test_images[:n])
output_img = ((output[0].to(torch.device('cpu'))).detach().numpy()).reshape(n,28,28)
for i in range(0,n):
axes = plt.subplot(2,n,i+1)
plt.imshow(loader.test_img[i],cmap = "gray")
axes.get_xaxis().set_visible(False) #removing axes
axes.get_yaxis().set_visible(False)
axes = plt.subplot(2,n,n+i+1)
plt.imshow(output_img[i],cmap="gray")
axes.get_xaxis().set_visible(False)
axes.get_yaxis().set_visible(False)
plt.show()
if train_need == 't' or train_need == 'tl':
# If the model was trained, it asks whether or not to save the model
save_status=input("Enter s to save the model: ").lower()
if save_status=='s':
torch.save(AutoEncoder.state_dict(),model_file)
if __name__ == "__main__":
main()