-
Notifications
You must be signed in to change notification settings - Fork 9
/
models.py
264 lines (205 loc) · 9.44 KB
/
models.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
import torch
import torch.nn.functional as F
import torchvision.models as models
import torch.nn as nn
from collections import OrderedDict
def conv_output_shape(h_w, kernel_size=(1, 1), stride=(1, 1), pad=(0, 0), dilation=1):
h = (h_w[0] + (2 * pad[0]) - (dilation * (kernel_size[0] - 1)) - 1)// stride[0] + 1
w = (h_w[1] + (2 * pad[1]) - (dilation * (kernel_size[1] - 1)) - 1)// stride[1] + 1
return h, w
def convtransp_output_shape(h_w, kernel_size=(1, 1), stride=(1, 1), pad=(0, 0), dilation=1):
h = (h_w[0] - 1) * stride[0] - 2 * pad[0] + kernel_size[0] + pad[0]
w = (h_w[1] - 1) * stride[1] - 2 * pad[1] + kernel_size[1] + pad[1]
return h, w
class LSTMEncoder(nn.Module):
def __init__(self, input_size, hidden_size, device='cpu', num_layers=1):
super(LSTMEncoder, self).__init__()
self.hidden_size = hidden_size
self.num_layers = num_layers
self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=False)
self.device = device
def forward(self, x):
# Initialize hidden state with zeros
S, B, z = x.shape
h0 = torch.zeros(self.num_layers, B, self.hidden_size).to(self.device)
# Initialize cell state
c0 = torch.zeros(self.num_layers, B, self.hidden_size).to(self.device)
all_timesteps, hidden_last_timestep = self.lstm(x, (h0, c0))
return hidden_last_timestep
class LSTMDecoder(nn.Module):
def __init__(self, input_size, hidden_size, device='cpu', num_layers=1):
super(LSTMDecoder, self).__init__()
self.lstm = nn.LSTM(input_size, hidden_size, num_layers, batch_first=False)
self.device = device
def forward(self, sizes, steps, hidden):
S, B, z = sizes
input = torch.zeros([max(steps), B, z], dtype=torch.float).to(self.device)
preds, hidden = self.lstm(input, hidden)
preds = preds.permute(1, 0, 2) # (Seq_len, B, z) --> (B, Seq_len, z)
return preds
# https://github.com/mateuszbuda/brain-segmentation-pytorch
class ModifiedUNet(nn.Module):
def __init__(self, args, in_channels=1, out_channels=1, bottleneck_out=None, init_features=32):
super(ModifiedUNet, self).__init__()
features = init_features
features = init_features
self.bottleneck_out = bottleneck_out
self.args = args
if not bottleneck_out:
self.bottleneck_out = features * 16
# encoder
self.encoder1 = ModifiedUNet._block(in_channels, features, name="enc1")
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.bn_enc1 = nn.BatchNorm2d(features)
self.encoder2 = ModifiedUNet._block(features, features * 2, name="enc2")
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.bn_enc2 = nn.BatchNorm2d(features * 2)
self.encoder3 = ModifiedUNet._block(features * 2, features * 4, name="enc3")
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
self.bn_enc3 = nn.BatchNorm2d(features * 4)
self.encoder4 = ModifiedUNet._block(features * 4, features * 8, name="enc4")
self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2)
self.bn_enc4 = nn.BatchNorm2d(features * 8)
# bottleneck
self.bottleneck = ModifiedUNet._block(features * 8, self.bottleneck_out, name="bottleneck")
self.avg_pool = torch.nn.AdaptiveAvgPool2d(output_size=1)
# VAE
self.mu = torch.nn.Linear(in_features=self.bottleneck_out, out_features=self.args.z_size)
self.sigma = torch.nn.Linear(in_features=self.bottleneck_out, out_features=self.args.z_size)
# LSTM
self.lstm_encoder = LSTMEncoder(input_size=args.z_size, hidden_size=args.z_size, device=args.device).to(args.device)
self.lstm_decoder = LSTMDecoder(input_size=args.z_size, hidden_size=args.z_size, device=args.device).to(args.device)
# decoder
# calculate what shape should be before upconv4
self.upconv6 = nn.ConvTranspose2d(
args.z_size, features * 16, kernel_size=2, stride=2
)
self.decoder6 = ModifiedUNet._block(features * 16, features * 16, name="dec6")
self.bn_dec6 = nn.BatchNorm2d(features * 16)
self.upconv5 = nn.ConvTranspose2d(
features * 16, features * 16, kernel_size=2, stride=2
)
self.decoder5 = ModifiedUNet._block(features * 16, features * 16, name="dec5")
self.bn_dec5 = nn.BatchNorm2d(features * 16)
self.upconv4 = nn.ConvTranspose2d(
features * 16, features * 8, kernel_size=2, stride=2
)
self.decoder4 = ModifiedUNet._block(features * 8, features * 8, name="dec4")
self.bn_dec4 = nn.BatchNorm2d(features * 8)
self.upconv3 = nn.ConvTranspose2d(
features * 8, features * 4, kernel_size=2, stride=2
)
self.decoder3 = ModifiedUNet._block(features * 4, features * 4, name="dec3")
self.bn_dec3 = nn.BatchNorm2d(features * 4)
self.upconv2 = nn.ConvTranspose2d(
features * 4, features * 2, kernel_size=2, stride=2
)
self.decoder2 = ModifiedUNet._block(features * 2, features * 2, name="dec2")
self.bn_dec2 = nn.BatchNorm2d(features * 2)
self.upconv1 = nn.ConvTranspose2d(
features * 2, features, kernel_size=2, stride=2
)
self.decoder1 = ModifiedUNet._block(features, features, name="dec1")
self.bn_dec1 = nn.BatchNorm2d(features)
self.conv = nn.Conv2d(
in_channels=features, out_channels=out_channels, kernel_size=1
)
def forward(self, x, idxs, steps):
enc1 = self.encoder1(x)
enc1 = self.bn_enc1(enc1)
enc1 = F.relu(enc1)
enc2 = self.pool1(enc1)
enc2 = self.encoder2(enc2)
enc2 = self.bn_enc2(enc2)
enc2 = F.relu(enc2)
enc3 = self.pool2(enc2)
enc3 = self.encoder3(enc3)
enc3 = self.bn_enc3(enc3)
enc3 = F.relu(enc3)
enc4 = self.pool3(enc3)
enc4 = self.encoder4(enc4)
enc4 = self.bn_enc4(enc4)
enc4 = F.relu(enc4)
bottleneck = self.bottleneck(self.pool4(enc4))
avg_pooled = self.avg_pool(bottleneck)
avg_pooled = avg_pooled.squeeze(-1).squeeze(-1)
# VAE
z_mu = self.mu.forward(avg_pooled)
z_sigma = self.sigma.forward(avg_pooled)
bs = x.shape[0]
eps = torch.randn(bs, self.args.z_size).to(self.args.device) * z_sigma + z_mu # Sampling epsilon from normal distributions
z_vector = z_mu + z_sigma * eps # z ~ Q(z|X)
# 1D to 2D
z_sequences = [z_vector[i['idx_from'] : i['idx_to']] for i in idxs]
z_sequences = torch.stack(z_sequences)
z_sequences = z_sequences.permute(1, 0, 2) # (B, Seq_len, z_vector) --> (Seq_len, B, z_vector)
# encode time, return last lastm hidden state
hidden_last_timestep = self.lstm_encoder.forward(z_sequences)
# decode and predict time
preds = self.lstm_decoder.forward(z_sequences.shape, steps, hidden_last_timestep)
# pick predictions that we need
picked_preds = []
for batch_idx, to_step in enumerate(steps):
picked_preds.append(preds[batch_idx][0:to_step])
masked_catted = torch.cat(picked_preds, dim=0)
# add H and W dims
masked_catted = masked_catted.unsqueeze(-1).unsqueeze(-1)
expand = self.upconv6(masked_catted)
expand = self.decoder6(expand)
expand = self.bn_dec6(expand)
expand = F.relu(expand)
expand = self.upconv5(expand)
expand = self.decoder5(expand)
expand = self.bn_dec5(expand)
expand = F.relu(expand)
dec4 = self.upconv4(expand)
dec4 = self.decoder4(dec4)
dec4 = self.bn_dec4(dec4)
dec4 = F.relu(dec4)
dec3 = self.upconv3(dec4)
dec3 = self.decoder3(dec3)
dec3 = self.bn_dec3(dec3)
dec3 = F.relu(dec3)
dec2 = self.upconv2(dec3)
dec2 = self.decoder2(dec2)
dec2 = self.bn_dec2(dec2)
dec2 = F.relu(dec2)
dec1 = self.upconv1(dec2)
dec1 = self.decoder1(dec1)
dec1 = self.bn_dec1(dec1)
dec1 = F.relu(dec1)
out = self.conv(dec1)
# out = torch.sigmoid(out)
return out, z_mu, z_sigma
@staticmethod
def _block(in_channels, features, name):
return nn.Sequential(
OrderedDict(
[
(
name + "conv1",
nn.Conv2d(
in_channels=in_channels,
out_channels=features,
kernel_size=3,
padding=1,
bias=False,
),
),
(name + "norm1", nn.BatchNorm2d(num_features=features)),
(name + "relu1", nn.ReLU(inplace=True)),
(
name + "conv2",
nn.Conv2d(
in_channels=features,
out_channels=features,
kernel_size=3,
padding=1,
bias=False,
),
),
(name + "norm2", nn.BatchNorm2d(num_features=features)),
(name + "relu2", nn.ReLU(inplace=True)),
]
)
)