-
Notifications
You must be signed in to change notification settings - Fork 0
/
vae_classifier.py
196 lines (164 loc) · 8.58 KB
/
vae_classifier.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
# Keras implementation of the paper:
# 3D MRI Brain Tumor Segmentation Using Autoencoder Regularization
# by Myronenko A. (https://arxiv.org/pdf/1810.11654.pdf)
# Author of this code: Suyog Jadhav (https://github.com/IAmSUyogJadhav)
from blocks import *
from utils import *
import torch
import torch.nn as nn
from collections import OrderedDict
import dataLoader as dl
# from group_norm import GroupNormalization
class Encoder(nn.Sequential):
def __init__(self, in_features=1, input_side_dim=48, model_depth=32):
super(Encoder, self).__init__()
next_padding = calc_same_padding(input_side_dim, 3, 1)
out_dim_0 = input_side_dim
out_dim_1 = calc_conv_shape(out_dim_0, 3, 0, 2)
out_dim_2 = calc_conv_shape(out_dim_1, 3, 0, 2)
out_dim_3 = calc_conv_shape(out_dim_2, 3, 0, 2)
# Define depth of model
init_d = model_depth
modules = [
('conv0', nn.Conv3d(in_features, init_d, kernel_size=3, stride=1, padding=next_padding)),
('sp_drop0', nn.Dropout3d(0.2)),
('green0', GreenBlock(init_d, init_d, out_dim_0)),
('downsize_0', nn.Conv3d(init_d, init_d * 2, kernel_size=3, stride=2, padding=1)),
# add padding to divide images by 2 exactly
('green10', GreenBlock(init_d * 2, init_d * 2, out_dim_1)),
('green11', GreenBlock(init_d * 2, init_d * 2, out_dim_1)),
('downsize_1', nn.Conv3d(init_d * 2, init_d * 4, kernel_size=3, stride=2, padding=1)),
('green20', GreenBlock(init_d * 4, init_d * 4, out_dim_2)),
('green21', GreenBlock(init_d * 4, init_d * 4, out_dim_2)),
('downsize_2', nn.Conv3d(init_d * 4, init_d * 8, kernel_size=3, stride=2, padding=1)),
('green30', GreenBlock(init_d * 8, init_d * 8, out_dim_3)),
('green31', GreenBlock(init_d * 8, init_d * 8, out_dim_3)),
('green32', GreenBlock(init_d * 8, init_d * 8, out_dim_3)),
('green33', GreenBlock(init_d * 8, init_d * 8, out_dim_3)),
]
for m in modules:
self.add_module(*m)
class Classifier(nn.Module):
def __init__(self, num_classes=5, input_side_dim=6, model_depth=32):
super(Classifier, self).__init__()
out_dim_0 = calc_conv_shape(input_side_dim, 1, 0, 1)
out_dim_1 = calc_conv_shape(out_dim_0, 1, 0, 1)
out_dim_2 = calc_conv_shape(out_dim_1, 1, 0, 1)
# print(out_dim_0, out_dim_1, out_dim_2)
next_padding = calc_same_padding(out_dim_2, 3, 1)
out_dim_3 = calc_conv_shape(out_dim_2, 3, next_padding, 1)
self.to_ground_truth = nn.Sequential(OrderedDict([
('conv0', nn.Conv3d(model_depth * 8, model_depth * 4, kernel_size=(1, 1, 1), stride=1, padding=0)),
('green0', GreenBlock(model_depth * 4, model_depth * 4, out_dim_0)),
('conv1', nn.Conv3d(model_depth * 4, model_depth * 2, kernel_size=(1, 1, 1), stride=1, padding=0)),
('green1', GreenBlock(model_depth * 2, model_depth * 2, out_dim_1)),
('conv2', nn.Conv3d(model_depth * 2, model_depth, kernel_size=(1, 1, 1), stride=1, padding=0)),
('green2', GreenBlock(model_depth, model_depth, out_dim_2)),
('conv3', nn.Conv3d(model_depth, model_depth, kernel_size=(3, 3, 3), stride=1, padding=next_padding)),
('relu', nn.LeakyReLU(inplace=True))
]))
# print('Classifier has {} features'.format(model_depth * out_dim_3 ** 3))
self.regressor = nn.Linear(in_features=model_depth * out_dim_3 ** 3, out_features=num_classes)
def forward(self, inputs):
conv_out = self.to_ground_truth(inputs)
return torch.nn.functional.relu(self.regressor(conv_out.view(conv_out.shape[0], -1)), inplace=True)
class VAERegularization(nn.Module):
def __init__(self, input_side_dim=6, model_depth=32):
super(VAERegularization, self).__init__()
# VAE regularization
self.reduce_dimension = nn.Sequential(OrderedDict([
# ('group_normR', GroupNormalization(in_features, groups=8)),
('norm0', nn.BatchNorm3d(model_depth * 8)),
('reluR0', nn.LeakyReLU(inplace=True)),
('convR0', nn.Conv3d(model_depth * 8, model_depth // 2, kernel_size=(3, 3, 3), stride=2, padding=1)),
]))
out_dim = 3
# print("out dim after VAE: {}".format(out_dim))
# REPARAMETERIZATION TRICK (needs flattening)
self.out_linear = nn.Linear(in_features=(model_depth // 2) * out_dim ** 3, out_features=model_depth * 8)
self.z_mean = nn.Linear(in_features=model_depth * 8, out_features=model_depth*4)
self.z_var = nn.Linear(in_features=model_depth * 8, out_features=model_depth*4)
self.reparameterization = Reparametrization()
def forward(self, inputs):
x = self.reduce_dimension(inputs)
x = self.out_linear(x.view(x.shape[0], -1))
z_mean = self.z_mean(x)
z_var = self.z_var(x)
del x
return self.reparameterization(z_mean, z_var), z_mean, z_var
class Decoder(nn.Module):
def __init__(self, model_depth=32, num_channels=1, input_side_dim=3):
super(Decoder, self).__init__()
self.model_depth = model_depth
self.input_side_dim = input_side_dim
self.reshape_block = nn.Sequential(OrderedDict([
('fc0', nn.Linear(in_features=model_depth*4, out_features=(model_depth // 2) * input_side_dim ** 3)),
('relu', nn.LeakyReLU(inplace=True)),
]))
blue_padding = calc_same_padding(48, 3, 1)
self.decode_block = nn.Sequential(OrderedDict([
('conv0', nn.Conv3d(model_depth // 2, model_depth * 8, kernel_size=1, stride=1)),
('up1', nn.Upsample(scale_factor=2)),
('upgreen0', UpGreenBlock(model_depth * 8, model_depth * 4, input_side_dim)),
('upgreen1', UpGreenBlock(model_depth * 4, model_depth * 2, 12)),
('upgreen2', UpGreenBlock(model_depth * 2, model_depth, 24)),
('blue_block', nn.Conv3d(model_depth, model_depth, kernel_size=3, stride=1, padding=blue_padding)),
('output_block', nn.Conv3d(in_channels=model_depth, out_channels=num_channels, kernel_size=1, stride=1))
]))
def forward(self, inputs):
x = self.reshape_block(inputs)
x = x.reshape([x.shape[0], self.model_depth // 2, self.input_side_dim, self.input_side_dim, self.input_side_dim])
x = self.decode_block(x)
return x
class BrainClassifierVAE(nn.Module):
def __init__(self, in_channels=1, input_side_dim=48, num_classes=16, model_depth=32):
super().__init__()
# ENCODING
self.encoder = Encoder(in_features=in_channels, input_side_dim=input_side_dim, model_depth=model_depth)
# input_side_dim = 6
# VAE regularization
self.internal_representation = VAERegularization(input_side_dim=6, model_depth=model_depth)
# DECODER
# The internal representation shrinks the dimension by a factor of 2
self.decoder = Decoder(
model_depth=model_depth,
num_channels=in_channels,
input_side_dim=3
)
# CLASSIFICATION
self.classifier = Classifier(num_classes=num_classes, input_side_dim=6, model_depth=model_depth)
def forward(self, inputs):
encoded = self.encoder(inputs)
out_features = self.classifier(encoded)
int_repr, z_mean, z_var = self.internal_representation(encoded)
del encoded
reconstructed_image = self.decoder(int_repr)
return out_features, reconstructed_image, inputs, z_mean, z_var
if __name__ == '__main__':
from losses import VAELoss
data, labels, input_shape = dl.get_Datas()
print(input_shape)
bigboi = BrainClassifierVAE(in_channels=53, input_side_dim=48, num_classes=8, model_depth=32)
input = torch.rand((10, 53, 48, 48, 48))
target = torch.rand((10, 8))
lr = 1e-4
weight_L2 = 0.1
weight_KL = 0.1
dice_e = 1e-8
optim = torch.optim.AdamW(bigboi.parameters(), lr=1e-4)
# Loss for features
loss_mse = torch.nn.MSELoss()
# Loss for VAE
loss_vae = VAELoss(
weight_KL=weight_KL,
weight_L2=weight_L2
)
out_features, reconstructed_image, input_image, z_mean, z_var = bigboi(input)
loss_mse_v = loss_mse(target, out_features)
loss_vae_v = loss_vae(reconstructed_image, input_image, z_mean, z_var)
loss = loss_mse_v + loss_vae_v
loss.backward()
print('loss: {}'.format(loss.item()))
print("features shape: {}".format(out_features.shape))
print("reconstructed shape: {}".format(reconstructed_image.shape))
print("z_mean: {} and z_var shapes: {}".format(z_mean.shape, z_var.shape))