-
Notifications
You must be signed in to change notification settings - Fork 1
/
CombNet.py
260 lines (195 loc) · 8.29 KB
/
CombNet.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
import torch.nn as nn
import torch.nn.functional as F
import torch
import torchvision.models as models
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
import os
class DataGen(Dataset):
def __init__(self):
self.base_dir = 'datasets'
self.classes = ['akiec', 'bcc', 'bkl', 'df', 'mel', 'nv', 'vasc']
self.imgs = []
self.target = [[0 for i in range(7)] for j in range(115 * 7)]
for c in self.classes:
self.imgs += os.listdir(os.path.join(self.base_dir, c))
for i in range(7 * 115):
self.target[i][int(i / 115)] = 1
# define transform
self.transform = transforms.Compose([ToTensor()])
# random indexing
rnd_index = np.random.permutation(len(self.imgs)).astype(np.int)
self.imgs = np.array(self.imgs)[rnd_index]
self.target = np.array(self.target)[rnd_index]
def __len__(self):
return len(self.imgs)
def __getitem__(self, index):
c = self.classes[np.argmax(self.target[index])]
img_path = os.path.join(self.base_dir, c, self.imgs[index])
img = Image.open(img_path).convert("RGB")
# image crop
img = img.crop((75, 0, 525, 450))
# resize
img = img.resize((224, 224))
one_img, one_target = self.transform((np.array(img), self.target[index]))
return one_img, one_target
class ToTensor(object):
""" Convert ndarrays in sample to Tensors. """
def __call__(self, data):
img = data[0]
target = data[1]
# swap color axis because
# numpy image: H x W x C
# torch image: C X H X W
img = img.transpose((2, 0, 1))
return torch.from_numpy(img), torch.from_numpy(target)
d = DataGen()
indices = torch.randperm(len(d)).tolist()
dataset_train = torch.utils.data.Subset(d, indices[:600])
dataset_test = torch.utils.data.Subset(d, indices[600:])
train_loader = DataLoader(dataset_train, batch_size=60, num_workers=1)
test_loader = DataLoader(dataset_test, batch_size=1, num_workers=1)
class FeaturePyramidStructure(nn.Module):
""" CombNet's backbone (using ResNet-18) """
def __init__(self):
super(FeaturePyramidStructure, self).__init__()
self.resnet18 = models.resnet18(pretrained=False)
self.layer_list = list(self.resnet18.children())
self.c1_module = self.layer_list[:3]
self.c5_module = self.layer_list[3:5]
self.c17_module = self.layer_list[5:9]
self.fc = nn.Linear(in_features=512, out_features=7)
def forward(self, x):
step = x
# conv1, batchnorm, relu
for layer in self.c1_module:
step = layer(step)
c1 = step
# maxpool, conv2_x
for layer in self.c5_module:
step = layer(step)
c5 = step
# conv3_x ~ conv5_x, avgpool
for layer in self.c17_module:
step = layer(step)
c17 = step
top_down_1 = F.softmax(self.fc(c17.view(c17.shape[0], -1)), dim=1)
return top_down_1, c5, c1
# top_down_1 is backbone's output
# c5 goes to SubNet2
# c1 goes to SubNet1
class SubNet1(nn.Module):
""" Backbone's sub-network1 for other scale features """
def __init__(self):
super(SubNet1, self).__init__()
self.conv1 = nn.Conv2d(64, 128, 3, stride=1, padding=1)
self.conv1_2 = nn.Conv2d(128, 128, 3, stride=1, padding=1)
self.conv2 = nn.Conv2d(128, 256, 3, stride=1, padding=1)
self.conv2_2 = nn.Conv2d(256, 256, 3, stride=1, padding=1)
self.conv3 = nn.Conv2d(256, 256, 3, stride=1, padding=1)
self.conv4 = nn.Conv2d(256, 256, 3, stride=1, padding=1)
self.pool = nn.MaxPool2d((2, 2), stride=2)
self.fc1 = nn.Linear(7 * 7 * 256, 512)
self.fc2 = nn.Linear(512, 7)
def forward(self, x):
c1_1 = F.relu(self.conv1(self.pool(x))) # (112, 112, 64) -> (56, 56, 128)
c1_2 = F.relu(self.conv1_2(c1_1)) # (56, 56, 128) -> (56, 56, 128)
c2_1 = F.relu(self.conv2(self.pool(c1_2))) # (56, 56, 128) -> (28, 28, 256)
c2_2 = F.relu(self.conv2_2(c2_1)) # (28, 28, 256) -> (28, 28, 256)
c3_1 = F.relu(self.conv3(self.pool(c2_2))) # (28, 28, 256) -> (14, 14, 256)
c3_2 = F.relu(self.conv4(c3_1)) # (14, 14, 256) -> (14, 14, 256)
c4 = self.pool(c3_2) # (14, 14, 256) -> (7, 7, 256)
c5 = F.relu(self.fc1(c4.view(c4.shape[0], -1))) # (7, 7, 256) -> 512
c6 = self.fc2(c5) # 512 -> 7
c7 = F.softmax(c6, dim=1)
return c7
class SubNet2(nn.Module):
""" Backbone's sub-network2 for other scale features """
def __init__(self):
super(SubNet2, self).__init__()
self.conv1 = nn.Conv2d(64, 128, 3, stride=1, padding=1)
self.conv1_2 = nn.Conv2d(128, 128, 3, stride=1, padding=1)
self.conv2 = nn.Conv2d(128, 128, 3, stride=1, padding=1)
self.pool = nn.MaxPool2d((2, 2), stride=2)
self.fc1 = nn.Linear(7 * 7 * 128, 512)
self.fc2 = nn.Linear(512, 7)
def forward(self, x):
c1_1 = F.relu(self.conv1(self.pool(x))) # (56, 56, 64) -> (28, 28, 128)
c1_2 = F.relu(self.conv1_2(c1_1)) # (28, 28, 128) -> (28, 28, 128)
c2_1 = F.relu(self.conv2(self.pool(c1_2))) # (28, 28, 128) -> (14, 14, 128)
c2_2 = F.relu(self.conv2(c2_1)) # (14, 14, 128) -> (14, 14, 128)
c3 = self.pool(c2_2) # (14, 14, 128) -> (7, 7, 128)
c4 = F.relu(self.fc1(c3.view(c3.shape[0], -1))) # (7, 7, 128) -> 512
c5 = self.fc2(c4) # 512 -> 7
c6 = F.softmax(c5, dim=1)
return c6
class CombNet(nn.Module):
""" Merge FeaturePyramidStructure, SubNet1, SubNet2 """
def __init__(self, device):
super(CombNet, self).__init__()
self.FPN = FeaturePyramidStructure()
self.sub1 = SubNet1()
self.sub2 = SubNet2()
self.FPN.to(device)
self.sub1.to(device)
self.sub2.to(device)
def forward(self, x):
topdown_1, f2, f3 = self.FPN(x)
topdown_2 = self.sub2(f2)
topdown_3 = self.sub1(f3)
return topdown_1, topdown_2, topdown_3
device = torch.device(3 if torch.cuda.is_available() else torch.device('cpu'))
combNet = CombNet(device)
combNet.to(device)
# For comparison with plain resnet-18
resnet18 = FeaturePyramidStructure()
resnet18.to(device)
combNet.train()
resnet18.train()
params = [p for p in combNet.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.01, momentum=0.9)
criterion = nn.CrossEntropyLoss()
params_res = [p for p in resnet18.parameters() if p.requires_grad]
optimizer_res = torch.optim.SGD(params_res, lr=0.01, momentum=0.9)
criterion_res = nn.CrossEntropyLoss()
# best
alpha = 0.5
beta = 0.7
# alpha = 0.1
# beta = 0.2
# For plotting
comb_log = []
res_log = []
for epoch in range(26):
for iter, (img, target) in enumerate(train_loader):
img = img.float().to(device)
img = img / 255
target = target.to(device)
optimizer.zero_grad()
optimizer_res.zero_grad()
loss1 = criterion(combNet(img)[0], torch.argmax(target, dim=1))
loss2 = criterion(combNet(img)[1], torch.argmax(target, dim=1))
loss3 = criterion(combNet(img)[2], torch.argmax(target, dim=1))
loss_res = criterion_res(resnet18(img)[0], torch.argmax(target, dim=1))
tloss = (loss1 + alpha * loss2 + beta * loss3)
tloss_res = loss_res
with torch.no_grad():
acc = torch.true_divide(torch.sum(torch.argmax(combNet(img)[0], dim=1) == torch.argmax(target, dim=1)),
len(target)) * 100
print('{}/{} combNet loss:{}, acc : {}'.format(epoch, iter, tloss, acc))
comb_log.append(acc)
acc_res = torch.true_divide(torch.sum(torch.argmax(resnet18(img)[0], dim=1) == torch.argmax(target, dim=1)),
len(target)) * 100
print(' resnet18 loss:{}, acc : {}'.format(tloss_res, acc_res))
res_log.append(acc_res)
tloss.backward()
tloss_res.backward()
optimizer.step()
optimizer_res.step()
plt.plot(np.arange(len(comb_log)), comb_log, label="combnet acc")
plt.plot(np.arange(len(res_log)), res_log, label="resnet18 acc")
plt.legend(loc='best')
plt.show()