-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel_analyzer.py
224 lines (182 loc) · 8.29 KB
/
model_analyzer.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
from base64 import encode
import torch
import torch.nn as nn
import torchvision.models as models
import torch.nn.functional as F
from PIL import Image
from torchvision import transforms
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import cv2
import numpy as np
from sklearn.preprocessing import minmax_scale
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
class ImgAttentionEncoder(nn.Module):
def __init__(self, embed_size):
super(ImgAttentionEncoder, self).__init__()
vggnet_feat = models.vgg19(pretrained=True).features
modules = list(vggnet_feat.children())[:-2]
self.cnn = nn.Sequential(*modules)
self.fc = nn.Sequential(nn.Linear(self.cnn[-3].out_channels, embed_size),
nn.Tanh())
def forward(self, image):
with torch.no_grad():
img_feature = self.cnn(image)
return img_feature
img_feature = img_feature.view(-1, 512, 196).transpose(1,2)
return img_feature
class ImgFeatureFusionEncoder2(nn.Module):
def __init__(self, embed_size,device):
super(ImgFeatureFusionEncoder2, self).__init__()
vggnet_feat = models.vgg19(pretrained=True).features
modules1 = list(vggnet_feat.children())[:18]
modules2 = list(vggnet_feat.children())[18:27]
modules3 = list(vggnet_feat.children())[27:-2]
self.cnn1 = nn.Sequential(*modules1)
self.cnn2 = nn.Sequential(*modules2)
self.cnn3 = nn.Sequential(*modules3)
self.device = device
self.fc = nn.Sequential(nn.Linear(self.cnn3[-3].out_channels, embed_size),
nn.Tanh())
def forward(self, image):
with torch.no_grad():
feature_layer_1 = self.cnn1(image) # 256,14,14
feature_layer_2 = self.cnn2(feature_layer_1) # 512,28,28
feature_layer_3 = self.cnn3(feature_layer_2) # 512,14,14
convoluted_1 = nn.Conv2d(256,512,1, bias=False).to(self.device)(feature_layer_1)
# pooled_layer1 = nn.MaxPool2d(4, stride=4)(convoluted_1).to(self.device)
# pooled_layer2 = nn.MaxPool2d(2, stride=2)(feature_layer_2)
pooled_layer1 = nn.AvgPool2d(4, stride=4)(convoluted_1).to(self.device)
pooled_layer2 = nn.AvgPool2d(2, stride=2)(feature_layer_2)
p_12_layer = torch.add(pooled_layer1,pooled_layer2)
fused_feature = torch.add(p_12_layer,feature_layer_3)
return fused_feature
img_feature = fused_feature.view(-1, 512, 196).transpose(1,2)
return img_feature
class ImgFeatureFusionEncoder(nn.Module):
def __init__(self, embed_size,device):
super(ImgFeatureFusionEncoder, self).__init__()
vggnet_feat = models.vgg19(pretrained=True).features
modules1 = list(vggnet_feat.children())[:18]
modules2 = list(vggnet_feat.children())[18:27]
modules3 = list(vggnet_feat.children())[27:-2]
self.cnn1 = nn.Sequential(*modules1)
self.cnn2 = nn.Sequential(*modules2)
self.cnn3 = nn.Sequential(*modules3)
self.device = device
self.fc = nn.Sequential(nn.Linear(self.cnn3[-3].out_channels, embed_size),
nn.Tanh())
def forward(self, image):
with torch.no_grad():
feature_layer_1 = self.cnn1(image)
feature_layer_2 = self.cnn2(feature_layer_1)
feature_layer_3 = self.cnn3(feature_layer_2)
convoluted_1 = nn.Conv2d(256,512,1, bias=False).to(self.device)(feature_layer_1)
# pooled_layer1 = nn.MaxPool2d(4, stride=4)(convoluted_1).to(self.device)
# pooled_layer2 = nn.MaxPool2d(2, stride=2)(feature_layer_2)
pooled_layer1 = nn.AvgPool2d(4, stride=4)(convoluted_1).to(self.device)
pooled_layer2 = nn.AvgPool2d(2, stride=2)(feature_layer_2)
p_12_layer = torch.add(pooled_layer1,pooled_layer2)
fused_feature = torch.add(p_12_layer,feature_layer_3)
return fused_feature
img_feature = fused_feature.view(-1, 512, 196).transpose(1,2)
return img_feature
class ImgFeatureFusionEncoder3(nn.Module):
def __init__(self, embed_size,device):
super(ImgFeatureFusionEncoder3, self).__init__()
vggnet_feat = models.vgg19(pretrained=True).features
modules1 = list(vggnet_feat.children())[:18]
modules2 = list(vggnet_feat.children())[18:27]
modules3 = list(vggnet_feat.children())[27:-2]
self.cnn1 = nn.Sequential(*modules1)
self.cnn2 = nn.Sequential(*modules2)
self.cnn3 = nn.Sequential(*modules3)
self.device = device
self.fc = nn.Sequential(nn.Linear(self.cnn3[-3].out_channels, embed_size),
nn.Tanh())
def forward(self, image):
with torch.no_grad():
feature_layer_1 = self.cnn1(image)
feature_layer_2 = self.cnn2(feature_layer_1)
feature_layer_3 = self.cnn3(feature_layer_2)
convoluted_1 = nn.Conv2d(256,512,1, bias=False).to(self.device)(feature_layer_1)
pooled_layer1 = nn.MaxPool2d(4, stride=4)(convoluted_1).to(self.device)
pooled_layer2 = nn.MaxPool2d(2, stride=2)(feature_layer_2)
# pooled_layer1 = nn.AvgPool2d(4, stride=4)(convoluted_1).to(self.device)
# pooled_layer2 = nn.AvgPool2d(2, stride=2)(feature_layer_2)
p_12_layer = torch.add(pooled_layer1,pooled_layer2)
fused_feature = torch.add(p_12_layer,feature_layer_3)
return fused_feature
img_feature = fused_feature.view(-1, 512, 196).transpose(1,2)
return img_feature
class SAN(nn.Module):
def __init__(self,attentionEncoder):
super(SAN, self).__init__()
self.img_encoder = attentionEncoder
# self.img_encoder = ImgFeatureFusionEncoder(embed_size,device)
def forward(self, img):
img_feature = self.img_encoder(img)
return img_feature
def image_loader(loader, image_name):
image = Image.open(image_name).convert('RGB')
image = loader(image).float()
image = torch.tensor(image, requires_grad=True)
image = image.unsqueeze(0)
return image
# data_transforms = transforms.Compose([transforms.ToTensor(),
# transforms.Normalize((0.485, 0.456, 0.406),
# (0.229, 0.224, 0.225))])
data_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406),
(0.229, 0.224, 0.225)),
])
def plot_image(image_tensor):
f, axarr = plt.subplots(23,23)
row = 0
col = 0
for i in range(image_tensor[0]):
axarr[row,col].imshow(image_tensor[i])
col += 1
if col == 23:
row += 1
col = 0
plt.show()
def train_model(encoder,image,name):
model_ft = SAN(attentionEncoder=encoder).to(device)
model_ft.eval()
x = model_ft(image)
x = x.detach().cpu().numpy()[0]
shape = x.shape
fig, axes = plt.subplots(23, 22)
plt.tight_layout()
x = x[:-6,:,:]
for idx, arch in enumerate(x):
i = (idx % 23)
j = (idx // 23)
shape = arch.shape
image_scaled = minmax_scale(arch.ravel(), feature_range=(0,1)).reshape(shape)
img = Image.fromarray(np.uint8(image_scaled * 255) , 'L')
img = np.asarray(img)
axes[i, j].imshow(img,cmap='gray',interpolation='none')
axes[i,j].axis('off')
axes[i,j].set_xticklabels([])
axes[i,j].set_yticklabels([])
plt.axis('off')
plt.subplots_adjust(wspace=0, hspace=0)
# plt.show()
# image_scaled = minmax_scale(x.ravel(), feature_range=(0,1)).reshape(shape)
# img = Image.fromarray(np.uint8(image_scaled * 255) , 'L')
# img.save("ResultImages/"+name+".PNG")
plt.savefig("FeatureImages/"+name+".png")
img = image_loader(data_transforms,"TestImage.jpg").to(device)
normal_encoder = ImgAttentionEncoder(1024)
fusion_encoder_avg = ImgFeatureFusionEncoder(1024,device=device)
fusion_encoder_max = ImgFeatureFusionEncoder3(1024,device);
fusion_encoder_concat = ImgFeatureFusionEncoder2(1024,device);
train_model(normal_encoder,img,"normal_encoder")
train_model(fusion_encoder_avg,img,"fusion_encoder_avg")
train_model(fusion_encoder_max,img,"fusion_encoder_max")
train_model(fusion_encoder_concat,img,"fusion_encoder_concat")