-
Notifications
You must be signed in to change notification settings - Fork 1
/
vgg_attention.py
252 lines (219 loc) · 8.49 KB
/
vgg_attention.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
# -*- coding: utf-8 -*-
__author__ = "Pau Rodríguez López, ISELAB, CVC-UAB"
__email__ = "[email protected]"
import cv2
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.serialization import load_lua
from torchvision.utils import save_image
class VGG_16(nn.Module):
"""
Main Class
"""
def __init__(self, nlabels):
"""
Constructor
"""
super().__init__()
self.nlabels = nlabels
self.block_size = [2, 2, 3, 3, 3]
self.conv_1_1 = nn.Conv2d(3, 64, 3, stride=1, padding=1)
self.conv_1_2 = nn.Conv2d(64, 64, 3, stride=1, padding=1)
self.conv_2_1 = nn.Conv2d(64, 128, 3, stride=1, padding=1)
self.conv_2_2 = nn.Conv2d(128, 128, 3, stride=1, padding=1)
self.conv_3_1 = nn.Conv2d(128, 256, 3, stride=1, padding=1)
self.conv_3_2 = nn.Conv2d(256, 256, 3, stride=1, padding=1)
self.conv_3_3 = nn.Conv2d(256, 256, 3, stride=1, padding=1)
self.conv_4_1 = nn.Conv2d(256, 512, 3, stride=1, padding=1)
self.conv_4_2 = nn.Conv2d(512, 512, 3, stride=1, padding=1)
self.conv_4_3 = nn.Conv2d(512, 512, 3, stride=1, padding=1)
self.conv_5_1 = nn.Conv2d(512, 512, 3, stride=1, padding=1)
self.conv_5_2 = nn.Conv2d(512, 512, 3, stride=1, padding=1)
self.conv_5_3 = nn.Conv2d(512, 512, 3, stride=1, padding=1)
self.conv_attention = nn.Conv2d(512, 3, 1)
nn.init.kaiming_normal_(self.conv_attention.weight)
self.conv_proc_detail = nn.Conv2d(512, 512, 3, stride=1, padding=1)
nn.init.kaiming_normal_(self.conv_proc_detail.weight)
self.fc6 = nn.Linear(512 * 7 * 7 + 512, 4096)
nn.init.kaiming_normal_(self.fc6.weight)
self.fc7 = nn.Linear(4096, 4096)
nn.init.kaiming_normal_(self.fc7.weight)
self.fc8 = nn.Linear(4096, self.nlabels)
nn.init.kaiming_normal_(self.fc8.weight)
'''def load_weights(self, path="pretrained/VGG_FACE.t7"):
""" Function to load luatorch weights
Args:
path: path for the luatorch weights
"""
model = load_lua(path, unknown_classes=True)
counter = 1
block = 1
for i, layer in enumerate(model.modules):
if hasattr(layer, "weight"):
if block <= 5:
self_layer = getattr(self, "conv_%d_%d" % (block, counter))
counter += 1
if counter > self.block_size[block - 1]:
counter = 1
block += 1
self_layer.weight.data[...] = layer.weight.view_as(self_layer.weight)[...]
self_layer.bias.data[...] = layer.bias.view_as(self_layer.bias)[...]
# else:
# self_layer = getattr(self, "fc%d" % (block))
# block += 1
# self_layer.weight.data[...] = layer.weight.view_as(self_layer.weight)[...]
# self_layer.bias.data[...] = layer.bias.view_as(self_layer.bias)[...]
'''
def get_vgg_parameters(self):
""" Function to obtain the vgg pretrained parameters. Useful for freezing.
Returns: pre-trained parameters
"""
parameters = []
for block in self.block_size:
for num in range(block):
layer = getattr(self, "conv_%d_$d"(block + 1, num + 1))
parameters += list(layer.parameters())
return parameters
def get_att_parameters(self):
""" Function to obtain the attention parameters.
Returns: attention params.
"""
parameters = []
parameters += list(self.conv_attention.parameters())
parameters += list(self.conv_proc_detail.parameters())
parameters += list(self.fc6.parameters())
parameters += list(self.fc7.parameters())
parameters += list(self.fc8.parameters())
return parameters
def attend(self, x):
""" Computes the attention mask on the input images
Args:
x: input images
Returns: attention mask
"""
b, c, h, w = x.size()
self.input_size = (h, w)
x = F.relu(self.conv_1_1(x))
x = F.relu(self.conv_1_2(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv_2_1(x))
x = F.relu(self.conv_2_2(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv_3_1(x))
x = F.relu(self.conv_3_2(x))
x = F.relu(self.conv_3_3(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv_4_1(x))
x = F.relu(self.conv_4_2(x))
x = F.relu(self.conv_4_3(x))
self.pool4 = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv_5_1(self.pool4))
x = F.relu(self.conv_5_2(x))
x = F.relu(self.conv_5_3(x))
self.pool5 = F.max_pool2d(x, 2, 2)
b, c, h, w = self.pool5.size()
att = self.conv_attention(self.pool5)
#print(att.size())
#att = att.view()
image = att.data
image = image.detach().cpu().numpy()
img = image[0]
import numpy as np
img = np.transpose(img, (2, 1, 0))
print(img.shape)
#img = img * 255
from PIL import Image
from scipy.misc import imsave
img = Image.fromarray(img.astype('uint8'), 'RGB')
img = img.resize((64, 128), Image.ANTIALIAS)
imsave('./output/viz.png', img)
print(image.shape)
exit(0)
return F.softmax(att, -1).view(b, 1, h, w)
def crop(self, x, multiple=7):
""" Adjusts the high resolution image feature size to be multiple of 7
Args:
x: input features
multiple: multiple to adjust to
Returns: cropped feature map
"""
b, c, h, w = x.size()
h_ = h % multiple
w_ = w % multiple
return x[:, :, 0:(h - h_), 0:(w - w_)]
def reprocess(self, x):
""" Reprocesses high resolution images
Args:
x: input image
"""
b, c, h, w = x.size()
self.input_size = (h, w)
x = F.relu(self.conv_1_1(x))
x = F.relu(self.conv_1_2(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv_2_1(x))
x = F.relu(self.conv_2_2(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv_3_1(x))
x = F.relu(self.conv_3_2(x))
x = F.relu(self.conv_3_3(x))
x = F.max_pool2d(x, 2, 2)
x = F.relu(self.conv_4_1(x))
x = F.relu(self.conv_4_2(x))
x = F.relu(self.conv_4_3(x))
self.pool4 = F.max_pool2d(x, 2, 2)
def classify(self, x_high, attention):
""" Final classifier. Applies attention and outputs class logits.
Args:
x_high: high res image
attention: attention mask
Returns:
"""
x_high = self.conv_proc_detail(self.crop(x_high))
b, c, h, w = x_high.size()
x_high = x_high.view(b, c, 7, h // 7, 7, w // 7)
attended = x_high * attention.view(b, 1, 7, 1, 7, 1)
attended = F.normalize(attended.view(b, c, -1).sum(-1), 2, -1)
x_low = F.normalize(self.pool5.view(b, 512 * 7 * 7), 2, -1)
concat = torch.cat([attended, x_low], -1)
fc6 = F.relu(self.fc6(concat), True)
fc6 = F.dropout(fc6, 0.5, inplace=True)
fc7 = F.relu(self.fc7(fc6), True)
fc7 = F.dropout(fc7, 0.5, inplace=True)
return self.fc8(fc7)
def forward(self, x1):
""" Model forward function
Args:
x1: low res image
x2: [high res] image
Returns:
"""
b1, c1, h1, w1 = x1.size()
#b2, c2, h2, w2 = x2.size()
att = self.attend(x1)
#print(att.size())
##att = att.view(att.size(0), 3, 64, 128))
#save_image(denorm(att.data), './attention/vggnetd.png')
#exit(0)
#if (h1, w1) != (h2, w2):
self.reprocess(x1)
return self.classify(self.pool4, att)
if __name__ == "__main__":
import numpy as np
#im = cv2.imread('images/ak.png') - np.array([129.1863, 104.7624, 93.5940]).reshape((1, 1, 3))
#im2 = cv2.resize(im, (448, 448))
#im = im.transpose((1, 2, 0)).reshape((1, 3, 224, 224))
#im2 = im2.transpose((1, 2, 0)).reshape((1, 3, 448, 448))
#im = torch.Tensor(im).cuda()
#im2 = torch.Tensor(im2).cuda()
#model = VGG_16(nlabels=751).cuda()
from torchvision import models
model = models.vgg16(pretrained=True).cuda()
print(model)
exit(0)
#print(model(im, im).max())
#print(model(im, im2).max())
def denorm(x):
out = (x + 1) / 2
return out.clamp(0, 1)