/
extract_res_fea.py
204 lines (179 loc) · 7.56 KB
/
extract_res_fea.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import *
import misc.utils as utils
from functools import reduce
import torchvision
from collections import OrderedDict
import argparse
from dataloaderp import *
from tqdm import tqdm
class Head(nn.Module):
'''
use res101's structure before block2
'''
def __init__(self, model='res101', freeze=True):
super(Head, self).__init__()
if model == 'res101':
base_model = torchvision.models.resnet101(
pretrained=True)
self.base_model = torch.nn.Sequential(OrderedDict([
('conv1', base_model.conv1),
('bn1', base_model.bn1),
('relu', base_model.relu),
('maxpool', base_model.maxpool),
('layer1', base_model.layer1),
]))
def forward(self, x):
out = self.base_model(x)
return out
class Extract(nn.Module):
def __init__(self, encoded_image_size=14, K=20, L=1024):
super(Extract, self).__init__()
self.head = Head(model='res101')
model = torchvision.models.resnet101(
pretrained=True) # pretrained ImageNet ResNet-101
self.features_model = torch.nn.Sequential(OrderedDict([
('layer2', model.layer2),
('layer3', model.layer3),
('layer4', model.layer4)
]))
self.adaptive_pool = nn.AdaptiveAvgPool2d(
(encoded_image_size, encoded_image_size))
# self.fine_tune2()
# self.my_resnet.eval()
del model
torch.cuda.empty_cache()
model = torchvision.models.resnet101(
pretrained=True) # pretrained ImageNet ResNet-101
self.miml_intermidate = torch.nn.Sequential(OrderedDict([
('layer2', model.layer2),
('layer3', model.layer3)]))
self.miml_last = torch.nn.Sequential(OrderedDict([
('layer4', model.layer4)]))
dim = 2048
map_size = 64
self.K = K
self.L = L
self.miml_sub_concept_layer = nn.Sequential(OrderedDict([
('conv1', nn.Conv2d(dim, 512, 1)),
('dropout1', nn.Dropout(0.5)), # (-1,512,14,14)
('conv2', nn.Conv2d(512, K*L, 1)),
# input need reshape to (-1,L,K,H*W)
('maxpool1', nn.MaxPool2d((K, 1))),
# reshape input to (-1,L,H*W), # permute(0,2,1)
('softmax1', nn.Softmax(dim=2)),
# permute(0,2,1) # reshape to (-1,L,1,H*W)
('maxpool2', nn.MaxPool2d((1, map_size)))
]))
del model
torch.cuda.empty_cache()
self.freeze()
self.freeze2()
def freeze(self):
for p in self.miml_intermidate.parameters():
p.requires_grad = False
for p in self.miml_last.parameters():
p.requires_grad = False
for p in self.miml_sub_concept_layer.parameters():
p.requires_grad = False
def freeze2(self):
for p in self.features_model.parameters():
p.requires_grad = False
def forward(self, images):
# Prepare the features
batch_size = images.shape[0]
head_out = self.head(images)
# miml
miml_features_out = self.miml_last(self.miml_intermidate(head_out))
# (-1,2048,8,8)
_, C, H, W = miml_features_out.shape
conv1_out = self.miml_sub_concept_layer.dropout1(
self.miml_sub_concept_layer.conv1(miml_features_out))
# shape -> (n_bags, (L * K), n_instances, 1)
conv2_out = self.miml_sub_concept_layer.conv2(conv1_out)
# shape -> (n_bags, L, K, n_instances)
conv2_out = conv2_out.reshape(-1, self.L, self.K, H*W)
# shape -> (n_bags, L, 1, n_instances),remove dim: 1
maxpool1_out = self.miml_sub_concept_layer.maxpool1(
conv2_out).squeeze(2)
# softmax
permute1 = maxpool1_out.permute(0, 2, 1)
softmax1 = self.miml_sub_concept_layer.softmax1(permute1)
permute2 = softmax1.permute(0, 2, 1)
# reshape = permute2.unsqueeze(2)
# predictions_instancelevel
reshape = permute2.reshape(-1, self.L, 1, H*W)
# shape -> (n_bags, L, 1, 1)
maxpool2_out = self.miml_sub_concept_layer.maxpool2(reshape)
attributes = maxpool2_out.squeeze()
# extract image features
# (batch_size, 2048, image_size/32, image_size/32)
images_features = self.features_model(head_out)
# (batch_size, 2048, encoded_image_size, encoded_image_size)
imgs_features = self.adaptive_pool(images_features)
# (batch_size, encoded_image_size, encoded_image_size, 2048)
imgs_features = imgs_features.permute(
0, 2, 3, 1).reshape(batch_size, -1, 2048)
return attributes, imgs_features
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# Data input settings
parser.add_argument('--input_json', type=str, default='/home/lkk/code/self-critical.pytorch/data/cocotalk.json',
help='path to the json file containing additional info and vocab')
parser.add_argument('--num_worker', type=int, default=0,
help='')
parser.add_argument('--input_sg_dir', type=str, default='/home/lkk/code/self-critical.pytorch/data/coco_img_sg',
help='scene graph')
parser.add_argument('--input_sg_voc', type=str, default='/home/lkk/code/self-critical.pytorch/data/coco_pred_sg_rela.npy',
help='scene graph voc')
parser.add_argument('--input_label_h5', type=str,
default='/home/lkk/code/self-critical.pytorch/data/cocotalk_label.h5')
parser.add_argument('--batch_size', type=int,
default=16)
parser.add_argument('--train_only', type=int, default=0,
help='if true then use 80k, else use 110k')
opt = parser.parse_args()
loader = DataLoaderRaw(opt)
model = Extract().cuda()
model.eval()
for i in tqdm(range(7081)):
data = loader.get_batch('train')
attrs, imgs = model(data['img'].cuda())
attrs, imgs = attrs.detach().cpu().numpy(), imgs.detach().cpu().numpy()
infos = data['infos']
for j in range(len(infos)):
name = str(infos[j]['id'])
attr = attrs[j]
img = imgs[j]
path = os.path.join(
'/home/lkk/code/self-critical.pytorch/data/cocoresnet', name)
np.savez(path, attr=attr, img=img)
for i in tqdm(range(313)):
data = loader.get_batch('val')
attrs, imgs = model(data['img'].cuda())
attrs, imgs = attrs.detach().cpu().numpy(), imgs.detach().cpu().numpy()
infos = data['infos']
for j in range(len(infos)):
name = str(infos[j]['id'])
attr = attrs[j]
img = imgs[j]
path = os.path.join(
'/home/lkk/code/self-critical.pytorch/data/cocoresnet', name)
np.savez(path, attr=attr, img=img)
for i in tqdm(range(313)):
data = loader.get_batch('test')
attrs, imgs = model(data['img'].cuda())
attrs, imgs = attrs.detach().cpu().numpy(), imgs.detach().cpu().numpy()
infos = data['infos']
for j in range(len(infos)):
name = str(infos[j]['id'])
attr = attrs[j]
img = imgs[j]
path = os.path.join(
'/home/lkk/code/self-critical.pytorch/data/cocoresnet', name)
np.savez(path, attr=attr, img=img)