-
Notifications
You must be signed in to change notification settings - Fork 8
/
model.py
224 lines (176 loc) · 8.71 KB
/
model.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
import torch
import torch.nn as nn
import torch.cuda
import torchvision.models as models
from configs import device_ids, dataset_train, dataset_eval, nn_type, vc_num, K, vMF_kappa, context_cluster, layer, meta_dir, categories, feature_num, rpn_configs
from configs import *
from Net2 import Net
from RPN import RegionProposalNetwork
def vgg16(layer):
net = models.vgg16(pretrained=True)
if layer == 'pool5':
num_layers = 31
elif layer == 'pool4':
num_layers = 24
elif layer == 'pool3':
num_layers = 17
model = nn.Sequential()
features = nn.Sequential()
for i in range(0, num_layers):
features.add_module('{}'.format(i), net.features[i])
model.add_module('features', features)
return model
def resnext(layer):
extractor = nn.Sequential()
net = models.resnext50_32x4d(pretrained=True)
if layer == 'last':
extractor.add_module('0', net.conv1)
extractor.add_module('1', net.bn1)
extractor.add_module('2', net.relu)
extractor.add_module('3', net.maxpool)
extractor.add_module('4', net.layer1)
extractor.add_module('5', net.layer2)
extractor.add_module('6', net.layer3)
extractor.add_module('7', net.layer4)
elif layer == 'second':
extractor.add_module('0', net.conv1)
extractor.add_module('1', net.bn1)
extractor.add_module('2', net.relu)
extractor.add_module('3', net.maxpool)
extractor.add_module('4', net.layer1)
extractor.add_module('5', net.layer2)
extractor.add_module('6', net.layer3)
else:
extractor = []
return extractor
# return backbone extractor based on nn_type and layer in configs
def get_backbone_extractor():
if nn_type == 'vgg':
return vgg16(layer).cuda(device_ids[0])
if nn_type == 'resnext':
return resnext(layer).cuda(device_ids[0]).eval()
print('Failed to get backbone extractor')
# return visual concept centers
def get_vc():
vc = np.zeros((vc_num, feature_num))
try:
if nn_type == 'vgg':
vc = np.load(meta_dir + 'init_{}/dictionary_vgg/dictionary_{}.pickle'.format(nn_type, layer), allow_pickle=True)
elif nn_type == 'resnext':
vc = np.load(meta_dir + 'init_{}/dictionary_resnext_pascal3d+/dictionary_{}_512_kap65.pickle'.format(nn_type, layer), allow_pickle=True)
except:
print('Failed to load VC')
vc = vc[:, :, np.newaxis, np.newaxis]
vc = torch.from_numpy(vc).type(torch.FloatTensor)
return vc.cuda(device_ids[0])
# return context cluster centers
def get_context():
context = np.zeros((0, feature_num))
for category in categories['train']:
try:
context = np.concatenate((context, np.load(meta_dir + 'init_{}/context_kernel_{}/{}_{}.npy'.format(nn_type, layer, category, context_cluster))), axis=0)
except:
print('Failed to load Context Kernels')
continue
context = context[:, :, np.newaxis, np.newaxis]
context = torch.from_numpy(context).type(torch.FloatTensor)
return context.cuda(device_ids[0])
def get_clutter_models():
clutter = np.zeros((0, vc_num))
try:
if nn_type == 'vgg':
clutter = np.load(meta_dir + 'init_{}/CLUTTER_MODEL_POOL4.pkl'.format(nn_type, nn_type, layer), allow_pickle=True)
for i in range(clutter.shape[0]):
clutter[i] = clutter[i] / clutter[i].sum()
elif nn_type == 'resnext':
for suf in ['_general', '_ijcv']: # the first clutter is the general one used for classification, the rest is used for segmentation
clutter = np.concatenate((clutter, np.load( meta_dir + 'init_{}/{}_{}_clutter_model{}.npy'.format(nn_type, nn_type, layer, suf)) ), axis=0)
except:
print('Failed to load Clutter Models')
clutter = clutter[:, :, np.newaxis, np.newaxis]
clutter = torch.from_numpy(clutter).type(torch.FloatTensor)
return clutter
def get_mixture_models(dim_reduction=True, tag='_it2'):
FG_Models = []
FG_prior = []
CNTXT_Models = []
CNTXT_prior = []
for category in categories['train']:
try:
load_path = meta_dir + 'init_{}/mix_model_vmf_pascal3d+_EM_all_context{}/mmodel_{}_K{}_FEATDIM512_{}_specific_view_{}.pickle'.format(nn_type, tag, category, K, layer, context_cluster)
alpha, beta, prior = np.load(load_path, allow_pickle=True)
except:
print('Failed to load Mixture Model: {}'.format(category.upper()))
FG_Models.append(None)
FG_prior.append(None)
CNTXT_Models.append(None)
CNTXT_prior.append(None)
continue
mix_fg = np.array(alpha)
mix_context = np.array(beta)
prior_fg = np.array(prior)[:, 0, :, :]
prior_context = np.array(prior)[:, 1, :, :]
# Reduce dimensions of the mixture model since most of the boundary regions only sampled one or two images during model building
if dim_reduction:
old_dim = prior_fg.shape
prior_whole = prior_fg + prior_context
h_cut=1
w_cut=1
while np.sum(prior_whole[:, h_cut:-h_cut, :].reshape(-1, 1)) / np.sum(prior_whole.reshape(-1, 1)) > 0.995:
h_cut += 1
while np.sum(prior_whole[:, :, w_cut:-w_cut].reshape(-1, 1)) / np.sum(prior_whole.reshape(-1, 1)) > 0.995:
w_cut += 1
mix_fg = mix_fg[:, :, w_cut:-w_cut, h_cut:-h_cut]
mix_context = mix_context[:, :, w_cut:-w_cut, h_cut:-h_cut]
prior_fg = prior_fg[:, h_cut:-h_cut, w_cut:-w_cut]
prior_context = prior_context[:, h_cut:-h_cut, w_cut:-w_cut]
new_dim = prior_fg.shape
print('Dim Reduction - {}: ({}, {}) --> ({}, {})'.format(category, old_dim[1], old_dim[2], new_dim[1], new_dim[2]))
mix_fg = np.transpose(mix_fg, [0, 1, 3, 2])
mix_context = np.transpose(mix_context, [0, 1, 3, 2])
# dealing with empty kernels mix_fg.shape = [8, 512, H, W]
mix_fg = np.transpose(mix_fg, [2, 3, 0, 1]) # mix_fg.shape = [H, W, 8, 512]
zero_map = (np.sum(mix_fg, axis=3) == 0)
vc_num = mix_fg.shape[3]
avg_feature = mix_fg.reshape(-1, vc_num).sum(0)
avg_feature = avg_feature / np.sum(avg_feature)
mix_fg[zero_map] = avg_feature
mix_fg = np.transpose(mix_fg, [2, 3, 0, 1])
#dealing with empty kernels mix_context.shape = [8, 512, H, W]
mix_context = np.transpose(mix_context, [2, 3, 0, 1]) # mix_context.shape = [H, W, 8, 512]
zero_map = (np.sum(mix_context, axis=3) == 0)
vc_num = mix_context.shape[3]
avg_feature = mix_context.reshape(-1, vc_num).sum(0)
avg_feature = avg_feature / np.sum(avg_feature)
mix_context[zero_map] = avg_feature
mix_context = np.transpose(mix_context, [2, 3, 0, 1])
# dealing with empty kernels prior_fg.shape = [8, H, W]
prior_fg[prior_fg == 0] = np.min(prior_fg[prior_fg > 0])
# dealing with empty kernels prior_context.shape = [8, H, W]
prior_context[prior_context == 0] = np.min(prior_context[prior_context > 0])
mix_fg = torch.from_numpy(mix_fg).type(torch.FloatTensor)
FG_Models.append(mix_fg.cuda(device_ids[0]))
mix_context = torch.from_numpy(mix_context).type(torch.FloatTensor)
CNTXT_Models.append(mix_context.cuda(device_ids[0]))
prior_fg = torch.from_numpy(prior_fg).type(torch.FloatTensor)
FG_prior.append(prior_fg.cuda(device_ids[0]))
prior_context = torch.from_numpy(prior_context).type(torch.FloatTensor)
CNTXT_prior.append(prior_context.cuda(device_ids[0]))
return [FG_Models, FG_prior, CNTXT_Models, CNTXT_prior]
# generate and return the entire compnet architecture
def get_compnet_head(mix_model_dim_reduction=True, mix_model_suffix=''):
net = Net(Feature_Extractor=get_backbone_extractor(),
VC_Centers=get_vc(),
Context_Kernels=get_context(),
Mixture_Models=get_mixture_models(dim_reduction=mix_model_dim_reduction, tag=mix_model_suffix),
Clutter_Models=get_clutter_models(),
vMF_kappa=vMF_kappa)
return net.cuda(device_ids[0])
# generate and return the entire rpn architecture
def get_rpn():
rpn = RegionProposalNetwork(in_channels=feature_num,
mid_channels=feature_num,
ratios=rpn_configs['ratios'],
anchor_scales=rpn_configs['anchor_scales'],
feat_stride=rpn_configs['feat_stride'])
return rpn.cuda(device_ids[0])