-
Notifications
You must be signed in to change notification settings - Fork 3
/
EXTD_cpu.py
184 lines (156 loc) · 6.36 KB
/
EXTD_cpu.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
#-*- coding:utf-8 -*-
from __future__ import division
from __future__ import absolute_import
from __future__ import print_function
import os
import torch
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as F
from torch.autograd import Variable
from layers import *
from data.config_EXTD import cfg
import numpy as np
class feature_extractor(nn.Module):
def __init__(self,C_out):
super(feature_extractor, self).__init__()
self.E = nn.Sequential(nn.Conv2d(3,C_out,3,2,1),
nn.BatchNorm2d(C_out))
def forward(self,x):
out = self.E(x)
out = F.relu(out, inplace=True)
return out
class inverted_residual_1(nn.Module):
def __init__(self,C_in,C_out):
super(inverted_residual_1, self).__init__()
self.IR1 = nn.Sequential(nn.Conv2d(C_in,C_in,3,1,padding=1,groups=C_in),
nn.BatchNorm2d(C_in),
nn.PReLU(),
nn.Conv2d(C_in,C_out,1,1,padding=0,groups=1),
nn.BatchNorm2d(C_out))
def forward(self,x):
return self.IR1(x)
class inverted_residual_2(nn.Module):
def __init__(self,C_in,h,stride=2):
super(inverted_residual_2, self).__init__()
self.IR2 = nn.Sequential(nn.Conv2d(C_in,h,1,1,padding=0,groups=1),
nn.BatchNorm2d(h),
nn.PReLU(),
nn.Conv2d(h,h,3,stride,padding=1,groups=h),
nn.BatchNorm2d(h),
nn.PReLU(),
nn.Conv2d(h,C_in,1,1,padding=0,groups=1),
nn.BatchNorm2d(C_in),
)
def forward(self,x):
return self.IR2(x)
class upsample(nn.Module):
def __init__(self, C_in, filters = 64):
super(upsample, self).__init__()
self.upsample = nn.Sequential(nn.Upsample(scale_factor=2, mode='bilinear'),
nn.Conv2d(C_in,filters,3,1,1,groups=C_in),
nn.Conv2d(filters,C_in,1,1,0,groups=1),
nn.BatchNorm2d(C_in),
nn.ReLU(inplace=True))
def forward(self,x):
return self.upsample(x)
class backbone(nn.Module):
def __init__(self,filters = 64):
super(backbone, self).__init__()
self.blocks = nn.Sequential(inverted_residual_1(filters,filters),
inverted_residual_2(filters,filters*2,1),
inverted_residual_2(filters,filters*2,1),
inverted_residual_2(filters,filters*2,1),
inverted_residual_2(filters,filters*2,2))
def forward(self, x):
return self.blocks(x)
def class_predictor(C_in,filters=2):
return nn.Conv2d(C_in,filters,3,1,1)
def bbox_predictor(C_in):
return nn.Conv2d(C_in,4,3,1,1)
class EXTD(nn.Module):
def __init__(self,phase):
super(EXTD, self).__init__()
self.phase = phase
self.num_classes = 2
self.E = feature_extractor(64)
self.F = backbone()
up,cls_pred,bbox_pred = [],[],[]
for i in range(5):
up.append(upsample(64))
self.up = nn.ModuleList(up)
for i in range(6):
if(i==5):
#maxout
cls_pred.append(class_predictor(64,4))
else:
cls_pred.append(class_predictor(64))
bbox_pred.append(bbox_predictor(64))
self.cls_pred, self.bbox_pred = nn.ModuleList(cls_pred), nn.ModuleList(bbox_pred)
if self.phase == 'test':
self.detect = Detect(cfg)
self.softmax = nn.Softmax(dim=-1)
def forward(self,x):
size = x.size()[2:]
x = self.E(x)
f,g = [],[]
#SSD
for i in range(6):
x = self.F(x)
f.append(x)
#FPN
g.append(self.up[0](f[-1])+f[-2])
for i in range(4):
g.append(self.up[i+1](g[-1])+f[-3-i])
#head
conf, loc, feature_maps = list(),list(), [f[-1]]+g
for i in range(6):
feature_map = feature_maps[i]
if(i==5):
#maxout
conf_x = self.cls_pred[i](feature_map).permute(0, 2, 3, 1).contiguous()
max_conf, _ = torch.max(conf_x[:,:, :, 0:3], dim=3, keepdim=True)
conf_x = torch.cat((max_conf, conf_x[:, :, :, 3:]), dim=3)
conf.append(conf_x)
else:
conf.append(self.cls_pred[i](feature_map).permute(0, 2, 3, 1).contiguous())
loc.append(self.bbox_pred[i](feature_map).permute(0, 2, 3, 1).contiguous())
features_maps_size = []
for i in range(len(loc)):
feat = []
feat += [loc[i].size(1), loc[i].size(2)]
features_maps_size += [feat]
self.priorbox = PriorBox(size, features_maps_size,cfg)
with torch.no_grad():
self.priors = Variable(self.priorbox.forward())
loc = torch.cat([o.view(o.size(0), -1) for o in loc], 1)
conf = torch.cat([o.view(o.size(0), -1) for o in conf], 1)
#print(self.priors[0], self.priors[-1])
if self.phase == 'test':
output = self.detect(
loc.view(loc.size(0), -1, 4), # loc preds
self.softmax(conf.view(conf.size(0), -1,self.num_classes)), # conf preds
self.priors.type(type(x.data)) # default boxes
)
else:
output = (
loc.view(loc.size(0), -1, 4),
conf.view(conf.size(0), -1, self.num_classes),
self.priors
)
return output
def xavier(self, param):
init.xavier_uniform(param)
def weights_init(self, m):
if isinstance(m, nn.Conv2d):
self.xavier(m.weight.data)
m.bias.data.zero_()
if isinstance(m,nn.ConvTranspose2d):
self.xavier(m.weight.data)
if 'bias' in m.state_dict().keys():
m.bias.data.zero_()
if isinstance(m,nn.BatchNorm2d):
m.weight.data[...] = 1
m.bias.data.zero_()
def build_extd(phase, num_classes=2):
return EXTD(phase)