dpn网络的pytorch实现方式

时间:2022-09-14 07:57:28

我就废话不多说了,直接上代码吧!

?
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
253
254
255
256
257
258
259
260
261
import torch
import torch.nn as nn
import torch.nn.functional as F
 
 
 
class CatBnAct(nn.Module):
 def __init__(self, in_chs, activation_fn=nn.ReLU(inplace=True)):
  super(CatBnAct, self).__init__()
  self.bn = nn.BatchNorm2d(in_chs, eps=0.001)
  self.act = activation_fn
 
 def forward(self, x):
  x = torch.cat(x, dim=1) if isinstance(x, tuple) else x
  return self.act(self.bn(x))
 
 
class BnActConv2d(nn.Module):
 def __init__(self, s, out_chs, kernel_size, stride,
     padding=0, groups=1, activation_fn=nn.ReLU(inplace=True)):
  super(BnActConv2d, self).__init__()
  self.bn = nn.BatchNorm2d(in_chs, eps=0.001)
  self.act = activation_fn
  self.conv = nn.Conv2d(in_chs, out_chs, kernel_size, stride, padding, groups=groups, bias=False)
 
 def forward(self, x):
  return self.conv(self.act(self.bn(x)))
 
 
class InputBlock(nn.Module):
 def __init__(self, num_init_features, kernel_size=7,
     padding=3, activation_fn=nn.ReLU(inplace=True)):
  super(InputBlock, self).__init__()
  self.conv = nn.Conv2d(
   3, num_init_features, kernel_size=kernel_size, stride=2, padding=padding, bias=False)
  self.bn = nn.BatchNorm2d(num_init_features, eps=0.001)
  self.act = activation_fn
  self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
 
 def forward(self, x):
  x = self.conv(x)
  x = self.bn(x)
  x = self.act(x)
  x = self.pool(x)
  return x
 
 
class DualPathBlock(nn.Module):
 def __init__(
   self, in_chs, num_1x1_a, num_3x3_b, num_1x1_c, inc, groups, block_type='normal', b=False):
  super(DualPathBlock, self).__init__()
  self.num_1x1_c = num_1x1_c
  self.inc = inc
  self.b = b
  if block_type is 'proj':
   self.key_stride = 1
   self.has_proj = True
  elif block_type is 'down':
   self.key_stride = 2
   self.has_proj = True
  else:
   assert block_type is 'normal'
   self.key_stride = 1
   self.has_proj = False
 
  if self.has_proj:
   # Using different member names here to allow easier parameter key matching for conversion
   if self.key_stride == 2:
    self.c1x1_w_s2 = BnActConv2d(
     in_chs=in_chs, out_chs=num_1x1_c + 2 * inc, kernel_size=1, stride=2)
   else:
    self.c1x1_w_s1 = BnActConv2d(
     in_chs=in_chs, out_chs=num_1x1_c + 2 * inc, kernel_size=1, stride=1)
  self.c1x1_a = BnActConv2d(in_chs=in_chs, out_chs=num_1x1_a, kernel_size=1, stride=1)
  self.c3x3_b = BnActConv2d(
   in_chs=num_1x1_a, out_chs=num_3x3_b, kernel_size=3,
   stride=self.key_stride, padding=1, groups=groups)
  if b:
   self.c1x1_c = CatBnAct(in_chs=num_3x3_b)
   self.c1x1_c1 = nn.Conv2d(num_3x3_b, num_1x1_c, kernel_size=1, bias=False)
   self.c1x1_c2 = nn.Conv2d(num_3x3_b, inc, kernel_size=1, bias=False)
  else:
   self.c1x1_c = BnActConv2d(in_chs=num_3x3_b, out_chs=num_1x1_c + inc, kernel_size=1, stride=1)
 
 def forward(self, x):
  x_in = torch.cat(x, dim=1) if isinstance(x, tuple) else x
  if self.has_proj:
   if self.key_stride == 2:
    x_s = self.c1x1_w_s2(x_in)
   else:
    x_s = self.c1x1_w_s1(x_in)
   x_s1 = x_s[:, :self.num_1x1_c, :, :]
   x_s2 = x_s[:, self.num_1x1_c:, :, :]
  else:
   x_s1 = x[0]
   x_s2 = x[1]
  x_in = self.c1x1_a(x_in)
  x_in = self.c3x3_b(x_in)
  if self.b:
   x_in = self.c1x1_c(x_in)
   out1 = self.c1x1_c1(x_in)
   out2 = self.c1x1_c2(x_in)
  else:
   x_in = self.c1x1_c(x_in)
   out1 = x_in[:, :self.num_1x1_c, :, :]
   out2 = x_in[:, self.num_1x1_c:, :, :]
  resid = x_s1 + out1
  dense = torch.cat([x_s2, out2], dim=1)
  return resid, dense
 
 
class DPN(nn.Module):
 def __init__(self, small=False, num_init_features=64, k_r=96, groups=32,
     b=False, k_sec=(3, 4, 20, 3), inc_sec=(16, 32, 24, 128),
     num_classes=1000, test_time_pool=False):
  super(DPN, self).__init__()
  self.test_time_pool = test_time_pool
  self.b = b
  bw_factor = 1 if small else 4
 
  blocks = OrderedDict()
 
  # conv1
  if small:
   blocks['conv1_1'] = InputBlock(num_init_features, kernel_size=3, padding=1)
  else:
   blocks['conv1_1'] = InputBlock(num_init_features, kernel_size=7, padding=3)
 
  # conv2
  bw = 64 * bw_factor
  inc = inc_sec[0]
  r = (k_r * bw) // (64 * bw_factor)
  blocks['conv2_1'] = DualPathBlock(num_init_features, r, r, bw, inc, groups, 'proj', b)
  in_chs = bw + 3 * inc
  for i in range(2, k_sec[0] + 1):
   blocks['conv2_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
   in_chs += inc
 
  # conv3
  bw = 128 * bw_factor
  inc = inc_sec[1]
  r = (k_r * bw) // (64 * bw_factor)
  blocks['conv3_1'] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'down', b)
  in_chs = bw + 3 * inc
  for i in range(2, k_sec[1] + 1):
   blocks['conv3_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
   in_chs += inc
 
  # conv4
  bw = 256 * bw_factor
  inc = inc_sec[2]
  r = (k_r * bw) // (64 * bw_factor)
  blocks['conv4_1'] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'down', b)
  in_chs = bw + 3 * inc
  for i in range(2, k_sec[2] + 1):
   blocks['conv4_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
   in_chs += inc
 
  # conv5
  bw = 512 * bw_factor
  inc = inc_sec[3]
  r = (k_r * bw) // (64 * bw_factor)
  blocks['conv5_1'] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'down', b)
  in_chs = bw + 3 * inc
  for i in range(2, k_sec[3] + 1):
   blocks['conv5_' + str(i)] = DualPathBlock(in_chs, r, r, bw, inc, groups, 'normal', b)
   in_chs += inc
  blocks['conv5_bn_ac'] = CatBnAct(in_chs)
 
  self.features = nn.Sequential(blocks)
 
  # Using 1x1 conv for the FC layer to allow the extra pooling scheme
  self.last_linear = nn.Conv2d(in_chs, num_classes, kernel_size=1, bias=True)
 
 def logits(self, features):
  if not self.training and self.test_time_pool:
   x = F.avg_pool2d(features, kernel_size=7, stride=1)
   out = self.last_linear(x)
   # The extra test time pool should be pooling an img_size//32 - 6 size patch
   out = adaptive_avgmax_pool2d(out, pool_type='avgmax')
  else:
   x = adaptive_avgmax_pool2d(features, pool_type='avg')
   out = self.last_linear(x)
  return out.view(out.size(0), -1)
 
 def forward(self, input):
  x = self.features(input)
  x = self.logits(x)
  return x
 
""" PyTorch selectable adaptive pooling
Adaptive pooling with the ability to select the type of pooling from:
 * 'avg' - Average pooling
 * 'max' - Max pooling
 * 'avgmax' - Sum of average and max pooling re-scaled by 0.5
 * 'avgmaxc' - Concatenation of average and max pooling along feature dim, doubles feature dim
 
Both a functional and a nn.Module version of the pooling is provided.
 
"""
 
def pooling_factor(pool_type='avg'):
 return 2 if pool_type == 'avgmaxc' else 1
 
 
def adaptive_avgmax_pool2d(x, pool_type='avg', padding=0, count_include_pad=False):
 """Selectable global pooling function with dynamic input kernel size
 """
 if pool_type == 'avgmaxc':
  x = torch.cat([
   F.avg_pool2d(
    x, kernel_size=(x.size(2), x.size(3)), padding=padding, count_include_pad=count_include_pad),
   F.max_pool2d(x, kernel_size=(x.size(2), x.size(3)), padding=padding)
  ], dim=1)
 elif pool_type == 'avgmax':
  x_avg = F.avg_pool2d(
    x, kernel_size=(x.size(2), x.size(3)), padding=padding, count_include_pad=count_include_pad)
  x_max = F.max_pool2d(x, kernel_size=(x.size(2), x.size(3)), padding=padding)
  x = 0.5 * (x_avg + x_max)
 elif pool_type == 'max':
  x = F.max_pool2d(x, kernel_size=(x.size(2), x.size(3)), padding=padding)
 else:
  if pool_type != 'avg':
   print('Invalid pool type %s specified. Defaulting to average pooling.' % pool_type)
  x = F.avg_pool2d(
   x, kernel_size=(x.size(2), x.size(3)), padding=padding, count_include_pad=count_include_pad)
 return x
 
 
class AdaptiveAvgMaxPool2d(torch.nn.Module):
 """Selectable global pooling layer with dynamic input kernel size
 """
 def __init__(self, output_size=1, pool_type='avg'):
  super(AdaptiveAvgMaxPool2d, self).__init__()
  self.output_size = output_size
  self.pool_type = pool_type
  if pool_type == 'avgmaxc' or pool_type == 'avgmax':
   self.pool = nn.ModuleList([nn.AdaptiveAvgPool2d(output_size), nn.AdaptiveMaxPool2d(output_size)])
  elif pool_type == 'max':
   self.pool = nn.AdaptiveMaxPool2d(output_size)
  else:
   if pool_type != 'avg':
    print('Invalid pool type %s specified. Defaulting to average pooling.' % pool_type)
   self.pool = nn.AdaptiveAvgPool2d(output_size)
 
 def forward(self, x):
  if self.pool_type == 'avgmaxc':
   x = torch.cat([p(x) for p in self.pool], dim=1)
  elif self.pool_type == 'avgmax':
   x = 0.5 * torch.sum(torch.stack([p(x) for p in self.pool]), 0).squeeze(dim=0)
  else:
   x = self.pool(x)
  return x
 
 def factor(self):
  return pooling_factor(self.pool_type)
 
 def __repr__(self):
  return self.__class__.__name__ + ' (' \
    + 'output_size=' + str(self.output_size) \
    + ', pool_type=' + self.pool_type + ')'

以上这篇dpn网络的pytorch实现方式就是小编分享给大家的全部内容了,希望能给大家一个参考,也希望大家多多支持服务器之家。

原文链接:https://blog.csdn.net/weixin_40123108/article/details/89682449