質問編集履歴
3
ご回答に対して修正した部分をのせました。
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -207,3 +207,59 @@
|
|
|
207
207
|
https://github.com/owruby/shake-shake_pytorch
|
|
208
208
|
|
|
209
209
|
から借りています。
|
|
210
|
+
|
|
211
|
+
|
|
212
|
+
|
|
213
|
+
|
|
214
|
+
|
|
215
|
+
ご指摘の点について
|
|
216
|
+
|
|
217
|
+
ShakeBlock内のforwardを
|
|
218
|
+
|
|
219
|
+
```
|
|
220
|
+
|
|
221
|
+
def forward(self, x, y):
|
|
222
|
+
|
|
223
|
+
h1 = self.branch1(x)
|
|
224
|
+
|
|
225
|
+
h2 = self.branch2(x)
|
|
226
|
+
|
|
227
|
+
h = ShakeShake.apply(h1, h2, self.training)
|
|
228
|
+
|
|
229
|
+
h0 = x if self.equal_io else self.shortcut(x)
|
|
230
|
+
|
|
231
|
+
|
|
232
|
+
|
|
233
|
+
return h + h0, h + h0
|
|
234
|
+
|
|
235
|
+
```
|
|
236
|
+
|
|
237
|
+
ShakeResNet内のforwardを
|
|
238
|
+
|
|
239
|
+
```
|
|
240
|
+
|
|
241
|
+
def forward(self, x):
|
|
242
|
+
|
|
243
|
+
h = self.c_in(x)
|
|
244
|
+
|
|
245
|
+
h, h = self.layer1(h, h)
|
|
246
|
+
|
|
247
|
+
h, h = self.layer2(h, h)
|
|
248
|
+
|
|
249
|
+
h, h = self.layer3(h, h)
|
|
250
|
+
|
|
251
|
+
h = F.relu(h)
|
|
252
|
+
|
|
253
|
+
h = F.avg_pool2d(h, 8)
|
|
254
|
+
|
|
255
|
+
h = h.view(-1, self.in_chs[3])
|
|
256
|
+
|
|
257
|
+
h = self.fc_out(h)
|
|
258
|
+
|
|
259
|
+
|
|
260
|
+
|
|
261
|
+
return h
|
|
262
|
+
|
|
263
|
+
```
|
|
264
|
+
|
|
265
|
+
とそれぞれ変更しましたが、同じエラーが出ます。
|
2
コード元のアドレスを追加しました。
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -199,3 +199,11 @@
|
|
|
199
199
|
|
|
200
200
|
|
|
201
201
|
というエラーが出ます。
|
|
202
|
+
|
|
203
|
+
|
|
204
|
+
|
|
205
|
+
コードは
|
|
206
|
+
|
|
207
|
+
https://github.com/owruby/shake-shake_pytorch
|
|
208
|
+
|
|
209
|
+
から借りています。
|
1
コードを追加しました。
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -31,3 +31,171 @@
|
|
|
31
31
|
というエラーが出ます。
|
|
32
32
|
|
|
33
33
|
forwardの引数は、必ず(self, x)のままでないといけないのでしょうか？
|
|
34
|
+
|
|
35
|
+
|
|
36
|
+
|
|
37
|
+
```
|
|
38
|
+
|
|
39
|
+
class ShakeBlock(nn.Module):
|
|
40
|
+
|
|
41
|
+
def __init__(self, in_ch, out_ch, stride=1):
|
|
42
|
+
|
|
43
|
+
super(ShakeBlock, self).__init__()
|
|
44
|
+
|
|
45
|
+
self.equal_io = in_ch == out_ch
|
|
46
|
+
|
|
47
|
+
self.shortcut = self.equal_io and None or Shortcut(in_ch, out_ch, stride=stride)
|
|
48
|
+
|
|
49
|
+
|
|
50
|
+
|
|
51
|
+
self.branch1 = self._make_branch(in_ch, out_ch, stride)
|
|
52
|
+
|
|
53
|
+
self.branch2 = self._make_branch(in_ch, out_ch, stride)
|
|
54
|
+
|
|
55
|
+
|
|
56
|
+
|
|
57
|
+
def forward(self, x, y):
|
|
58
|
+
|
|
59
|
+
h1 = self.branch1(x)
|
|
60
|
+
|
|
61
|
+
h2 = self.branch2(x)
|
|
62
|
+
|
|
63
|
+
h = ShakeShake.apply(h1, h2, self.training)
|
|
64
|
+
|
|
65
|
+
h0 = x if self.equal_io else self.shortcut(x)
|
|
66
|
+
|
|
67
|
+
|
|
68
|
+
|
|
69
|
+
return h + h0
|
|
70
|
+
|
|
71
|
+
|
|
72
|
+
|
|
73
|
+
def _make_branch(self, in_ch, out_ch, stride=1):
|
|
74
|
+
|
|
75
|
+
return nn.Sequential(
|
|
76
|
+
|
|
77
|
+
nn.ReLU(inplace=False),
|
|
78
|
+
|
|
79
|
+
nn.Conv2d(in_ch, out_ch, 3, padding=1, stride=stride, bias=False),
|
|
80
|
+
|
|
81
|
+
nn.BatchNorm2d(out_ch),
|
|
82
|
+
|
|
83
|
+
nn.ReLU(inplace=False),
|
|
84
|
+
|
|
85
|
+
nn.Conv2d(out_ch, out_ch, 3, padding=1, stride=1, bias=False),
|
|
86
|
+
|
|
87
|
+
nn.BatchNorm2d(out_ch))
|
|
88
|
+
|
|
89
|
+
|
|
90
|
+
|
|
91
|
+
|
|
92
|
+
|
|
93
|
+
class ShakeResNet(nn.Module):
|
|
94
|
+
|
|
95
|
+
def __init__(self, depth, num_classes):
|
|
96
|
+
|
|
97
|
+
super(ShakeResNet, self).__init__()
|
|
98
|
+
|
|
99
|
+
n_units = (depth - 2) / 6
|
|
100
|
+
|
|
101
|
+
w_base = 32
|
|
102
|
+
|
|
103
|
+
in_chs = [16, w_base, w_base * 2, w_base * 4]
|
|
104
|
+
|
|
105
|
+
|
|
106
|
+
|
|
107
|
+
self.in_chs = in_chs
|
|
108
|
+
|
|
109
|
+
|
|
110
|
+
|
|
111
|
+
self.c_in = nn.Conv2d(3, in_chs[0], 3, padding=1)
|
|
112
|
+
|
|
113
|
+
self.layer1 = self._make_layer(n_units, in_chs[0], in_chs[1])
|
|
114
|
+
|
|
115
|
+
self.layer2 = self._make_layer(n_units, in_chs[1], in_chs[2], 2)
|
|
116
|
+
|
|
117
|
+
self.layer3 = self._make_layer(n_units, in_chs[2], in_chs[3], 2)
|
|
118
|
+
|
|
119
|
+
self.fc_out = nn.Linear(in_chs[3], num_classes)
|
|
120
|
+
|
|
121
|
+
|
|
122
|
+
|
|
123
|
+
# Initialize paramters
|
|
124
|
+
|
|
125
|
+
for m in self.modules():
|
|
126
|
+
|
|
127
|
+
if isinstance(m, nn.Conv2d):
|
|
128
|
+
|
|
129
|
+
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
|
130
|
+
|
|
131
|
+
m.weight.data.normal_(0, math.sqrt(2. / n))
|
|
132
|
+
|
|
133
|
+
elif isinstance(m, nn.BatchNorm2d):
|
|
134
|
+
|
|
135
|
+
m.weight.data.fill_(1)
|
|
136
|
+
|
|
137
|
+
m.bias.data.zero_()
|
|
138
|
+
|
|
139
|
+
elif isinstance(m, nn.Linear):
|
|
140
|
+
|
|
141
|
+
m.bias.data.zero_()
|
|
142
|
+
|
|
143
|
+
|
|
144
|
+
|
|
145
|
+
def forward(self, x):
|
|
146
|
+
|
|
147
|
+
h = self.c_in(x)
|
|
148
|
+
|
|
149
|
+
h = self.layer1(h, h)
|
|
150
|
+
|
|
151
|
+
h = self.layer2(h)
|
|
152
|
+
|
|
153
|
+
h = self.layer3(h)
|
|
154
|
+
|
|
155
|
+
h = F.relu(h)
|
|
156
|
+
|
|
157
|
+
h = F.avg_pool2d(h, 8)
|
|
158
|
+
|
|
159
|
+
h = h.view(-1, self.in_chs[3])
|
|
160
|
+
|
|
161
|
+
h = self.fc_out(h)
|
|
162
|
+
|
|
163
|
+
|
|
164
|
+
|
|
165
|
+
return h
|
|
166
|
+
|
|
167
|
+
|
|
168
|
+
|
|
169
|
+
def _make_layer(self, n_units, in_ch, out_ch, stride=1):
|
|
170
|
+
|
|
171
|
+
layers = []
|
|
172
|
+
|
|
173
|
+
for i in range(int(n_units)):
|
|
174
|
+
|
|
175
|
+
layers.append(ShakeBlock(in_ch, out_ch, stride=stride))
|
|
176
|
+
|
|
177
|
+
in_ch, stride = out_ch, 1
|
|
178
|
+
|
|
179
|
+
|
|
180
|
+
|
|
181
|
+
return nn.Sequential(*layers)
|
|
182
|
+
|
|
183
|
+
```
|
|
184
|
+
|
|
185
|
+
|
|
186
|
+
|
|
187
|
+
仮にh = self.layer1(h, h)として
|
|
188
|
+
|
|
189
|
+
ShakeBlock内のforward(self, x, y)を呼び出していますが、
|
|
190
|
+
|
|
191
|
+
実行すると、
|
|
192
|
+
|
|
193
|
+
|
|
194
|
+
|
|
195
|
+
result = self.forward(*input, **kwargs)
|
|
196
|
+
|
|
197
|
+
TypeError: forward() takes 2 positional arguments but 3 were given
|
|
198
|
+
|
|
199
|
+
|
|
200
|
+
|
|
201
|
+
というエラーが出ます。
|