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。
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -102,4 +102,32 @@
|
|
|
102
102
|
|
|
103
103
|
コードは
|
|
104
104
|
https://github.com/owruby/shake-shake_pytorch
|
|
105
|
-
から借りています。
|
|
105
|
+
から借りています。
|
|
106
|
+
|
|
107
|
+
|
|
108
|
+
ご指摘の点について
|
|
109
|
+
ShakeBlock内のforwardを
|
|
110
|
+
```
|
|
111
|
+
def forward(self, x, y):
|
|
112
|
+
h1 = self.branch1(x)
|
|
113
|
+
h2 = self.branch2(x)
|
|
114
|
+
h = ShakeShake.apply(h1, h2, self.training)
|
|
115
|
+
h0 = x if self.equal_io else self.shortcut(x)
|
|
116
|
+
|
|
117
|
+
return h + h0, h + h0
|
|
118
|
+
```
|
|
119
|
+
ShakeResNet内のforwardを
|
|
120
|
+
```
|
|
121
|
+
def forward(self, x):
|
|
122
|
+
h = self.c_in(x)
|
|
123
|
+
h, h = self.layer1(h, h)
|
|
124
|
+
h, h = self.layer2(h, h)
|
|
125
|
+
h, h = self.layer3(h, h)
|
|
126
|
+
h = F.relu(h)
|
|
127
|
+
h = F.avg_pool2d(h, 8)
|
|
128
|
+
h = h.view(-1, self.in_chs[3])
|
|
129
|
+
h = self.fc_out(h)
|
|
130
|
+
|
|
131
|
+
return h
|
|
132
|
+
```
|
|
133
|
+
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。
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -98,4 +98,8 @@
|
|
|
98
98
|
result = self.forward(*input, **kwargs)
|
|
99
99
|
TypeError: forward() takes 2 positional arguments but 3 were given
|
|
100
100
|
|
|
101
|
-
というエラーが出ます。
|
|
101
|
+
というエラーが出ます。
|
|
102
|
+
|
|
103
|
+
コードは
|
|
104
|
+
https://github.com/owruby/shake-shake_pytorch
|
|
105
|
+
から借りています。
|
1
コードを追加しました。
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -14,4 +14,88 @@
|
|
|
14
14
|
TypeError: forward() takes 2 positional arguments but 4 were given
|
|
15
15
|
|
|
16
16
|
というエラーが出ます。
|
|
17
|
-
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？
|
|
17
|
+
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？
|
|
18
|
+
|
|
19
|
+
```
|
|
20
|
+
class ShakeBlock(nn.Module):
|
|
21
|
+
def __init__(self, in_ch, out_ch, stride=1):
|
|
22
|
+
super(ShakeBlock, self).__init__()
|
|
23
|
+
self.equal_io = in_ch == out_ch
|
|
24
|
+
self.shortcut = self.equal_io and None or Shortcut(in_ch, out_ch, stride=stride)
|
|
25
|
+
|
|
26
|
+
self.branch1 = self._make_branch(in_ch, out_ch, stride)
|
|
27
|
+
self.branch2 = self._make_branch(in_ch, out_ch, stride)
|
|
28
|
+
|
|
29
|
+
def forward(self, x, y):
|
|
30
|
+
h1 = self.branch1(x)
|
|
31
|
+
h2 = self.branch2(x)
|
|
32
|
+
h = ShakeShake.apply(h1, h2, self.training)
|
|
33
|
+
h0 = x if self.equal_io else self.shortcut(x)
|
|
34
|
+
|
|
35
|
+
return h + h0
|
|
36
|
+
|
|
37
|
+
def _make_branch(self, in_ch, out_ch, stride=1):
|
|
38
|
+
return nn.Sequential(
|
|
39
|
+
nn.ReLU(inplace=False),
|
|
40
|
+
nn.Conv2d(in_ch, out_ch, 3, padding=1, stride=stride, bias=False),
|
|
41
|
+
nn.BatchNorm2d(out_ch),
|
|
42
|
+
nn.ReLU(inplace=False),
|
|
43
|
+
nn.Conv2d(out_ch, out_ch, 3, padding=1, stride=1, bias=False),
|
|
44
|
+
nn.BatchNorm2d(out_ch))
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
class ShakeResNet(nn.Module):
|
|
48
|
+
def __init__(self, depth, num_classes):
|
|
49
|
+
super(ShakeResNet, self).__init__()
|
|
50
|
+
n_units = (depth - 2) / 6
|
|
51
|
+
w_base = 32
|
|
52
|
+
in_chs = [16, w_base, w_base * 2, w_base * 4]
|
|
53
|
+
|
|
54
|
+
self.in_chs = in_chs
|
|
55
|
+
|
|
56
|
+
self.c_in = nn.Conv2d(3, in_chs[0], 3, padding=1)
|
|
57
|
+
self.layer1 = self._make_layer(n_units, in_chs[0], in_chs[1])
|
|
58
|
+
self.layer2 = self._make_layer(n_units, in_chs[1], in_chs[2], 2)
|
|
59
|
+
self.layer3 = self._make_layer(n_units, in_chs[2], in_chs[3], 2)
|
|
60
|
+
self.fc_out = nn.Linear(in_chs[3], num_classes)
|
|
61
|
+
|
|
62
|
+
# Initialize paramters
|
|
63
|
+
for m in self.modules():
|
|
64
|
+
if isinstance(m, nn.Conv2d):
|
|
65
|
+
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
|
66
|
+
m.weight.data.normal_(0, math.sqrt(2. / n))
|
|
67
|
+
elif isinstance(m, nn.BatchNorm2d):
|
|
68
|
+
m.weight.data.fill_(1)
|
|
69
|
+
m.bias.data.zero_()
|
|
70
|
+
elif isinstance(m, nn.Linear):
|
|
71
|
+
m.bias.data.zero_()
|
|
72
|
+
|
|
73
|
+
def forward(self, x):
|
|
74
|
+
h = self.c_in(x)
|
|
75
|
+
h = self.layer1(h, h)
|
|
76
|
+
h = self.layer2(h)
|
|
77
|
+
h = self.layer3(h)
|
|
78
|
+
h = F.relu(h)
|
|
79
|
+
h = F.avg_pool2d(h, 8)
|
|
80
|
+
h = h.view(-1, self.in_chs[3])
|
|
81
|
+
h = self.fc_out(h)
|
|
82
|
+
|
|
83
|
+
return h
|
|
84
|
+
|
|
85
|
+
def _make_layer(self, n_units, in_ch, out_ch, stride=1):
|
|
86
|
+
layers = []
|
|
87
|
+
for i in range(int(n_units)):
|
|
88
|
+
layers.append(ShakeBlock(in_ch, out_ch, stride=stride))
|
|
89
|
+
in_ch, stride = out_ch, 1
|
|
90
|
+
|
|
91
|
+
return nn.Sequential(*layers)
|
|
92
|
+
```
|
|
93
|
+
|
|
94
|
+
仮にh = self.layer1(h, h)として
|
|
95
|
+
ShakeBlock内のforward(self, x, y)を呼び出していますが、
|
|
96
|
+
実行すると、
|
|
97
|
+
|
|
98
|
+
result = self.forward(*input, **kwargs)
|
|
99
|
+
TypeError: forward() takes 2 positional arguments but 3 were given
|
|
100
|
+
|
|
101
|
+
というエラーが出ます。
|