模型设计速查表

神经网络中的数据流形式

1class ReLUConvBN(nn.Module):
2    def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
3        super(ReLUConvBN, self).__init__()
4        self.op = nn.Sequential(
5            nn.ReLU(inplace=False),
6            nn.Conv2d(C_in, C_out, kernel_size, stride=stride, padding=padding, bias=False),
7            nn.BatchNorm2d(C_out, affine=affine)
8        )
9
10    def forward(self, x):
11        return self.op(x)

1class SepConv(nn.Module):
2    def __init__(self, C_in, C_out, kernel_size, stride, padding, affine=True):
3        super(SepConv, self).__init__()
4        self.op = nn.Sequential(
5            nn.ReLU(inplace=False),
6            nn.Conv2d(C_in, C_in, kernel_size=kernel_size, stride=stride, 
7                      padding=padding, groups=C_in, bias=False),
8            nn.Conv2d(C_in, C_in, kernel_size=1, padding=0, bias=False),
9            nn.BatchNorm2d(C_in, affine=affine),
10            nn.ReLU(inplace=False),
11            nn.Conv2d(C_in, C_in, kernel_size=kernel_size, stride=1, 
12                      padding=padding, groups=C_in, bias=False),
13            nn.Conv2d(C_in, C_out, kernel_size=1, padding=0, bias=False),
14            nn.BatchNorm2d(C_out, affine=affine)
15        )
16
17    def forward(self, x):
18        return self.op(x)

1class DilConv(nn.Module):
2    def __init__(self, C_in, C_out, kernel_size, stride, padding, dilation, affine=True):
3        super(DilConv, self).__init__()
4        self.op = nn.Sequential(
5            nn.ReLU(inplace=False),
6            nn.Conv2d(C_in, C_in, kernel_size=kernel_size, stride=stride, padding=padding, 
7            dilation=dilation, groups=C_in, bias=False),
8            nn.Conv2d(C_in, C_out, kernel_size=1, padding=0, bias=False),
9            nn.BatchNorm2d(C_out, affine=affine)
10        )
11    def forward(self, x):
12        return self.op(x)

1class FactorizedReduce(nn.Module):
2    def __init__(self, C_in, C_out, affine=True):
3        super(FactorizedReduce, self).__init__()
4        assert C_out % 2 == 0
5        self.relu = nn.ReLU(inplace=False)
6        self.conv_1 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False)
7        self.conv_2 = nn.Conv2d(C_in, C_out // 2, 1, stride=2, padding=0, bias=False)
8        self.bn = nn.BatchNorm2d(C_out, affine=affine)
9    def forward(self, x):
10        x = self.relu(x)
11        out = torch.cat([self.conv_1(x), self.conv_2(x[:,:,1:,1:])], dim=1)
12        out = self.bn(out)
13        return out