多尺度注意力机制代码

class MultiScaleAttention(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(MultiScaleAttention, self).__init__()
        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
        self.conv2 = nn.Conv2d(in_channels, out_channels, kernel_size=5, padding=2)
        self.conv3 = nn.Conv2d(in_channels, out_channels, kernel_size=7, padding=3)
        self.fc = nn.Linear(3*out_channels, out_channels)

    def forward(self, x):
        feat1 = F.relu(self.conv1(x))
        feat2 = F.relu(self.conv2(x))
        feat3 = F.relu(self.conv3(x))
        multi_scale_feat = torch.cat((feat1, feat2, feat3), dim=1)
        attention_weight = F.softmax(multi_scale_feat, dim=1)
        attention_out = torch.matmul(attention_weight, multi_scale_feat)
        output = self.fc(attention_out)
        return output