将特征映射附加到PyTorch中的网络中间层

Question

我的PyTorch网络架构如下所示。它接收大小为16x8x2048的输入，并给出大小为256x128x3的输出。我有一个特征图，我想在下面的架构中的每个上采样层之后，沿着通道维度附加它。在添加之前，我会将功能映射缩放到适当的分辨率。如何将这些要素图附加到网络的中间层

model = []
model += [Conv2dBlock(2048, 256, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)]
model +=  [nn.Upsample(scale_factor=2, mode='bilinear')] 
model +=  [Conv2dBlock(256, 128, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)]
model +=  [nn.Upsample(scale_factor=2, mode='bilinear')]
model +=  [Conv2dBlock(128, 64, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)]                        
model +=  [nn.Upsample(scale_factor=2, mode='bilinear')] 
model +=  [Conv2dBlock(64, 32, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)]
model +=  [nn.Upsample(scale_factor=2, mode='bilinear')] 
model +=  [Conv2dBlock(32, 32, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)]                        
model += [Conv2dBlock(32, 3, 3, 1, 1, norm='none', activation=activ, pad_type=pad_type)]                        
model = nn.Sequential(*model)

Answer 1

因此，解决方案只是改变了网络的编写方式，正如https://stackoverflow.com/users/10229754/mercury所指出的那样。以下更改显示了一种写入网络的方式，以便可以附加额外的输入。在forward()函数中，可以使用torch.cat()-

conv1 = Conv2dBlock(2063, 256, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)
up1 = nn.Upsample(scale_factor=2, mode='bilinear')
conv2 = Conv2dBlock(256, 128, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)
up2 = nn.Upsample(scale_factor=2, mode='bilinear')
conv3 = Conv2dBlock(128, 64, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)
up3 = nn.Upsample(scale_factor=2, mode='bilinear')
conv4 = Conv2dBlock(64, 32, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)
up4 = nn.Upsample(scale_factor=2, mode='bilinear')
conv5 = Conv2dBlock(32, 32, 3, 1, 1, norm='bn', activation=activ, pad_type=pad_type)
conv6 = Conv2dBlock(32, 3, 3, 1, 1, norm='none', activation=activ, pad_type=pad_type)