我正在尝试使用全连接层合成一个1×1卷积层。这意味着一个全连接神经网络来决定1×1卷积层的参数。我的做法如下：
class Network(nn.Module):def __init__(self, len_input, num_kernels): self.input_layers = nn.Sequential( nn.Linear(len_input, num_kernels * 2), nn.ReLU(), nn.Linear(num_kernels * 2, num_kernels), nn.ReLU() ) self.synthesized_conv = nn.Conv2d(in_channels=3, out_channels=num_kernels, bias=False, kernel_size=1) self.conv_layers = nn.Sequential( nn.ReLU(), nn.Conv2d(in_channels=num_kernels, out_channels=3, kernel_size=1) )def forward(self, x1, img): x = self.input_layer(x1.float()) with torch.no_grad(): self.synthesized_conv.weight = nn.Parameter(x.reshape_as(self.synthesized_conv.weight)) generated = self.conv_layer(self.synthesized_conv(img)) return generated 你可以看到我初始化了一个名为”synthesized_conv”的1×1卷积层，并尝试用一个名为”self.input_layers”的全连接网络输出通过引用调用来替换其参数。然而，梯度似乎并没有流经全连接网络，而只是流经卷积层。以下是全连接层的参数直方图：
这个直方图强烈表明全连接部分根本没有学习。这很可能是因为全连接网络输出对卷积参数的更新不当。有人能帮我解释如何在不破坏自动梯度图的情况下实现吗？
回答：
问题在于你一次又一次地重新定义模型的weight属性。一个更直接的解决方案是使用函数式方法，即torch.nn.functional.conv2d：
class Network(nn.Module): def __init__(self, len_input, num_kernels): super().__init__() self.input_layers = nn.Sequential( nn.Linear(len_input, num_kernels * 2), nn.ReLU(), nn.Linear(num_kernels * 2, num_kernels * 3), nn.ReLU()) self.synthesized_conv = nn.Conv2d( in_channels=3, out_channels=num_kernels, kernel_size=1) self.conv_layers = nn.Sequential( nn.ReLU(), nn.Conv2d(in_channels=num_kernels, out_channels=3, kernel_size=1)) def forward(self, x1, img): x = self.input_layers(x1.float()) w = x.reshape_as(self.synthesized_conv.weight) generated = F.conv2d(img, w) return generated此外，我认为你的input_layers将需要输出总共num_kernels * 3个组件，因为你合成的卷积总共有三个通道。
这里是一个测试示例：
>>> model = Network(10,3)>>> out = model(torch.rand(1,10), torch.rand(1,3,16,16))>>> out.shape(torch.Size([1, 3, 16, 16]), <ThnnConv2DBackward at 0x7fe5d8e41450>)当然，synthesized_conv的参数永远不会改变，因为它们从未被用来推断输出。你可以完全移除self.synthesized_conv：
class Network(nn.Module): def __init__(self, len_input, num_kernels): super().__init__() self.input_layers = nn.Sequential( nn.Linear(len_input, num_kernels * 2), nn.ReLU(), nn.Linear(num_kernels * 2, num_kernels*3), nn.ReLU()) self.syn_conv_shape = (num_kernels, 3, 1, 1) self.conv_layers = nn.Sequential( nn.ReLU(), nn.Conv2d(in_channels=num_kernels, out_channels=3, kernel_size=1)) def forward(self, x1, img): x = self.input_layers(x1.float()) generated = F.conv2d(img, x.reshape(self.syn_conv_shape)) return generated