【pytorch】——conv，convtranspose2d的权重排布方式

pytorch

convtranspose2d的outc和inc是交换了位置的，需要注意一下。这样自己在保存的时候，可能需要用permute交换两个的维度。直接看具体的例子：

import torch
import torch.nn as nn

conv = nn.Conv2d(in_channels=32, out_channels=16, kernel_size=3, stride=1)
transpose = nn.ConvTranspose2d(in_channels=32, out_channels=16, kernel_size=3, stride=1)

print("conv weight shape:{}".format(conv.weight.shape))
print("transpose2d weight shape:{}".format(transpose.weight.shape))

log

conv weight shape:torch.Size([16, 32, 3, 3])
transpose2d weight shape:torch.Size([32, 16, 3, 3])