省流

• nn.Conv2d 需要的输入张量格式为 (batch_size, channels, height, width)，但您的示例输入张量 x 是 (batch_size, height, width, channels)。因此，需要对输入张量进行转置。

• 注意，TensorFlow使用"NHWC"（批次、高度、宽度、通道）格式，而PyTorch使用"NCHW"（批次、通道、高度、宽度）格式

异常报错

RuntimeError: Given groups=1, weight of size [16, 3, 2, 3], 
expected input[8, 65, 66, 3] to have 3 channels, 
but got 65 channels instead

异常截图

异常代码

def down_shifted_conv2d(x , num_filters , filters_size = [2,3],stride = 1, **kwargs):batch_size,H,W,channels = x.shapepadding = (0,0,int(((filters_size[1]) - 1) / 2 ) , int((int(filters_size[1]) - 1) / 2),int(filters_size[0]) - 1 , 0,0,0)x_paded = nn.functional.pad(x, padding)print(x_paded.shape)conv_layer = nn.Conv2d(in_channels=channels, out_channels=num_filters, kernel_size=filters_size,stride=stride, **kwargs)return conv_layer(x_paded)
# Example usage
x = torch.randn(8, 64, 64, 3)  # Example input with batch size 8, height and width 64, and 3 channels
num_filters = 16
output = down_shifted_conv2d(x, num_filters)
print(output.shape)

原因解释

• 在pytorch中，“nn.Conv2d”需要输入的张量格式为（batch_size,channels,height,width），原图输入的x的格式是（batch_size,height ,weight,channel）所以需要对tensor进行转置。

• 矩阵交换维度的函数permute，按照编号，将新的顺序填好即可。

def down_shifted_conv2d(x , num_filters , filters_size = [2,3], stride = 1, **kwargs):batch_size, H, W, channels = x.shape# Transpose the input tensor to (batch_size, channels, height, width)x = x.permute(0, 3, 1, 2)# Paddingpadding = (int((filters_size[1] - 1) / 2), int((filters_size[1] - 1) / 2),filters_size[0] - 1, 0)x_paded = F.pad(x, padding)

修正代码

def down_shifted_conv2d(x , num_filters , filters_size = [2,3],stride = 1, **kwargs):batch_size,H,W,channels = x.shape# 按照顺序对4个维度分别进行填充padding = (0,0,int(((filters_size[1]) - 1) / 2 ) , int((int(filters_size[1]) - 1) / 2),int(filters_size[0]) - 1 , 0,0,0)x_paded = nn.functional.pad(x, padding)x_paded = x_paded.permute(0,3,1,2)# 进行卷积conv_layer = nn.Conv2d(in_channels=channels, out_channels=num_filters, kernel_size=filters_size,stride=stride, **kwargs)return conv_layer(x_paded)
# Example usage
x = torch.randn(8, 64, 64, 3)  
num_filters = 16
output = down_shifted_conv2d(x, num_filters)
print(output.shape)

执行结果