Using ConvTranspose2d and Conv2d on the same input

1

Hello,

I am currently working on HoughNet (https://arxiv.org/abs/2007.02355) paper which has its code publicly available. I tried to change ConvTranspose2d in (https://github.com/nerminsamet/houghnet/blob/master/src/lib/models/networks/houghnet_large_hourglass.py#L280) with Conv2d. However, I get this error message:


RuntimeError: Given groups=1, weight of size [9, 1, 17, 17], expected input[2, 9, 128, 192] to have 1 channels, but got 9 channels instead

The original configuration is as follows:

Sequential(
  (0): ConvTranspose2d(9, 1, kernel_size=(17, 17), stride=(1, 1), padding=(8, 8), bias=False)
) 
input_size = torch.Size([2, 9, 1, 128, 192])
weight_size = torch.Size([9, 1, 17, 17])

The error occurs when the configuration is as follows:

Sequential(
  (0): Conv2d(9, 1, kernel_size=(17, 17), stride=(1, 1), padding=(8, 8), bias=False)
) 
input_size = torch.Size([2, 9, 1, 128, 192])
weight_size = torch.Size([9, 1, 17, 17])

As far as I understand, Conv2d should be applicable on this input. I tried replicate this on a small trial code. There I have to have a 4D input and I wasn’t able to replicate this error. Whenever, I have 4D inputs however, I can replace my conv layers with transposed conv and vice versa. Any idea why this might occur?

Thanks.

2

The weight_size of the nn.Conv2d module is wrong, as it would be:

conv = nn.Conv2d(9, 1, kernel_size=(17, 17), stride=(1, 1), padding=(8, 8), bias=False)
print(conv.weight.shape)
> torch.Size([1, 9, 17, 17])

Are you changing the .weight attribute manually in your code?

3

I simply changed the ConvTranspose2d with Conv2d and I get this error. However, the weight initially is assigned manually: https://github.com/nerminsamet/houghnet/blob/master/src/lib/models/networks/houghnet_large_hourglass.py#L276

4

I would like to reduce 9 channels to 1, do I have to have the matrix ordering swapped manually as well?

5

As shown in my minimal code snippet, the weight shape is wrong in your code, so you would have to permute it to the expected shape, i.e. swap dim0 with dim1.

1 Like