lolemacs / soft-sharing Goto Github PK

Hello,
I understand that this is quite an old repo. But I wanted to try my luck here.

How can the SConv2d handle depth-wise convolution. Whenever I have tried to include the groups parameter with the final F.conv2d function, it has thrown a shape mismatch error.
For clarity:
I want to replace this convolution operation
nn.conv2d : Conv2d(486, 486, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=486, bias=False)

And this is my corresponding SConv2d object:
SConv2d(torch.Size([1, 486, 486, 3, 3]), stride=1, padding=1) with torch.Size([1, 1, 1, 1, 1]) coefficients.

Now, whenever I try to use a group value with the SConv2d function, it breaks down with shape mismatch.

RuntimeError: Given groups=486, weight of size [486, 486, 3, 3], expected input[2, 486, 112, 112] to have 236196 channels, but got 486 channels instead

I would be grateful for any suggestions.

Hi Savarese,

Thanks a lot for sharing the codes. I have run the SWRN-28-10-6 model with cutout regularization and all the 50000 training images, yet I could not obtain the 2.7% test error reported in the paper. I am wondering if I still need more modifications to achieve that. I appreciate it if you can give some useful suggestions.

Thanks!
Zhijie

In main.py you used name "descriptor". But in your layer implementation I think it is called "coefficients".

Is "descriptor" a special term in Pytorch or you actually did apply weight decay on those coefficients?