Need equivariant ResNet & VGG like architectural networks implemented in E2CNN library

[quva-lab]

Need equivariant ResNet & VGG like architectural networks implemented in E2CNN library

[namankhetan]

Hii @mkool67 I think you haven't included the group pooling layer after the sterrable CNN's block which is required to get the equivariant output from the network as they have mentioned in their paper.

[mkool67]

Thank you @namankhetan for pointing this out. I looked the example code again and I can't see the group pooling layer in the example code. So, it will be difficult for me to build equivariant model. Can you include code for ResNet & VGG like architectures with groop pooling layer.

[namankhetan]

@mkool67 Sure!..will add them in examples folder of the repository. I will also include test cases for you to have a better understanding.

[namankhetan]

hey @mkool67 link to the pull request #46 , let me know any further concers, if any.

[mkool67]

Thank you @namankhetan, It worked for me. It is exactly what I was looking for.

[Gabri95]

Hi @mkool67

The WideResNet in the examples does not include a group pooling layer since the last layer already maps to trivial (invariant) representations through convolution (R2Conv).
This is just a design choice but is not the only one possible (e.g. you could map to regular representation with convolution and then apply group pooling).

Thanks a lot @namankhetan for building these models!

If you are fine, I will close this issue. Feel free to reopen it if you have any question

Gabriele