About the equivalence of wide_resnet about e2cnn HOT 5 CLOSED

Hi @hcy226

The reason is that the final avg_pool2d is necessary to obtain invariance to rotations.
By removing the average pooling, the output still has some spatial resolution; a rotation of the input leads to a rotation of this output feature, i.e. the model is equivariant, not invariant.

The code, however, is testing for invariance, not equivariance.
This means that, when you compare y with e.g. y90, you should first rotate y90 back by 90 degrees.

Hope this helps,
Gabriele

from e2cnn.

Thanks @Gabri95 ! Here comes another question: why is the shape of x in the wrn 2^n+1,e.g. 513*513/33*33 in case I use pooling or dilation? Can I test my e2cnn-based model by the input shape of 256*256 or 512*512 ?

from e2cnn.

Hi @hcy226

This is to ensure the pooling with stride 2 doesn't break equivariance to 90 degrees rotations.
Check Figure 2 here https://arxiv.org/abs/2004.09691

Best,
Gabriele

from e2cnn.

Hi @hcy226

This is to ensure the pooling with stride 2 doesn't break equivariance to 90 degrees rotations. Check Figure 2 here https://arxiv.org/abs/2004.09691

Best, Gabriele

Thanks！
In fact, when I test the equivariance of my own network based on e2cnn with the input shape [1,3,256,256], I find that average pooling and max pooling in e2cnn does not break the equivariance, but conv with stride>1 will do, which is different from the paper. May I know the reason why is that?

from e2cnn.

Hi @hcy226

I am not sure what you mean: stride>1 is expected to break equivariance when the input has even size, as explained in the paper I linked to.
Could you be more precise?

Best,
Gabriele

from e2cnn.