AttributeError when input type has no element_size() method

[Lukas Eisert]

The bug

Hi, I see a problem with the code calculating the size of the layers:

In layer_info.py line 109:

if hasattr(inputs[0], "size") and callable(inputs[0].size):
    return list(inputs[0].size()), inputs[0].element_size()

I have the problem, that I use a package with modified tensors which have no "element_size" method.
I.e. the code crashes at that point.

Expected behavior

What about this

if hasattr(inputs[0], "size") and callable(inputs[0].size):
    if hasattr(inputs[0], "element_size") and callable(inputs[0].element_size):
        return list(inputs[0].size()), inputs[0].element_size()
    else:
        #Maybe add a warning here
        return list(inputs[0].size()), 0

[Tyler Yep]

Can you link the full code used to get this error? If I do not have a test case I cannot support it going forward.

PRs fixing this are welcome

[Lukas Eisert]

Hi!

My code is not public atm. I could write a test case if I find some time later this week!

In my case Im using E2-equivariant CNN layers from here https://github.com/QUVA-Lab/e2cnn and they use a modified tensor to handle the equivariance (https://github.com/QUVA-Lab/e2cnn/blob/master/e2cnn/nn/geometric_tensor.py). There is a size() but no element_size() method defined. This said, I could also ask them to add this to their code.

But I think it would be nice if your code would work also for non-supported layers as much as possible and not only return with an error.

[Tyler Yep]

Thanks, please submit a test case when you get the chance.