ConvCaps p_in a_in view

[tomahawk810]

I think there is an issue in the way the input tensor x is reshaped in order to extract a_in and p_in.
It seems to me that the dimensions of a_in and p_in require a permutation before applying Tensor.view().

Note that I changed the training batch size to 16, also I am using A, B, C, D = 32, 32, 32, 32.

Transformation before view:

After this line:

Matrix-Capsules-EM-PyTorch/model/capsules.py

Line 253 in 9cb3fc0

p_in = x[:, :, :, :, :, :self.B*self.psize].contiguous()

I get this:

p_in.shape
Out[2]: torch.Size([16, 3, 3, 6, 6, 512])

View:

The view is done in the following way:

Matrix-Capsules-EM-PyTorch/model/capsules.py

Line 255 in 9cb3fc0

p_in = p_in.view(b*oh*ow, self.K*self.K*self.B, self.psize)

To do the view in this way, p_in.shape should be:
torch.Size([16, 6, 6, 3, 3, 512])

Do you agree? I am new to Pytorch, so I might misunderstand the way Tensor.view() works.

[yl-1993]

@tomahawk810

Sorry about the late reply, as I am busy preparing cvpr submission.

As stated in

Matrix-Capsules-EM-PyTorch/model/capsules.py

Line 192 in 9cb3fc0

Output: (b, H', W', K, K, B*(P*P+1))

The shape of x after adding patches is (b, H', W', K, K, B*(P*P+1)).
Therefore, the shape of p_in is (b, H', W', K, K, B*P*P).
For view, actually it can be split into two steps:

p_in = p_in.view(b*H'*W', K*K, B, P*P)
p_in = p_in.view(b*H'*W', K*K, B*P*P)

[yl-1993]

@tomahawk810 I get your point now.
I check the code and you are right about the output shape.
The issue lies in the add_patches function.

I have fixed the problem now. Could you please help review the PR #4?

Thanks for pointing out!

[yl-1993]

@tomahawk810 I merge the PR since there is no comment for two weeks.
Please continue to comment if you feel anything is weird.