Deviations of this implementation from the original paper

[InnovArul]

I have come across the following main deviation between the paper(https://arxiv.org/abs/1710.09829) and the implementation of this repo:

In the paper: there are 32 primary capsule layer with 6x6 capsules each of 8 dimensions each. Hence, we needs to have 32 independent convolution layers with 8 output channels. In the repo: it is implemented to be having 8 independent convolution layers with 32 output channels.

reference:

https://photos.app.goo.gl/FeCg4ejNdF3eVPvh6

https://github.com/cedrickchee/capsule-net-pytorch/blob/master/capsule_layer.py#L52

I have noticed this issue in another repo as well. I'm not sure if there is a misunderstanding in my interpretation of the paper. https://github.com/gram-ai/capsule-networks/issues/23

Can you please check and comment on this?

[nabsabraham]

This is a good question! I wonder if it has to do with weight sharing? It says in the paper that "each capsule in the [6 × 6] grid is sharing their weights with each other" so I'm wondering if its only per capsule dimension. Any input is useful! This is just my conjecture

[InnovArul]

I believe the sharing of weights in [6 x 6] grid is achieved by default due to the use of Convolution layer with 8 dimensions (Channels). Don't you think so?