About the input dimension of fc1 of decoder

[zijundeng]

Hello, thanks for your code, which helps me strengthen the understanding of the capsnet. I have a question as follows.

For all the capsules, it seems better to share fc1 parameters of decoder (so the input dimension is 16 instead of 10 * 16). And this repo uses input dimension of 16. What is your consideration?

[laubonghaudoi]

Thanks for your correction. According to section 4.1 in the paper, which says The output of the digit
capsule is fed into a decoder, we can infer that only the output of one capsule, rather than outputs of all 16 capsules, is sent to the decoder. I doubt if there is any difference in the final outcome between these two implementations (since all other outputs are zeros). I'm working on it and might push an update later, sorry for the mistake.

[zijundeng]

客气了哈哈

[laubonghaudoi]

Fixed in #3.