The mask operation is different between training and testing?

[AlexHex7]

In the paper, there is a sentence

During training, we mask out all but the activity vector of the correct digit capsule.

so, I think it will mask all but the capsule (1x16 vector) which is match the ground-truth during training. And the code now is about testing time, it will mask all but the longest capsule (1x16 vector).

[timomernick]

Sorry, I don't quite understand this. Can you elaborate?

[AlexHex7]

First, I'm not sure whether it is right.

In my opinion, the code of reconstruction_loss does is that choose the longest capsule, and then mask all other 9 capsules. I think this is in testing period.
For the training period, what should be done is that choose the capsule which is stand for the input digit, and mask all other 9 capsules. For example, if the input digit is 0, then for the digitcaps (10 x 16), the first capsule whose length is stand for the probability of digit 0's existing should be chosen. And the other 9 capsule which are stand for digit 1 to 9 should be mask.

But the code does not has the operation of training period.

[timomernick]

Oh, I think I see. At train time I am using the longest capsule and masking the others, but maybe I should be using the capsule corresponding to the digit from the training sample. I hope to look into this soon, as I don't understand the difference.