Missing weight sharing? about capsnet-tensorflow HOT 3 CLOSED

Glad that we share the same idea.
As for the choice of 32-channels of PrimaryCaps, I suppose it's kind of similar to the output channel in traditional conv nets, where we usually select 256, 512 etc. In this case, we choose 8x32=256 as the output channel, since the capsule needs to be a vector and we choose 8 as number of units in one vector, so finally, the channel of Primary caps would be 32. In a word, i think there is no specific reason, but following conventional instinct building a network with conv ops.
best,

from capsnet-tensorflow.

Hi,

I think it means that each capsule in the [6,6] grid shares the weights of conv op which have the shape 8 x 9x9 x 256 for each of 32 channels of the PrimaryCaps(i.e. sharing feature detector spatially), and it's not about sharing weights with DigitCaps.

from capsnet-tensorflow.

Thanks. I thought the same way at the beginning.

Never mind. In this way, we actually treat the 6632 capsules as individual ones. I am just not sure whether they have a specific reason to put a 32-layer representation here.

from capsnet-tensorflow.