Routing algorithm about capsnet-tensorflow

[laubonghaudoi]

Agree. I still have difficulties in understanding how b_ij and W_ij are updated, there is little information on this in the paper. Hinton has mentioned to give up back-propagation in his speeches but what I have seen so far in this implementation is still using Adam to update weights by back-propagating gradients from the two capsule layers.

[naturomics]

@bshao001 You're right, I misunderstood the routing algorithm. Thank you for your pointing out. The interesting thing is, the loss can go down steadily (not effective as the result of the paper). I'm fixing it.

[13331151]

I don't think the routing process could be unfolded into 6 layers, because u_ij is only calculated once in one forward pass.

[InnerPeace-Wu]

Here is some of my idea about log prior or bias. @bshao001 I think the static way will still work fine, the key is we should not update bias during routing. Since we only use dynamic routing to compute digit caps in this case, and with static way, the digit caps could be updated. Here is my implementation of Capsnet for reference.

[bshao001]

@naturomics
不幸被我言中，原作者的代码果然用了tf.while_loop来实现核心的_update_routing方法。