search

southworks / CapsNet-CNTK

A CNTK implementation of CapsNet based on Geoffrey Hinton's paper Dynamic Routing Between Capsules
6 stars 5 forks source link

Confusing implementation #2

Closed akanimax closed 6 years ago

akanimax commented 6 years ago

In the module CapsLayer.py, inside the function DigitCaps(input, num_capsules, dim_out_vector, routings=3, name='DigitCaps'): for computing the predictions given by the prev layer capsules 'u' for the subsequent layer 'u_hat_ij', the implementation does elementwise multiplication (hadamard product) and then performs a reduce. Aren't we supposed to project every 'u' using a linear matrix projection into 'u_hat_ij'?

rnrneverdies commented 6 years ago

I think the current implementation is equivalent to a matmul. Don't you? The main issue here is that CNTK does not currently support multidimensional matmul which would be more computationally efficiently.

I've originally tried to use the numpy matmul via a CNTK custom operator but is significantly slower on GPU than the current implementation. Although it is faster on CPU, I leaned towards the GPU power mainly due to the training time.

If you have any improvement suggestion don't hesitate to comment on it.