Capsule network in < 80 lines.
Research-friendly implementation of capsule networks from Dynamic Routing Between Capsules in pytorch.
Capsules are groups of neurons each describing an entity with positional characteristics.
Capsules are meant to form a 'soft parsing tree' for a scene with strength of confidence encoded by norm of capsule's activations.
Video of lecture: https://www.youtube.com/watch?v=x5Vxk9twXlE
- completely readable and very compact
- capsule layers are perfectly stackable
- auto inference of number of capsules after convolutional stem
- memory efficiency:
almost 3x less GPU memory foorprint compared to other implementation (2Gb vs 5.9Gb for batch size of 256) - blazingly fast:
15 sec/epoch vs 747 sec/epoch on single V100 vs other implementation
and, well, I didn't even use torch.jit.trace and did not use fp16, which will provide an additional boost in efficiency.
Two most important changes that made this possible are:
- all convolutional capsules are packed into one fat convolution instead of splitting into several convolutions and then reshaping and concatenating,
einopstakes care of making that efficiently. Additionallyeinopsresolves weight management for capsules. - ugly routing implementation made efficient - all split/concat and all unnecessary repeats are eliminated by proper usage of
einsumandeinops
There is a ton of implementations for CapsNets.
Most of them are hardly readable, almost all are inefficient and a large fraction is simply wrong. All of them require to make some computations specific for each dataset and disallow easy tweaking of parts.
@michaelklachko suggested to rewrite his implementation of routing algorithm to pytorch using ein-notation.
This turned out to be a very nice exercise.
Additionally to improvements in routing done by Michael,
I've minified all tricky places in encoder/decoder with einops layers.