CapsNet Tweak

vous permet de modifier ou de tordre les dimensions de la capsule d'un réseau de sortie Capsule.

Allows you to run a part of a Capsule Network model on the device and modify the output capsule parameters to produce different images.

Could provide some insights into what was learned for each parameter.

Capsule Networks are based on Geoffrey Hinton's paper Dynamic Routing Between Capsules

I extracted the decoder after training and saved some of the digit capsule layer output to be able to run inference on it on the device (offline).

A decoder is found at the end of a Capsule Network and is used to reconstruct the source image given the capsule outputs from the digit caps layer. During training, its loss contributes to the final loss of the model and this ends up , hopefully, giving each parameter something meaningful and a lot of times an easily recognizable feature.

Implemented in Tensorflow - source: https://github.com/JsFlo/CapsNet