Scaling the Scattering Transform: Deep Hybrid Networks

• Pure tensorflow implementation of the scattering transform and hybrid networks.
• No need for external dependencies, only tensorflow functions are used
• Depending on the settings (e.g. use_XLA), the code is between 3x to 5x slower than pyscatwave.
• Since only tensorflow primitives are involved, it is possible to backprop through the Scattering Transform (this functionality will soon be added to pyscatwave).

Sources:

• Scaling the Scattering Transform: Deep Hybrid Networks
• pyscatwave

This repository is a simple adapatation of pyscatwave with tensorflow primitives only.

Copyright (c) 2017, Eugene Belilovsky (INRIA), Edouard Oyallon (ENS) and Sergey Zagoruyko (ENPC) All rights reserved.

Requirements

python modules

• tensorflow 1.0
• h5py
• matplotlib
• opencv 3
• numpy
• tqdm
• colorama
• terminaltables
• parmap

You should also install the dependencies of pyscatwave to run the tests.

Running the code

Follow these instructions

Running tests

Follow these instructions