CuPy/PyTorch Scattering implementation

A scattering network is a Convolutional Network with filters predefined to be wavelets that are not learned and it can be used in vision task such as classification of images. The scattering transform can drastically reduce the spatial resolution of the input (e.g. 224x224->14x14) with demonstrably neglible loss in dicriminative power.

The software uses PyTorch + NumPy FFT on CPU, and PyTorch + CuPy + CuFFT on GPU.

Previous (lua-based) versions of the code can be found at https://github.com/edouardoyallon/scatwave

If using this code for your research please cite our paper:

E. Oyallon, E. Belilovsky, S. Zagoruyko Scaling the Scattering Transform: Deep Hybrid Networks

You can find experiments from the paper in the following repository: https://github.com/edouardoyallon/scalingscattering/

We used PyTorch for running experiments in https://arxiv.org/abs/1703.08961, but it is possible to use scattering with other frameworks (e.g. Chainer, Theano or Tensorflow) if one copies Scattering outputs to CPU (or run on CPU and convert to numpy.ndarray via .numpy()).

We do some simple timings and comparisons to the previous (multi-core CPU) implementation of scattering (ScatnetLight). We benchmark the software using a 1080 GPU. Below we show input sizes (WxHx3xBatchSize) and speed:

32 × 32 × 3 × 128 (J=2)- 0.03s (speed of 8x vs ScatNetLight)

256 × 256 × 3 × 128 (J=2) - 0.71 s (speed up of 225x vs ScatNetLight)

The software was tested on Linux with anaconda Python 2.7 and various GPUs, including Titan X, 1080s, 980s, K20s, and Titan X Pascal.

The first step is to install pytorch following instructions from http://pytorch.org, then you can run pip:

pip install -r requirements.txt
python setup.py install

Example:

import torch
from scatwave.scattering import Scattering

scat = Scattering(M=32, N=32, J=2).cuda()
x = torch.randn(1, 3, 32, 32).cuda()

print scat(x).size()

All contributions are welcome.

Edouard Oyallon, Eugene Belilovsky, Sergey Zagoruyko