Skip to content
Repo containing code to do the learnable scatternet/invariant convolutional layer
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
experiments
images
scatnet_learn
tests
MANIFEST.in
README.rst
main.py
requirements.txt
setup.py

README.rst

The Learnable ScatterNet

This repo implements the network described in the paper A Learnable ScatterNet: Locally Invariant Convolutional Layers. In particular, it is a way to replicate the results from Table 3 using PyTorch:

images/table3.png

The learnable ScatterNet is a DTCWT based scatternet. It differs from the original translation invariant scatternet introduced in Invariant Scattering Convolution Networks by adding a learned mixing layer between the scattering orders. To do this we've programmed a differentiable scatternet, allowing us to pass gradients through scattering layers. Additionally, because it is based on the DTCWT, we are restricted in the number of orientations we can use in the Scattering. For more information see the paper.

The results in the table above highlight the benefits of our implementation. ScatterNets A is the original translation invariant scatternet before a 4 layer convolutional network. ScatterNet B introduces the learned mixing matrices in between scattering orders. ScatterNet C is again the original translation invariant scatternet but with a learned convolutional layer before it and ScatterNet D has a learned layer before and in between scattering orders.

A block diagram of what we're doing is shown below (Figure 1 from the paper).

images/figure1.png

Installation

This repo uses my pytorch implementation of the dtcwt: pytorch_wavelets. You can install this however just by pip installing the requirements.txt. From the command line, the following 3 commands will allow you to run the experiments:

git clone https://github.com/fbcotter/scatnet_learn
pip install -r requirements.txt
pip install .

Running Experiments

The whole suite of tests to create Table 1 can be run by running the experiments/paper_experiments.py file. Note that this is written to work on a multi-gpu system, and loads each gpu with different nets - i.e. it is very intensive and can take several hours to run. It is recommended to try to run individual nets first.

Development and Expansion

The code is designed to be reusable, so you can design your own networks using the original ScatNet or Learnable ScatNet layers. For example, if you wanted to create a standard DTCWT ScatterNet frontend with no learned mixing, you can with the following code:

from scatnet_learn import ScatLayerj1
import torch.nn as nn
from collections import OrderedDict
C = 3
# A standard scatlayer expands the channel input from C to 7C - one
# lowpass and 6 oriented bandpass.
frontend = nn.Sequential(OrderedDict([
            ('order1', ScatLayer()),
            ('order2', ScatLayer())]))
x = torch.randn(1,C,64,64)
y = frontend(x)
y.shape
>>> (1, 147, 16, 16)

If you want to use the proposed expansions from the paper, you should instead use the InvariantLayerj1 class. This wraps the ScatLayerj1 class and adds the learned mixing matrix afterwards.

from scatnet_learn import InvariantLayerj1
frontend = nn.Sequential(OrderedDict([
            ('order1', InvariantLayerj1(C)),
            ('order2', InvariantLayerj1(7*C))]))

By default the mixing matrix will preserve the translation invariant scatternet tendency to increase the channel dimension by 7. However, if you want to change the number of output channels you can:

C = 3
frontend = nn.Sequential(OrderedDict([
            ('order1', InvariantLayerj1(C, F=7*C)),
            ('order2', InvariantLayerj1(7*C, F=7*C))]))
x = torch.randn(1,C,64,64)
y = frontend(x)
y.shape
>>> (1, 21, 16, 16)

Experiments

In the experiments folder, there are scripts for running cifar and mnist experiments using the newly proposed learnable scatternet layer. These were the scripts used to generate the results tables in the paper.

You can’t perform that action at this time.