Repo to replicate experiments on learning convolutional filters in the wavelet domain
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


GainLayer Experiments

This repo implements the network described in the paper Deep Learning in the Wavelet Domain. In particular, it is a way to replicate the results from Table 1:


The WaveLeNet uses two convolutional layers which learn coefficients in the wavelet space, rather than in the pixel space. This allows us to selectively attenuate/enhance frequency regions of the 2-D Fourier plane:



This repo uses my pytorch implementation of the dtcwt: pytorch_wavelets. You can install this however just by pip installing the requirements.txt.

You can pip install this directory as well, but you do not need to to run experiments.


The whole suite of tests to create Table 1 can be run by running the 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.

python -h

Gives full information on all the options available to run, but perhaps a good simple first test is:

python OUTDIR/run1 --net_type lenet_gainlayer  \
    --dataset cifar100  \
    --eval_period 4 \
    --optim adam --lr 0.001 \
    --trainsize 10000
python OUTDIR/run2 --net_type lenet \
    --dataset cifar100 \
    --eval_period 4 \
    --optim adam --lr 0.001 \
    --trainsize 10000

This runs both the LeNet and WaveLeNet architectures on cifar100 with a subset of the training set.