leaf-audio-pytorch

Pytorch port of Google Research's LEAF Audio paper published at ICLR 2021.

This port is not completely finished, but the Leaf() frontend is fully ported over, functional and validated to have similar outputs to the original tensorflow implementation. A few small things are missing, such as the SincNet and SincNet+ implementations, a few different pooling layers, etc.

PLEASE leave issues, pull requests, comments, or anything you find in using this repository that may be of value to others who will try to use this.

Installation

From the root directory of this repo, run:

pip install -e .

Usage

leaf_audio_pytorch mirrors it's original respository; imports and arguments are the same.

import leaf_audio_pytorch.frontend as frontend

leaf = frontend.Leaf()

Installation for Developing

If you are looking to develop on this repo, the requirements.txt contains everything needed to run the torch and tf implementations of leaf audio simultaneously.

NOTE: There is some weird dependency stuff going on with the original leaf-audio repo. Seems like its a dependency issue with lingvo and waymo-open-dataset. These below commands are a workaround.

Install the packages required:

pip install -r requirements.txt --no-deps

Install the leaf-audio repo from Git SSH:

pip install git+ssh://git@github.com/google-research/leaf-audio.git --no-deps

Then add the leaf_audio_pytorch package as well

python setup.py develop

At this point everything should be good to go! The scripts in test/ contains some testing code to validate the torch implementation mirrors tf.

Some Things to Keep in Mind (PLEASE READ)

When writing this port, I ran a debugger of the torch and tf implementations side by side and validated that each layer and operation mirrors the tensorflow implementation (to within a few significant digits, i.e. a tensor's values may variate by 0.001). There is one notable exception: The depthwise convolution within the GaussianLowpass pooling layer has a larger variation in tensor values, but the ported operation still produces similar outputs. I'm not sure why this operation is producing different values, but i'm currently looking into it. Please do your own due diligence in using this port and making sure this works as expected.
As of March 29, I finished the initial version of the port, but I have not tested Leaf() in a traning setting yet. Calling .backward() on Leaf() throws no errors, meaning backprop works as expected. However, I do not yet know how this will function during training.
As PyTorch and Tensorflow follow different tensor ordering conventions, Leaf() does all of its operations and outputs tensors with channels first.

Reference

All credit and attribution goes to Neil Zeghidour and the Google Research team who wrote the paper and created the Tensorflow implementation.