It's Raw! Audio Generation with State-Space Models

Karan Goel, Albert Gu, Chris Donahue, Christopher Ré

ICML 2022

Samples from a SaShiMi model trained on MNIST. Green parts represent the priming signal. Red parts are generated by the model.

Listen to our audio samples in this link. (Mirror)

This repository provides a re-implementation of this paper in PyTorch, developed as part of the course METU CENG 796 - Deep Generative Models. This re-implementation is provided by:

• İlker Işık, [email protected]
• Muhammed Can Keleş, [email protected]

Please see the following Jupyter Notebooks for more information:

• main.ipynb: Summary of the paper, implementation notes and our experimental results.
• Ablation.ipynb: Our experiments with 2-Layer SaShiMi on Youtube Mix dataset.
• S4-MNIST-Demo.ipynb: A demo on MNIST dataset that showcases the S4 layer, which is the main building block in SaShiMi.
• SaShiMi-MNIST-Demo.ipynb: Training and testing the SaShiMi model on MNIST dataset.

Installation

PyTorch is required. See PyTorch installation page for more info. Here's how to install PyTorch with pip:

pip3 install torch torchvision torchaudio

Following libraries are required for dataset handling:

pip3 install numpy scipy

Optional: PyKeOps can be installed for more memory-efficient Cauchy kernel computation. Install PyKeOps using pip:

pip3 install pykeops

If that doesn't work, try:

pip3 install pykeops[full]

Downloading the Data

This Hugging Face repository contains the pre-trained models together with the YouTube Mix dataset, which are not included in this repository due to their large file sizes. To download these, you can simply run download_data.sh. wget and unzip programs must be installed.

./download_data.sh

Note that MNIST dataset comes from PyTorch's MNIST class, it will be downloaded automatically at runtime when needed.

Training Script

Since the training on the full dataset takes a lot of time, we have created a dedicated training script for that instead of putting it into the Jupyter Notebooks.

The script named train.py will train the SaShiMi model on the Youtube Mix dataset located at ./datasets/youtube-mix/train/ and save the model checkpoints at ./models/ym-8l/ after each epoch:

python3 train.py

We used variations of train.py for the ablation experiments.

In the first run, this script will begin the training from scratch and continue indefinitely until a keyboard interrupt occurs. In the successive runs, it will load the latest saved checkpoint and continue.

This script doesn't have a command line interface, nor is it configurable via other files. Since the script is really simple, it is intended to be modified. For example, you can interrupt the training at any time and change the learning rate by modifying the following part:

# Update LR
for group in optimizer.param_groups:
    group["lr"] = 0.001

Unit Tests

This repository contains numerous unit tests for both S4 and SaShiMi.

Run all unit tests with:

python3 -m unittest

We also have a GitHub Actions Workflow for running these tests.

Cauchy Kernel Benchmark

S4/cauchy.py can be run as a standalone script. It will perform the same Cauchy kernel computation using the naive and PyKeOps method, and then compare the results.

Run the following script to get more information about the command line arguments:

python3 S4/cauchy.py -h

If you run the benchmark with a large enough sequence length, the naive method will fail due to out of memory error. PyKeOps, on the other hand, should be able handle this with no problems:

python3 S4/cauchy.py -l 64000

It also reports the maximum difference between the matrices computed by these two methods. Currently, this value is quite large (~0.001); however, all unit tests that compare these two methods pass. This might be caused by the random initialization of the inputs in the benchmark.