fad_pytorch

Original FAD paper (PDF)

Install

pip install fad_pytorch

Features:

• runs in parallel on multiple processors and multiple GPUs (via accelerate)
• supports multiple embedding methods:
  • VGGish and PANN, both mono @ 16kHz
  • OpenL3 and (LAION-)CLAP, stereo @ 48kHz
• uses publicly-available pretrained checkpoints for music (+other sources) for those models. (if you want Speech, submit a PR or an Issue; I don’t do speech.)
• favors ops in PyTorch rather than numpy (or tensorflow)
• fad_gen supports local data read or WebDataset (audio data stored in S3 buckets)
• runs on CPU, CUDA, or MPS

Instructions:

This is designed to be run as 3 command-line scripts in succession. The latter 2 (fad_embed and fad_score) are probably what most people will want:

1. fad_gen: produces directories of real & fake audio (given real data). See fad_gen documentation for calling sequence.
2. fad_embed [options] <real_audio_dir> <fake_audio_dir>: produces directories of embeddings of real & fake audio
3. fad_score [options] <real_emb_dir> <fake_emb_dir>: reads the embeddings & generates FAD score, for real (“$r$”) and fake (“$f$”):

$$ FAD = || \mu_r - \mu_f ||^2 + tr\left(\Sigma_r + \Sigma_f - 2 \sqrt{\Sigma_r \Sigma_f}\right)$$

Documentation

See the Documentation Website.

Comments / FAQ / Troubleshooting

• “RuntimeError: CUDA error: invalid device ordinal”: This happens when you have a “bad node” on an AWS cluster. Haven’t yet figured out what causes it or how to fix it. Workaround: Just add the current node to your SLURM --exclude list, exit and retry. Note: it may take as many as 5 to 7 retries before you get a “good node”.
• “FAD scores obtained from different embedding methods are wildly different!” …Yea. It’s not obvious that scores from different embedding methods should be comparable. Rather, compare different groups of audio files using the same embedding method, and/or check that FAD scores go down as similarity improves.
• “FAD score for the same dataset repeated (twice) is not exactly zero!” …Yea. There seems to be an uncertainty of around +/- 0.008. I’d say, don’t quote any numbers past the first decimal point.

Contributing

This repo is still fairly “bare bones” and will benefit from more documentation and features as time goes on. Note that it is written using nbdev, so the things to do are:

1. Fork this repo
2. Clone your fork to your (local) machine
3. Install nbdev: python3 -m pip install -U nbdev
4. Make changes by editing the notebooks in nbs/, not the .py files in fad_pytorch/.
5. Run nbdev_export to export notebook changes to .py files
6. For good measure, run nbdev_install_hooks and nbdev_clean - especially if you’ve added any notebooks.
7. Do a git status to see all the .ipynb and .py files that need to be added & committed
8. git add those files and then git commit, and then git push
9. Take a look in your fork’s GitHub Actions tab, and see if the “test” and “deploy” CI runs finish properly (green light) or fail (red light)
10. Once you get green lights, send in a Pull Request!

Feel free to ask me for tips with nbdev, it has quite a learning curve. You can also ask on fast.ai forums and/or fast.ai Discord

Citations / Blame / Disclaimer

This repo is 2 weeks old. I’m not ready for this to be cited in your papers. I’d hate for there to be some mistake I haven’t found yet. Perhaps a later version will have citation info. For now, instead, there’s:

Disclaimer: Results from this repo are still a work in progress. While every effort has been made to test model outputs, the author takes no responsbility for mistakes. If you want to double-check via another source, see “Related Repos” below.

Related Repos

There are [several] others, but this one is mine. These repos didn’t have all the features I wanted, but I used them for inspiration:

• https://github.com/gudgud96/frechet-audio-distance
• https://github.com/google-research/google-research/tree/master/frechet_audio_distance: Goes with Original FAD paper
• https://github.com/AndreevP/speech_distances