[Audio] SSL Pretraining framework for flow-matching model for audio processing

[Kuray107]

What does this PR do ?

This PR adds a ssl-pretraining framework and a flow matching AudioToAudio model

Collection: audio

Changelog

• Add a flow matching generative model in audio/models/enhancement.py
• Create audio/modules/ssl_pretrain_masking to partially-masked out input in the ssl pre-training stage.
• Create audio/parts/submodules/transformerunet.py to implement a Transformer U-Net structure.
• Create audio/parts/submodules/flow.py for training/inferencing with the flow matching generative model
• Create audio/parts/utils/callbacks.py to add processed audio in the training logger.

Usage

Model can be first be pre-trained using

python examples/audio/audio_to_audio_train.py --config-name flow_matching_generative_ssl_pretraining.yaml

• Please update the train_ds.manifest_filepath and validation_ds.manifest_filepath with the correct filepaths

After the ssl-pretraining stage is completed, the pretrained model could be finetuned using

python examples/audio/audio_to_audio_train.py --config-name flow_matching_generative_finetuning.yaml

• Remember to update the init_from_nemo_model with the pretrained checkpoint path in the config file.

Alternatively, one can also just train a flow matching generative from scratch:

python examples/audio/audio_to_audio_train.py --config-name flow_matching_generative.yaml

Model can be evaluated using

python examples/audio/audio_to_audio_eval.py \
    model_path=${PATH_TO_NEMO_FILE} \
    dataset_manifest=${PATH_TO_DATASET_MANIFEST} \
    output_dir=${PATH_TO_OUTPUT_DIR} \
    input_key=noisy_filepath \
    target_key=clean_filepath \
    metrics=[sdr,estoi]

Sampler setting can be easily overridden, e.g., to change the number of sampling steps as

python examples/audio/audio_to_audio_eval.py ... ++sampler.num_steps=${NUM_STEPS}

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• [] Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

[racoiaws]

Is this identity? I would expect identity to be weight = torch.eye and bias = torch.zeros

[racoiaws]

Oh, this is identity in the sense that output of the whole module is identity

[racoiaws]

Is this classifier-free guidance? Is there a reason to have it enabled in the default "generative model" config?

[Kuray107]

p_cond is only effective during the training process. In the inference stage one could choose whether to use classifier-free guidance or not.

[anteju]

LGTM