[TTS] Add univnet scheduler

[ArtyomZemlyak]

What does this PR do ?

Add scheduler for univnet model.

Collection: [TTS]

Changelog

• Add get_warmup_steps into UnivNetModel
• Change configure_optimizers in UnivNetModel for scheduler support

Usage

Standart usage of scheduler

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.

Additional Information

• Related to # (issue)

[ArtyomZemlyak]

@okuchaiev
Hi!
Can you review?

[XuesongYang]

LGTM. Thanks. Out of my curiosity, after applying these schedulers to univnet, did you observe better performance over the one without explicitly applying schedulers?

[ArtyomZemlyak]

@XuesongYang
Some improvements exists. But need more training tests.

sched - CosineAnnealing with warmup ratio 0.02
Yellow - its finetuning from another language and voice.

[XuesongYang]

@XuesongYang Some improvements exists. But need more training tests.

Thank you for sharing! From your plots, it seems promising by tweaking the scheduler's parameters.

As a side note: we have shifted the vocoder model to hifigan from univnet. So probably we will not invest to improve univnet. But again, you are welcome to share any practical and better configuration setups of univnet if you achieve better performance. We can then pick it up, train/fine-tune it again, and update our univnet NGC model checkpoints. Please reach out in the issues ;)