LORA Implementation for Parameter Efficient Fine-Tuning on new datasets

[sidhantls]

Addresses #158 and the request for PEFT mentioned in the README.

Feature

This PR adds PEFT support with Low-Rank adapters (LORA) for fine-tuning Parler-TTS on new datasets.

LORA is applied to the Parler-TTS decoder Transformer where PEFT is applied to Linear projection layers. With the T5 encoder of Parler-Mini frozen, PEFT trains only 8% params, compared to 50% without PEFT

Benefits

• PEFT enables Fine-tuning of Parler-Mini on 8GB GPU (with limited offloading?).
  • On my windows machine, it takes 29.52s/it with PEFT and 117.50s/it without PEFT for fine-tuning.
• PEFT enables Fine-Tuning of Parler-TTS Large, 2.3B on Google Collab. Without, I get OOM error.

Results:

• Training plots with PEFT and without PEFT can be found in this Weights and Biases Report along with generated audio samples by the two models. This is fine-tuning on Jenny-6H.
• Audio Samples Comparison: Audio samples sounds very similar. In some cases, the PEFT model produces better (better pronunciation and no cut-off at the end): Refer to Index 47, 49

Reproduce:

• Google Collab: Notebook. To turn off PEFT, set USE_PEFT=false

To do:

• Analysis to see if Parameter Efficient Fine-Tuning will be useful to the community @ylacombe @sanchit-gandhi
• Complete PR Code: Saving Model:
  • After PEFT training is complete, it requires the adapters to be removed and saved. I have implemented this, but it is an in-place operation. So, how are we going to save checkpoints? For the final model, we can use this in-place operation with replaces adapters with regular layers so that the model can be loaded the same way without any other code changes

[sidhantls]

to do: remove this line