Documentation for adapter fine-tuning

[marcoyang1998]

Add documentation for #1512.

[YuriiMytiai]

Hi!
Do you have accuracy metrics for the case when you fine-tune whole ASR model on the same data?

[marcoyang1998]

Hi, fine-tuning the whole model gives us 13.31/13.39. You may want to have a look at #1484 for more reference numbers.

[1215thebqtic]

Hi,

I conducted experiments with my own data and found that the training time for adapter-finetune and full-finetune are almost the same. I print weight values of different epochs of adapter-finetune, and only adapters' weights change. Is this normal? Thank you!

[JinZr]

yes this is normal. you need to look into the implementation and technical details of adapter based finetuning to gain deeper understanding. best regards jin

…

On May 28, 2024, at 15:26, 1215thebqtic ***@***.***> wrote: Hi, I conducted experiments with my own data and found that the training time for adapter-finetune and full-finetune are almost the same. I print weight values of different epochs of adapter-finetune, and only adapters' weights change. Is this normal? Thank you! — Reply to this email directly, view it on GitHub <#1545 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOON42CM5RETC7FIRVZABYTZEQWQVAVCNFSM6AAAAABEROZKC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMZUGUZDCOBZGY>. You are receiving this because you are subscribed to this thread.

[1215thebqtic]

yes this is normal. you need to look into the implementation and technical details of adapter based finetuning to gain deeper understanding. best regards jin
…
On May 28, 2024, at 15:26, 1215thebqtic @.***> wrote: Hi, I conducted experiments with my own data and found that the training time for adapter-finetune and full-finetune are almost the same. I print weight values of different epochs of adapter-finetune, and only adapters' weights change. Is this normal? Thank you! — Reply to this email directly, view it on GitHub <#1545 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOON42CM5RETC7FIRVZABYTZEQWQVAVCNFSM6AAAAABEROZKC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMZUGUZDCOBZGY. You are receiving this because you are subscribed to this thread.

The doc said adapter-finetune is much faster than full-finetune, so I'm confused.

https://k2-fsa.github.io/icefall/recipes/Finetune/adapter/finetune_adapter.html#fine-tune-with-adapter

[marcoyang1998]

Could you show me the average time for finishing one epoch for both of your experiments?

In my experiment, adapter-based training took ~25 min per epoch, and the full-finetune took ~30 min per epoch.

[1215thebqtic]

adapter-finetune is 60min/epoch, full-finetune is 46min/epoch. Adapter finetune costs more time. I change base_lr=0.00005, lr_epochs=100, lr_batches=100000 for full-finetune. I use default values for adapter finetune.

[marcoyang1998]

Adapter finetune costs more time.

That's unusual, how big is your model? BTW, your base-lr seems very small (even for full fine-tuning), which might lead to poor performance.

[1215thebqtic]

For model with adapter: Number of model parameters: 169281897, A total of 1234624 trainable parameters (0.729% of the whole model)