FSD50K Speech Model Fine-tuning Tutorial

[FlorentMeyer]

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you make sure to update the docs?
• Did you write any new necessary tests?

What does this PR do?

Add FSD50K Speech Model Fine-tuning Tutorial.

PR review

Did you have fun?

A lot 🙃

[codecov]

Codecov Report

Merging #201 (9794cfc) into main (89b94ba) will not change coverage.
Report is 1 commits behind head on main.
The diff coverage is n/a.

Additional details and impacted files

@@         Coverage Diff         @@
##           main   #201   +/-   ##
===================================
  Coverage    73%    73%           
===================================
  Files         2      2           
  Lines       382    382           
===================================
  Hits        280    280           
  Misses      102    102           

[rohitgr7]

hey @FlorentMeyer, mind check the file you uploaded, looks like it's too big and there might be some redundant stuff here. Might clean it up?

[Borda]

not sure what happen but GH does not want to show me the diff :/

[FlorentMeyer]

Good evening,

I should mention that the code in the converted notebook was exactly the same as in this Colab notebook (having removed the !pip installs). I also kept the output, but reading other people's examples I suppose that the outputs printed inside the docs are the ones obtained by running the .py converted notebooks on your side.

My last commit therefore makes these changes to the linked Colab notebook:

• remove all cells outputs
• remove conditions on bash instructions (a single bash command inside an if was causing a syntax error due to the absence of Python code)
• remove %% magic
• comment out the tensorboard cell (which was responsible for creating such a large file, I am sorry I hadn't checked it before)

Changes to the .yaml file:

• add gdown as a requirement
• remove brackets around my name (didn't know they were special characters)

I'm just not sure whether the Pandas dataframes with the audio players will get rendered.

[rohitgr7]

mind use torchmetrics functional metrics?
https://torchmetrics.readthedocs.io/en/stable/classification/average_precision.html#functional-interface

[FlorentMeyer]

At the time I wrote the code, torchmetrics.functional.average_precision's target took "integer labels" therefore not accepting multi-hot labels. Just let me check whether this was fixed and if I get the same results as with scikit-learn!

[FlorentMeyer]

OK, the new implementations of multilabel_average_precision give the same results as scikit-learn

[rohitgr7]

let's use that :)

[rohitgr7]

also, it looks like you have true values for preds, I'd recommend using test_step instead to show the metrics.

[FlorentMeyer]

What do you mean? Something like this?

With on_step=False, on_epoch=True to only log the end of the epoch according to https://pytorch-lightning.readthedocs.io/en/stable/extensions/logging.html#logging-from-a-lightningmodule:

The above config for validation applies for test hooks as well.

Suggested change

	mAP_micro = average_precision_score(
	# In __init__:
	self.mAP = torchmetrics.classification.MultilabelAveragePrecision()
	# In test_step:
	self.mAP(preds, y)
	self.log('mAP', self.mAP)

[FlorentMeyer]

Then the class version of torchmetrics should be prefered to functional I'd say?

[rohitgr7]

okay.. it's fine.. let's use functional metrics here since you already have all the targets and predictions.
modular metrics are useful, when you are aggregating the metrics, let's say on step-level

[FlorentMeyer]

I see there are still problems with:

1. the git+<my_repo> in requirements.txt
2. the building of the docs saying some cells are missing IDs even though I used jupytext as requested

[FlorentMeyer]

OK I also saw that there were bizarre things happening in the notebook, looks like the pre-commit hooks are moving stuff around causing duplication every time I pull them into my own code before being able to push again (example) and it's easy to miss things when reading a notebook as a .py file.

Anyway I read the whole file carefully and this should be fixed now. Also all cells have an ID so I'm not sure where the "cells are missing IDs" error comes from :/

[FlorentMeyer]

Small up!

[Borda]

lets remove all the ID