Implement downstream evaluation

[palonso]

I propose leveraging and extending mir_ref for the evaluation of all the tasks of interest.

We could create a predict() method for speed up as much as possible the extraction of embeddings from our models

[raraz15]

I have been reviewing mir_ref. Here are some of the potential challenges I've identified:

1. Training of the probes is done with tensorflow. It would be much easier to stick to PyTorch lightning, which would avoid the weight conversion dependency. In addition, during fine-tuning in BEST-RQ and MusicFM, an unfrozen SSL model is used. Implementing this would also be much easier without changing the framework.
2. We would have more control over the downstream tasks and how the evaluations are implemented. I identified 4 downstream tasks implemented in mir_ref that are of relevance. We could implement them ourselves without much difficulty.
3. Avoiding the need to understand and implement someone else's code. What needs to be done here is not the same as adapting an image transformer to audio, I believe it would take unnecessary time. Although continuing this project would be nice, we do not know how much support we can count on given the timeline of the BSC grant.

I am open to discussion and talking about more on this. Let me know what you think, please.

[palonso]

Finally, we decided to go for a custom implementation of the downstream validation started on #38.
Let's repurpose this issue for additional features that should be implemented in the downstream package:

• Extract embeddings from an arbitrary layer.
• Store configuration files next to the embeddings.
• Support grid search, parameter ranges, or more advanced hyperparameter optimization techniques.
• LR scheduler in the downstream probe.
• weight decay in the downstream probe.