Design debate: Should we use hydra's object instantiation feature?

[vinamarora8]

In my opinion, good research release code is simple, easy to read, and transparent. Instantiating models and tokenizers using hydra.utils.instantiate(), however, is very opaque and requires readers to understand hydra's configuration loading system.
This discussion is only regarding examples/poyo/train.py. I like to use this hydra feature when developing code and doing experiments too. But I don't see it having a place in released research code.

[mazabou]

Currently hydra.utils.instantiate() is used for the model, the transforms, and the metrics.

1. For the model, I agree we can call POYOPlus directly. The only advantage of having the class in the config file is that it makes it super clear what the destination for the model is. which is actually pretty neat. if we remove it, it might not be as explicit. of course we can always add a comment.

2. For the transforms, composing augmentations is pretty neat imo. what would be the alternative here?

3. For the metrics, they are hidden, so should we care about this?