Full callback handling

[schwobr]

Is your feature request related to a problem? Please describe.
I started deep learning using fastAI, and despite all its drawbacks there is one thing I found very handy when I wanted to tweek the training loop : callbacks. This is far less important here as we have control over training_step, but I was for instance wondering how I could do something when train begins (like log my model's graph). Maybe I'm missing something but it seems to me that there is no simple way to do that.
The fact that there is a Callback class defined (used for early stopping and checkpointing) makes me think that it was somewhat planned at some point, but was never fully integrated (or I may be missing something).

Describe the solution you'd like
Implement callback handling with a few methods (the one defined in Callback) that can be called at specific points during the training loop. Early stopping and checkpointing can be added as default callbacks and integrated within this new framework. The idea is to have something simple to interact with the training loop at specific points.

Describe alternatives you've considered
Let things as they are and just add the callback directly in pl.LightningModule as optional methods to be implemented. They then just need to be called at the right time during the training loop. It doesn't change much compared to traditional callbacks but it may be closer to the design of pytorch-lightning.

If you find that there is any value in this idea I could try and write a PR in the following weeks.

[williamFalcon]

in lightning you do whatever in any hook you want.

https://pytorch-lightning.readthedocs.io/en/latest/Trainer/hooks/

[schwobr]

Indeed, I missed that one. I guess I would add a on_train_begin and on_train_end to HookModel then. Or is there some other way to interact with train beginning and ending ?

[williamFalcon]

those are the ways! but if we need more hooks, feel free to submit a PR!

[schwobr]

Thanks, I'll look into it !