[feature request] ModuleTrainer class

[ncullen93]

Could potentially add a ModuleTrainer or ModelTrainer class that works similarly to SuperModule but can take in one or more normal nn.Module classes.. This would allow for support of pre-trained networks and more flexible training structures, while also allowing seamless integration with all other pytorch code

@recastrodiaz

[recastrodiaz]

Sounds like a great idea! ;)
I'll get back to you with some questions Monday or sooner if possible.

[ncullen93]

check out what I just put in modules/model_trainer.py . Added initializations as well.

[recastrodiaz]

Amazing! This is definitely what I had in mind! I saw that you added support for multiple forwards inputs*. That's something I was really looking for!

The History callback class seems to be looking for non existing attributes in the model. Maybe the History constructor should accept a ModuleTrainer as init parameter? (Although this sounds like a circular dependency...). I've removed it from the callbacks from now in ModuleTrainer __init__ function:

self._callbacks = [self.history]

#-->

self._callbacks = []

History class:
/Users/rodrigo/Libs/torchsample/torchsample/callbacks.py in on_train_begin(self, logs)
    168     def on_train_begin(self, logs=None):
    169         self.losses = []
--> 170         if self.model._has_regularizers:
    171             self.regularizer_losses = []
    172         if self.model._has_lagrangian_constraints:

Thanks again for the excellent work!

[ncullen93]

should be fixed can you check?

[ncullen93]

closed bc ModuleTrainer was added

[recastrodiaz]

Can confirm ModuleTrainer works!