auto_scale_batch_size not working with datamodule

[carmocca]

🐛 Bug

The Trainer expects the LightningModule to have self.batch_size (see scale_batch_size() in training_tricks.py). However, if one is using the new LightningDataModule, that should be the class with self.batch_size defined.

To Reproduce

assert hasattr(lightning_data_module, "batch_size")
trainer = Trainer(auto_scale_batch_size=True)
trainer.fit(lightning_module, datamodule=lightning_data_module)

pytorch_lightning.utilities.exceptions.MisconfigurationException: Field batch_size not found in both `model` and `model.hparams`

Expected behavior

auto_scale_batch_size should work using LightningDataModule

Environment

* Packages:
	- numpy:             1.18.5
	- pyTorch_debug:     False
	- pyTorch_version:   1.6.0
	- pytorch-lightning: 0.9.1rc1
	- tensorboard:       2.2.0
	- tqdm:              4.48.2

[awaelchli]

I knew this bug report will eventually come :) same will happen for auto_lr_find.
We need to generalize our lightning_hasattr and getattr helper functions to include the datamodule. In total, we have

model.attribute_name
model.hparams.attribute_name
model.dm.attribute_name

all of these should be considered by both lr_find and auto_scale_batch_size

[tugot17]

@awaelchli So currently there is no solution to this issue?

[awaelchli]

no, you need to set it manually, sorry.
But I'll try to make a PR this weekend if not today. Should be a relatively easy fix unless I missed something.

[SkafteNicki]

@awaelchli I completely agree that it should be a simple fix, since all the attribute getting/setting is handled by our own lightning_getattr ect functions.

[awaelchli]

@carmocca As you may already know, we fixed this. Just a note in case you didn't see it, you now need to call .tune() instead of .fit():

trainer.tune(lightning_module, datamodule=lightning_data_module)

This is to better distinguish the training from the tuning step. However, it may be subject to change since there are some refactors happening right now.

[tugot17]

@awaelchli
In which version tune will be officially introduced?

[awaelchli]

If it stays, in v1.0

[carmocca]

I am not familiar with the tuning interface. Few questions:

Right now, I can use auto_scale_batch_size by passing the option to the trainer and calling fit. This does the auto scale procedure and then starts training

If I understand correctly, when tune is released the following will be the necessary:

should_auto_scale_bs = # comes from the user
trainer = Trainer(auto_scale_batch_size=should_auto_scale_bs)

if should_auto_scale_bs:
    trainer.tune(...)
trainer.fit(...)

Im assuming tune doesnt run fit automatically when it is finished.

What would happen if the code ran trainer.tune(...) outside of the if and auto_scale_batch_size was False?
Also, shouldn't auto_scale_batch_size be a parameter of tune instead of Trainer? (Maybe it is, I just don't know where is the tune discussion).

[awaelchli]

I do not know how to answer these questions, @williamFalcon added the tune method so it would be best to ask him how he sees it being used in the future.

[SkafteNicki]

@carmocca I think the process in the future will be:

trainer = Trainer(auto_scale_batch_size = should_auto_scale_bs)
trainer.tune(model) # will do nothing if should_auto_scale_bs=False
trainer.fit(model)

it should still be easy for the user to use these features. The moving of the tuning from fit to tune is to disentangle the hyperparameter tuning from the actual optimization of the network. This will make it easier for us in the future to implement more tuning algorithms.