Misleading exception raised during batch scaling

[tejasvi]

Use batch_size from model.hparams.batch_size instead of model.batch_size

[justusschock]

Hi @tejasvi . Thanks for contributing, but I probably wouldn't merge this, sice we recently moved away from hparams in #1896

[Borda]

@tejasvi mind update it to reflect we new structure of arguments?

[pep8speaks]

Hello @tejasvi! Thanks for updating this PR.

• In the file pytorch_lightning/trainer/training_tricks.py:

Line 259:58: W291 trailing whitespace

Comment last updated at 2020-06-08 16:17:27 UTC

[justusschock]

What about something like:

splitted_batch_arg_name = batch_arg_name.split('.')
contains_arg = model
for name in splitted_batch_arg_name[:-1]:
    contains_arg = getattr(contains_arg, name)

batch_size = getattr(contains_arg, splitted_batch_arg_name[-1])

if value:
    setattr(contains_arg, splitted_batch_arg_name[-1], value)

This way one could either specify the name to something like 'batch_size' (which would correspond to model.batch_size) or to 'hparams.batch_size' (which would correspond to model.hparamsbatch_size).

The issue is, that arguments are not always attached and hparams can also be under other names (like 'hyperparameters'). These issues would all be resolved with something like this I guess :)

[SkafteNicki]

We had a similar problem, when the learning rate finder was implemented, namely that some users like to store their learning rate in something like model.hparams.optimizers.lr instead of just model.hparams.lr. We solved it by introducing these two utility functions:
https://github.com/PyTorchLightning/pytorch-lightning/blob/c3cf33d1de7de3444e3a4a7d45bc613d8e01a7c0/pytorch_lightning/trainer/lr_finder.py#L485-L501
Which can be repurposed for this PR (move to utilities.parsing?). A similar nested_getattr needs to be implemented.

[Borda]

we are not registering in the model anymore...

[williamFalcon]

sorry, merged by accident. mind re-opening?
(merge has been reversed)

[Borda]

sorry, merged by accident. mind re-opening?
(merge has been reversed)

merged cannot be reopened, it is already in master
and reverting and immediate merge does not solve the issue much
@williamFalcon pls ping me next time...

[Borda]

@tejasvi mind open a new PR?