Use callable object for patching dataloaders

[shoarora]

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
• Did you read the contributor guideline?
• Did you make sure to update the docs?
  • not sure if there's anywhere I need to
• Did you write any new necessary tests?
• If you made a notable change (that affects users), did you update the CHANGELOG?

What does this PR do?

Fixes #968 .

Changes how we patch dataloaders when passed to trainer.fit() so that it's pickleable

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

[shoarora]

@williamFalcon @Borda

Do we need tests for this?
Not sure if this needs a CHANGELOG change either. would appreciate review!

[shoarora]

Getting a very unhelpful error, will keep investigating

Exception                                 Traceback (most recent call last)
<ipython-input-15-bef80bcdc7b4> in <module>()
     16 # trainer = Trainer(num_tpu_cores=8)
     17 
---> 18 trainer.fit(model, loader)

2 frames
/usr/local/lib/python3.6/dist-packages/torch/multiprocessing/spawn.py in join(self, timeout)
    110                 raise Exception(
    111                     "process %d terminated with exit code %d" %
--> 112                     (error_index, exitcode)
    113                 )
    114 

Exception: process 0 terminated with exit code 1

[shoarora]

Update: now if you pass your dataloaders to trainer.fit():

• non-slurm ddp works in a VM/linux context
• non-slurm ddp does not work in colab
  • the colab stack trace (my previous message) is unhelpful, and I'm having trouble figuring out how to get to the bottom of it
  • on the other hand, ddp in a colab context is an impractical scenario
• colab TPU worked regardless of this PR
• non-colab TPU works

Not sure what a change of this level warrants in terms of testing and documentation for pl. Please let me know!

I added a gpu-ddp test for passing all the loaders to fit() and confirmed that this branch allows the test to pass

[williamFalcon]

this is awesome! will pr review in a few hours!
thank u

[williamFalcon]

i think ddp in colab needs the spawn flag or something

[Borda]

LGTM, just curious if simple functools would help too which does very similar wrapping

from functools import partial

dataloader = partial(dataloader)

[shoarora]

LGTM, just curious if simple functools would help too which does very similar wrapping

from functools import partial

dataloader = partial(dataloader)

partial is specifically for wrapping functions/callables and freezing sets of arguments, so you couldn't call it on a dataloader:

>>> from functools import partial
>>> x = partial(loader)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: the first argument must be callable