Failing test: test_running_test_pretrained_model_ddp

[neggert]

I think this is another problem stemming from the fact that we don't have a way to pass data back from torch.multiprocessing.spawn. Needs more investigation.

def test_running_test_pretrained_model_ddp(tmpdir):
        """Verify `test()` on pretrained model."""
        ...
        # run test set
        new_trainer = Trainer(**trainer_options)
>       new_trainer.test(pretrained_model)
tests/test_restore_models.py:60:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
pytorch_lightning/trainer/trainer.py:1189: in test
    self.run_evaluation(test_mode=True)
pytorch_lightning/trainer/evaluation_loop.py:299: in run_evaluation
    if test_mode and not self.is_overriden('test_step'):
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
self = <pytorch_lightning.trainer.trainer.Trainer object at 0x7f845ec23f90>, f_name = 'test_step', model = None
    def is_overriden(self, f_name, model=None):
        if model is None:
            model = self.get_model()
        super_object = LightningModule
        # when code pointers are different, it was overriden
>       is_overriden = getattr(model, f_name).__code__ is not getattr(super_object, f_name).__code__
E       AttributeError: 'NoneType' object has no attribute 'test_step'
pytorch_lightning/trainer/model_hooks.py:20: AttributeError

[williamFalcon]

https://pytorch.org/docs/stable/notes/multiprocessing.html#reuse-buffers-passed-through-a-queue

torch.multiprocessing is a drop in replacement for Python’s python:multiprocessing module. It supports the exact same operations, but extends it, so that all tensors sent through a python:multiprocessing.Queue, will have their data moved into shared memory and will only send a handle to another process.

Looks like we can use python:multiprocessing.Queue?

[Borda]

@williamFalcon Could we check these two commits on GPU - 20d15c8 and 5dd2afe?