pin_memory should be true only if gpus are specified

[igormq]

Hey folks,

First of all, thank you for this amazing package. It has saved many hours of coding :)

I was using for the first time the pl data modules when I hit the following line
cifar10_datamodule.py#L132. The problem is: when I am training/evaluating/testing on CPU, the data module try to pin the data and try to reserve a batch size memory amount in gpu. Nothing goes wrong if I have GPU with enough memory ran, but, when I try to run in CPU mode with all my GPUs in use I get the following error:

RuntimeError: Caught RuntimeError in pin memory thread for device 0.
Original Traceback (most recent call last):
  File "/lib/python3.7/site-packages/torch/utils/data/_utils/pin_memory.py", line 31, in _pin_memory_loop
    data = pin_memory(data)
  File "/lib/python3.7/site-packages/torch/utils/data/_utils/pin_memory.py", line 55, in pin_memory
    return [pin_memory(sample) for sample in data]
  File "lib/python3.7/site-packages/torch/utils/data/_utils/pin_memory.py", line 55, in <listcomp>
    return [pin_memory(sample) for sample in data]
  File "lib/python3.7/site-packages/torch/utils/data/_utils/pin_memory.py", line 47, in pin_memory
    return data.pin_memory()
RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1595629427478/work/aten/src/THC/THCCachingHostAllocator.cpp:278

The fix is simple and we can borrow the code from the pl core implementation as in pytorch_lightning/accelerators/accelerator_connector.py#L90

Bests.

[github-actions]

Hi! thanks for your contribution!, great first issue!

[nateraw]

@igormq mind sending a PR? if not I can address this.

[nateraw]

I think the main thing here is to just include the pin_memory flag in the datamodules. I think by default it makes sense to set it to True either way? please correct me if I'm wrong 😄

[briankosw]

@nateraw I think even beyond the pin_memory flag, some other flags should also be included such as the shuffle and drop_last flag. What do you think? I can work on a pull request for updating the flags in the datamodules.

[nateraw]

@briankosw sure! for the shuffle and drop_last flags, lets just make sure we use sensible defaults. 😄 Thanks!!