transfer_batch_to_gpu not moving List[torch.tensor] and like datastructures to GPU

[Laksh1997]

🐛 Bug

trainer.transfer_batch_to_gpu(torch.randn(3, 3), 0) gives:

tensor([[-1.0600,  0.3306,  1.1276],
        [ 1.0012, -0.2687, -0.5493],
        [-0.5619, -1.7161, -0.8625]], device='cuda:0')

But
trainer.transfer_batch_to_gpu([ torch.randn(3, 3) ], 0) gives

[tensor([[ 1.3631,  0.3408, -1.1074],
         [-1.1176, -0.8056,  0.2937],
         [-0.4235,  0.7321, -0.8811]])]

which is not the expected behaviour as we are supposed to move all tensors in the data structure onto GPU.

Same issue occurs for tuple and dictionary datastructures.

Also - would it be possible to get gpu transfer for argparse.Namespace datastructures, or class datastructures?

IE if I make a class like so:

class Batch(object):
    def __init__(self, 
                 src_input_ids,
                 tgt_input_ids,
                 src_attn_mask,
                 tgt_attn_mask,
                 src_seg_ids,
                 tgt_labels
                ):
        self.__dict__.update(locals())

it would be cool if I could directly pass this in with gpu transfer.

Thanks a lot !

To Reproduce

Just copy one of the examples to get the Trainer class.

Then do trainer=Trainer(...) and run the above code

Code sample

Expected behavior

Environment

Python 3.6
Pytorch 1.3.1
Cuda 10.1
PytorchLightning master (as of today)

[awaelchli]

I cannot reproduce this on master. For me
trainer.transfer_batch_to_gpu([ torch.randn(3, 3) ], 0)

returns

[tensor([[ 1.6940, -0.8183,  0.6075],
        [ 0.5152,  0.2925, -0.3388],
        [ 0.9064,  0.2509,  0.7852]], device='cuda:0')]

as expected. Same with tuples and dicts. Was it perhaps fixed in the meantime? Could you try again @Laksh1997?

[williamFalcon]

we can reopen if it's still an issue