Fastai: auc_roc_score metric fails with GPU tensors

[benvenutto]

Describe the bug

When using 'auc_roc_score' as one of the tabular model metrics, at the end of the 1st epoch you get the following error:

RuntimeError: Expected object of backend CUDA but got backend CPU for sequence element 1 in sequence argument at position #1 'tensors'

Execution of the notebook fails.

Provide your installation details

=== Software ===
python        : 3.6.8
fastai        : 1.0.51.dev0
fastprogress  : 0.1.20
torch         : 1.1.0a0+be364ac
nvidia driver : 418.39
torch cuda    : 10.1.105 / is available
torch cudnn   : 7500 / is enabled

=== Hardware ===
nvidia gpus   : 1
torch devices : 1
  - gpu0      : 8119MB | GeForce GTX 1070 Ti

=== Environment ===
platform      : Linux-4.15.0-47-generic-x86_64-with-debian-stretch-sid
distro        : Ubuntu 16.04 xenial
conda env     : Unknown
python        : /opt/conda/bin/python
sys.path      :
/opt/conda/lib/python36.zip
/opt/conda/lib/python3.6
/opt/conda/lib/python3.6/lib-dynload
/opt/conda/lib/python3.6/site-packages
/opt/conda/lib/python3.6/site-packages/torchvision-0.2.3-py3.6.egg
/Docker/workdir/fastai

To Reproduce

1. load the Tabular example notebook in the Fastai library (examples directory)
2. add 'auc_roc_score' to the list of metrics on line 9:

learn = tabular_learner(data, layers=[200,100], metrics=[accuracy, auc_roc_score])

1. run the notebook

Expected behavior

The notebook should complete execution, with the additional metric.

Screenshots

Additional context

Running inside 'nvcr.io/nvidia/pytorch:19.03-py3' Docker container, under Ubuntu 18.04.

[sgugger]

This is now fixed in master. Please note that auc_roc_score won't work directly, you have to use the AUROC() metric (unless your outputs have the same size as your targets, but that's not the case in this tabular example).

[benvenutto]

Many thanks @sgugger I tried that earlier, and got an error from _init_ saying it got 3 params not 1, will do a bit more reading :)

[frasermince]

@benvenutto I'm currently getting the same thing. Any luck?

[benvenutto]

This seems to work:

learner = tabular_learner(train_data, layers=[2000, 100], metrics=[accuracy, AUROC()])

You need to instantiate AUROC as it's a class, not function, which wraps the metric. Which is exactly as @sgugger replied above, I wasn't paying attention :(