Ignite: Issue with traced models

[vfdev-5]

If user passes a traced model into create_supervised* methods and specifies a device, there will be the following error:

traced_model = torch.jit.trace(model, z)

evaluator = create_supervised_evaluator(traced_model, metrics, device='cuda')
> RuntimeError: to is not supported on TracedModules

We need to check if model have the method .to...

[vfdev-5]

Probably, we should wait until this issue is done.

EDIT: https://github.com/pytorch/pytorch/pull/15340

[vfdev-5]

As pytorch/pytorch#15340 has been merged, create_supervised* methods will work on ScriptModule and traced models.
We need to checked with tests on CI ...
@TheCodez if you would like to contribute, you can add a test with traced and scripted models passing to create_supervised* method with device cpu and the test should pass for pytorch-nightly build (can be check with torch.__version__)

[TheCodez]

the test should pass for pytorch-nightly build (can be check with torch.__version__)

Do you mean only executing the test if on the nightly, or check if it throws a RuntimeError if not on the nightly?

[vfdev-5]

@TheCodez yes something like this.

[TheCodez]

@vfdev-5 which torch.__version__ to check for nightly?

[vfdev-5]

@TheCodez you can do it like this:

# Check if torch-nightly
if "dev" in torch.__version__:
     ...
else:
    ...