model is moved too early in create_supervised_trainer/evaluator

Hi, first of all, thanks for the great library! I have a general “bug” report or question on the specific implementation of create_supervised_trainer and create_supervised_evaluator. As seen for example here: https://github.com/pytorch/ignite/blob/master/ignite/engine/__init__.py#L64

The moment a supervised evaluator/trainer is created the model is moved to the specific device and not during the inference stage.

I agree, that this version doesn’t have an impact in general, but I would argue that the model should be moved in the _inference function and not when I create the evaluator/trainer. A weird example but similar to my case: I have multiple evaluators and I would like to run one on the CPU and one on the GPU. The reason being that a different library doesn’t support GPUs yet and I would like to add it in an “ignite” style to keep the code similar. Now, the run steps will not work, as the model is not moved again in the inference stage.

I can imagine that there is a reason why it is the way it is, but I am wondering if a future version couldn’t do the model.to(device) step in the _inference function. 😃

Thanks!

[vfdev-5]

yes, please

On Tue, Mar 10, 2020, 09:47 kai-tub [e-mail] wrote:

Since this is labeled as “help wanted” should I issue a PR with the changes mentioned by @justusschock https://github.com/justusschock?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/pytorch/ignite/issues/830?email_source=notifications&email_token=AASYOHYP6NPNVYLKAVAO7BDRGX5BVA5CNFSM4LCQXVF2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEOKQP7Q#issuecomment-596969470, or unsubscribe https://github.com/notifications/unsubscribe-auth/AASYOH4PIKK5KUO7IVOBZ7LRGX5BVANCNFSM4LCQXVFQ .

[kai-tub]

Seems like it does not make sense at all to send model to a device in create_supervised_trainer while optimizer is already defined.

Yes, seems like it would be better to not move the model in trainer.

Concerning your use-case:

In my case, I only used these evaluators after the training procedure. So there was no need to move the model back in-between steps. But I initialized both at the same time and that is how I saw the “bug”.

I don’t see a good way how we could still move the model in the function. Right now, I would favor removing the model moving process and updating the docs.

But I can imagine that a lot of code uses this behavior to save one line of moving the model before initializing the optimizer. What is your opinion of trying to catch an error due to different devices and adding more details to the exception? Or should the docs suffice? Or do you want to add a deprecation warning?

Do you have an alternative approach?