Introduce a variable skip_unrolling in class Metric

[simeetnayan81]

Fixes #2940

Description:
Introduce a variable skip_unrolling in class Metric as discussed here https://discord.com/channels/831462531327328276/1110662056622964860/1253769540710567977

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

[vfdev-5]

Thanks for the PR @simeetnayan81
Left a comment about the implementation

[vfdev-5]

Let's also update above docstring by adding the new argument. Please also check CONTRIBUTING guideline the part about how to add .. versionadded:: tag in the bottom of the docstring. Version to put should be 0.5.1

[simeetnayan81]

Should I add .. versionchanged:: instead?

[vfdev-5]

Yes, correct, it should be versionchanged tag, https://github.com/pytorch/ignite/blob/master/CONTRIBUTING.md#writing-documentation

[simeetnayan81]

On it. Thanks.

[simeetnayan81]

Have made the changes. Kindly review.

[vfdev-5]

Thanks for the update! I added few minor comments. Let's add new feature tests and run the CI to see if any failures

[simeetnayan81]

Tests should be added to end of the test_metric.py file?

[vfdev-5]

Yes, you can add it in the end of the file

[simeetnayan81]

skip_unrolling = False is already covered by all the prior tests. I have added a test for when skip_unrolling = True. Kindly review and let me know the changes.

[vfdev-5]

Looks good @simeetnayan81 , thanks, let's just add an example of usage of the flag in the docstring and it will be good to go, once CI is green (except unrelated failures)

[simeetnayan81]

@vfdev-5 Before adding the example in the docstring, I wanted to confirm, to make skip_unrolling effective for the loss function, we might also need to change this.
https://github.com/pytorch/ignite/blob/master/ignite/metrics/loss.py#L77
Prev:

def __init__(
        self,
        loss_fn: Callable,
        output_transform: Callable = lambda x: x,
        batch_size: Callable = len,
        device: Union[str, torch.device] = torch.device("cpu"),
    ):
        super(Loss, self).__init__(output_transform, device=device)

Change to:

def __init__(
        self,
        loss_fn: Callable,
        output_transform: Callable = lambda x: x,
        batch_size: Callable = len,
        device: Union[str, torch.device] = torch.device("cpu"),
        skip_unrolling=False
    ):
        super(Loss, self).__init__(output_transform, device=device, skip_unrolling=skip_unrolling)

[vfdev-5]

@simeetnayan81 yes, you are right, we need to add this new arg to all metrics defining a constructor. Let's update Loss metric here and update other metrics in a follow-up PR.

[simeetnayan81]

Things to do in a follow-up PR.

• Add test for updated Loss class
• Update other sub-classes of Metric with skip_unrolling arg as required, add tests and docstring

[vfdev-5]

Thanks for the updates and the TODO. Can we do this point here ?

Add test for updated Loss class

[simeetnayan81]

Alright @vfdev-5

[simeetnayan81]

Have made the changes, the new test works locally.

[simeetnayan81]

The test is failing because list[torch.Tensor, torch.Tensor] is supported on python 3.9 and above. Let me modify it a bit.

[vfdev-5]

Thanks @simeetnayan81 , lgtm