ddp precision recall

[fco-dv]

Fixes #1283

Description:

• Enable the case for Recall/Precision in multi_label, not averaged configuration for DDP

• Remove associated warnings, tests updated by scaling output size with idist.get_world_size()

• usage of metric_device should be checked in

  ignite/tests/ignite/metrics/test_precision.py

  Line 771 in 002b595

  def _test_distrib_integration_multilabel(device):

  
  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)

[sdesrozis]

@fco-dv well done !

Please could you update docstring ?

ignite/ignite/metrics/recall.py

Line 53 in 63a6f1b

In multilabel cases, if average is False, current implementation does not work with distributed computations.

ignite/ignite/metrics/precision.py

Line 112 in 63a6f1b

In multilabel cases, if average is False, current implementation does not work with distributed computations.

[vfdev-5]

Thanks for the PR @fco-dv ! Please, take a look at the comment on how to improve the test.

[vfdev-5]

First of all, could you please fix the issue we spotted during the call with unused metric_device that should go to Precision definition (line 795).

Ultimately, we would like to enable the tests in the same way as test_multilabel_input_NCHW where we run _test(average=True, ...) and _test(average=False, ...). In case of average False, there is no equivalent option in sklearn (as probably it does not make much sense neither). So, please, take a look how we test things in test_multilabel_input_NCHW and do same thing here. Then we can simply remove this dummy test below.

PS.: average=False and is_multilabel=True gives as a result for precision = multilabel precision per sample

import torch
from ignite.metrics import Precision

pr = Precision(average=False, is_multilabel=True)

pr.update((torch.tensor(y_pred[:5, :]), torch.tensor(y_true[:5, :])))
pr.update((torch.tensor(y_pred[5:, :]), torch.tensor(y_true[5:, :])))
print(pr._true_positives, pr._positives, pr.compute(), pr.compute().mean().item())
> (tensor([2., 2., 2., 2., 2., 2., 1., 0.], dtype=torch.float64),
 tensor([2., 3., 2., 2., 2., 2., 1., 1.], dtype=torch.float64),
 tensor([1.0000, 0.6667, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 0.0000],
        dtype=torch.float64),
 0.8333333333333333)

[vfdev-5]

@fco-dv could you please udpate the tests according the suggestion: #1646 (comment)

[vfdev-5]

LGTM! Thanks @fco-dv !