fixes metric hashing (#478)

* fixes metric hashing

* Update torchmetrics/metric.py

* Apply suggestions from code review

* chlog

* fix test

* fix decorator

* fix decorator

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: thomas chaton <thomas@grid.ai>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: Nicki Skafte <skaftenicki@gmail.com>
Co-authored-by: Jirka <jirka.borovec@seznam.cz>
(cherry picked from commit 3257c60)

CHANGELOG.md

@@ -33,6 +33,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Fixed bug where compositional metrics where unable to sync because of type mismatch ([#454](https://github.com/PyTorchLightning/metrics/pull/454))
 
+
+- Fixed metric hashing ([#478](https://github.com/PyTorchLightning/metrics/pull/478))
+
+
 ## [0.5.0] - 2021-08-09
 
 ### Added

tests/bases/test_hashing.py

@@ -0,0 +1,22 @@
+import pytest
+
+from tests.helpers.testers import DummyListMetric, DummyMetric
+
+
+@pytest.mark.parametrize(
+    "metric_cls",
+    [
+        DummyMetric,
+        DummyListMetric,
+    ],
+)
+def test_metric_hashing(metric_cls):
+    """Tests that hases are different.
+
+    See the Metric's hash function for details on why this is required.
+    """
+    instance_1 = metric_cls()
+    instance_2 = metric_cls()
+
+    assert hash(instance_1) != hash(instance_2)
+    assert id(instance_1) != id(instance_2)

tests/bases/test_metric.py

@@ -160,7 +160,7 @@ class B(DummyListMetric):
 
     b1 = B()
     b2 = B()
-    assert hash(b1) == hash(b2)
+    assert hash(b1) != hash(b2)  # different ids
     assert isinstance(b1.x, list) and len(b1.x) == 0
     b1.x.append(tensor(5))
     assert isinstance(hash(b1), int)  # <- check that nothing crashes

torchmetrics/metric.py

@@ -502,7 +502,12 @@ def _filter_kwargs(self, **kwargs: Any) -> Dict[str, Any]:
         return filtered_kwargs
 
     def __hash__(self) -> int:
-        hash_vals = [self.__class__.__name__]
+        # we need to add the id here, since PyTorch requires a module hash to be unique.
+        # Internally, PyTorch nn.Module relies on that for children discovery
+        # (see https://github.com/pytorch/pytorch/blob/v1.9.0/torch/nn/modules/module.py#L1544)
+        # For metrics that include tensors it is not a problem,
+        # since their hash is unique based on the memory location but we cannot rely on that for every metric.
+        hash_vals = [self.__class__.__name__, id(self)]
 
         for key in self._defaults:
             val = getattr(self, key)