fixes metric hashing

tests/bases/test_hashing.py

@@ -0,0 +1,23 @@
+import pytest
+
+from tests.helpers.testers import DummyMetric, DummyNonTensorMetric
+
+pytest.mark.parametrize(
+    "metric_cls",
+    [
+        (DummyMetric,),
+        (DummyNonTensorMetric,),
+    ],
+)
+
+
+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/helpers/testers.py

@@ -511,6 +511,20 @@ def compute(self):
         pass
 
 
+class DummyNonTensorMetric(Metric):
+    name = "DummyNonTensor"
+
+    def __init__(self):
+        super().__init__()
+        self.add_state("x", [], dist_reduce_fx=None)
+
+    def update(self):
+        pass
+
+    def compute(self):
+        pass
+
+
 class DummyListMetric(Metric):
     name = "DummyList"
 

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 the hash of an module to be unique.
+        # they rely 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)