Fix: Allow hashing of metrics with lists in their state

pytorch_lightning/metrics/metric.py

@@ -332,8 +332,23 @@ def _filter_kwargs(self, **kwargs):
     def __hash__(self):
         hash_vals = [self.__class__.__name__]
 
+        # Torch tensors are hashable, but based on their
+        # underlying pointer. Since we do a structural
+        # equality check in __eq__, we circumvent this
+        # by hashing the _sum_ of tensors' contents
+        def format_tensor(t):
+            if not isinstance(t, torch.Tensor):
+                return t
+            return float(t.detach().cpu().sum())
+
         for key in self._defaults.keys():
-            hash_vals.append(getattr(self, key))
+            val = getattr(self, key)
+            # Special case: allow list values, so long
+            # as their elements are hashable
+            if hasattr(val, '__iter__') and not isinstance(val, torch.Tensor):
+                hash_vals.extend((format_tensor(x) for x in val))
+            else:
+                hash_vals.append(format_tensor(val))
 
         return hash(tuple(hash_vals))
 

tests/metrics/test_metric.py

@@ -154,6 +154,27 @@ def compute(self):
     assert a.compute() == 5
 
 
+def test_hash():
+
+    class A(Dummy):
+        pass
+
+    class B(DummyList):
+        pass
+
+    a1 = A()
+    a2 = A()
+    assert hash(a1) == hash(a2)
+
+    b1 = B()
+    b2 = B()
+    assert hash(b1) == hash(b2)
+    assert isinstance(b1.x, list) and len(b1.x) == 0
+    b1.x.append(torch.tensor(5))
+    assert isinstance(hash(b1), int)  # <- check that nothing crashes
+    assert isinstance(b1.x, list) and len(b1.x) == 1
+
+
 def test_forward():
 
     class A(Dummy):