make sure metriccollection is deterministic

CHANGELOG.md

@@ -41,6 +41,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Changed behaviour of `confusionmatrix` for multilabel data to better match `multilabel_confusion_matrix` from sklearn ([#134](https://github.com/PyTorchLightning/metrics/pull/134))
 - Updated FBeta arguments ([#111](https://github.com/PyTorchLightning/metrics/pull/111))
 - Changed `reset` method to use `detach.clone()` instead of `deepcopy` when resetting to default ([#163](https://github.com/PyTorchLightning/metrics/pull/163))
+- Metrics passed as dict to `MetricCollection` will now always be in deterministic order ([#173](https://github.com/PyTorchLightning/metrics/pull/173))
 
 ### Deprecated
 

tests/bases/test_collections.py

@@ -156,3 +156,12 @@ def test_metric_collection_prefix_arg(tmpdir):
     out = new_metric_collection(5)
     for name in names:
         assert f"new_prefix_{name}" in out, 'prefix argument not working as intended with clone method'
+
+
+def test_metric_collection_same_order():
+    m1 = DummyMetricSum()
+    m2 = DummyMetricDiff()
+    col1 = MetricCollection({"a": m1, "b": m2})
+    col2 = MetricCollection({"b": m2, "a": m1})
+    for k1, k2 in zip(col1.keys(), col2.keys()):
+        assert k1 == k2

torchmetrics/collections.py

@@ -35,6 +35,7 @@ class MetricCollection(nn.ModuleDict):
             * dict: if metrics are passed in as a dict, will use each key in the
               dict as key for output dict. Use this format if you want to chain
               together multiple of the same metric with different parameters.
+              Note that the keys in the output dict will be sorted alphabetically.
 
         prefix: a string to append in front of the keys of the output dict
 
@@ -78,7 +79,9 @@ def __init__(
         super().__init__()
         if isinstance(metrics, dict):
             # Check all values are metrics
-            for name, metric in metrics.items():
+            # Make sure that metrics are added in deterministic order
+            for name in sorted(metrics.keys()):
+                metric = metrics[name]
                 if not isinstance(metric, Metric):
                     raise ValueError(
                         f"Value {metric} belonging to key {name}"