metrics: test reset + improve coverage

Author: mdekstrand

src/lenskit/metrics/_collect.py

@@ -159,9 +159,6 @@ def measure_collection(
                 key_kwargs = keys | dict(zip(outputs.key_fields, key))
                 list_test = test.lookup_projected(key)
 
-                if out is None:
-                    out = ItemList()
-
                 if list_test is None:
                     no_test_count += 1
                     list_test = ItemList([])
@@ -228,12 +225,7 @@ def _wrap_metric(
     assert isinstance(m, Metric), f"invalid type for metric {m}"
 
     if label is None:
-        if isinstance(m, Metric):
-            wl = m.label
-        elif isinstance(m, type):
-            wl = m.__name__  # type: ignore
-        else:
-            wl = type(m).__name__
+        wl = m.label
     else:
         wl = label
 

tests/eval/test_measurement_collector.py

@@ -14,10 +14,9 @@
 
 from lenskit.basic import PopScorer
 from lenskit.data import ItemList, ItemListCollection
-from lenskit.metrics import NDCG, Recall
+from lenskit.metrics import NDCG, ListLength, Recall, RecipRank
 from lenskit.metrics._base import FunctionMetric, ListMetric, Metric
 from lenskit.metrics._collect import MeasurementCollector, _wrap_metric
-from lenskit.metrics.basic import ListLength
 from lenskit.splitting import split_temporal_fraction
 
 _log = logging.getLogger(__name__)
@@ -247,3 +246,30 @@ def measure_list(self, recs, test):
 
     summary = acc.summary_metrics()
     assert summary["test.mean"] == 2
+
+
+def test_reset():
+    acc = MeasurementCollector()
+    acc.add_metric(ListLength())
+    acc.add_metric(RecipRank())
+
+    acc.measure_list(ItemList([1, 2, 3, 4, 5]), ItemList([4]))
+    acc.measure_list(ItemList([5, 4, 3, 2, 1]), ItemList([1]))
+
+    lms = acc.list_metrics()
+    assert len(lms) == 2
+    assert np.all(lms["N"] == 5)
+    sms = acc.summary_metrics()
+    assert sms["N.mean"] == approx(5.0)
+    assert sms["RecipRank.mean"] == approx(0.225)
+
+    acc.reset()
+    acc.measure_list(ItemList([1, 2, 3, 4, 5, 10]), ItemList([2]))
+    acc.measure_list(ItemList([5, 4, 3, 2, 1, 10]), ItemList([2]))
+
+    lms = acc.list_metrics()
+    assert len(lms) == 2
+    assert np.all(lms["N"] == 6)
+    sms = acc.summary_metrics()
+    assert sms["N.mean"] == approx(6.0)
+    assert sms["RecipRank.mean"] == approx(0.625)