save

lilac/data/dataset_duckdb.py

@@ -1873,6 +1873,7 @@ def select_groups(
     if not leaf.categorical and (leaf_is_float or leaf_is_integer):
       if named_bins is None:
         # Auto-bin.
+        print(stats)
         named_bins = _auto_bins(stats, NUM_AUTO_BINS)
 
       sql_bounds = []
@@ -3915,6 +3916,9 @@ def _normalize_bins(bins: Optional[Union[Sequence[Bin], Sequence[float]]]) -> Op
 
 
 def _auto_bins(stats: StatsResult, num_bins: int) -> list[Bin]:
+  if stats.min_val is None or stats.max_val is None:
+    return [('0', None, None)]
+
   min_val = cast(float, stats.min_val)
   max_val = cast(float, stats.max_val)
   value_range = max_val - min_val

lilac/data/dataset_select_groups_test.py

@@ -263,6 +263,18 @@ def test_auto_bins_for_float(make_test_data: TestDataMaker) -> None:
   assert res.bins
 
 
+def test_auto_bins_for_missing_float(make_test_data: TestDataMaker) -> None:
+  items: list[Item] = [{'feature': 1.0}] + [{'feature': float('nan')}] * 5
+  dataset = make_test_data(items)
+  # The 1.0 row was just to get the right type inference going; desired dataset is a bunch of NaNs.
+  dataset.delete_rows(filters=[('feature', 'equals', 1.0)])
+
+  res = dataset.select_groups('feature')
+  assert res.counts == [(None, 5)]
+  assert res.too_many_distinct is False
+  assert res.bins == [('0', None, None)]
+
+
 def test_map_dtype(make_test_data: TestDataMaker) -> None:
   items = [
     {'column': {'a': 1.0, 'b': 2.0}},