Label smooth threshold fix

thinc/tests/test_util.py

@@ -115,6 +115,14 @@ def test_to_categorical(label_smoothing):
     ):
         to_categorical(numpy.asarray([0, 0, 0]), label_smoothing=0.01),
 
+    error_msg = ("For 5 number of classes "
+                 "label_smoothing parameter has to be less than "
+                 "0.8, but found 0.8.")
+    with pytest.raises(
+        ValueError, match=error_msg
+    ):
+        to_categorical(numpy.asarray([0, 1, 2, 3, 4]), label_smoothing=0.8)
+
 
 def test_convert_recursive():
     is_match = lambda obj: obj == "foo"

thinc/util.py

@@ -212,16 +212,15 @@ def to_categorical(
     *,
     label_smoothing: float = 0.0,
 ) -> FloatsXd:
-    if not 0.0 <= label_smoothing < 0.5:
-        raise ValueError(
-            "label_smoothing should be greater or "
-            "equal to 0.0 and less than 0.5, "
-            f"but {label_smoothing} was provided."
-        )
 
     if n_classes is None:
         n_classes = int(numpy.max(Y) + 1)  # type: ignore
 
+    if label_smoothing < 0.0:
+        raise ValueError(
+            "Label-smoothing parameter has to be greater than or equal to 0"
+        )
+
     if label_smoothing == 0.0:
         if n_classes == 0:
             raise ValueError("n_classes should be at least 1")
@@ -234,6 +233,13 @@ def to_categorical(
             )
         nongold_prob = label_smoothing / (n_classes - 1)
 
+    if (1 - label_smoothing) < nongold_prob:
+        raise ValueError(
+            f"For {n_classes} number of classes "
+            "label_smoothing parameter has to be less than "
+            f"{1 - nongold_prob}, but found {label_smoothing}."
+        )
+
     xp = get_array_module(Y)
     label_distr = xp.full((n_classes, n_classes), nongold_prob, dtype="float32")
     xp.fill_diagonal(label_distr, 1 - label_smoothing)