Add conversion options from non-TFDS to TFDS valid features.

PiperOrigin-RevId: 650913004

Author: The TensorFlow Datasets Authors

tensorflow_datasets/core/utils/huggingface_utils.py

@@ -238,6 +238,12 @@ def convert_hf_value(
       # Ensure RGB format for PNG encoding.
       return hf_value.convert('RGB')
     case feature_lib.Tensor():
+      if isinstance(hf_value, float):
+        # In some cases, for example when loading jsonline files using pandas,
+        # empty non-float values, such as strings, are converted to float nan.
+        # We spot those occurrences as the feature.np_dtype is not float.
+        if np.isnan(hf_value) and not dtype_utils.is_floating(feature.np_dtype):
+          return _get_default_value(feature)
       return hf_value
 
   raise TypeError(

tensorflow_datasets/core/utils/huggingface_utils_test.py

@@ -211,6 +211,12 @@ def test_convert_value_raises(hf_value, feature):
             ),
             {'foo': b''},
         ),
+        # nan, but the feature type is not float
+        (
+            np.nan,
+            feature_lib.Text(),
+            b'',
+        ),
     ],
 )
 def test_convert_value(hf_value, feature, expected_value):