Internal change.

PiperOrigin-RevId: 561693602

tensorflow_datasets/core/dataset_builders/huggingface_dataset_builder.py

@@ -191,11 +191,17 @@ def _convert_value(hf_value: Any, feature: feature_lib.FeatureConnector) -> Any:
       }
   elif isinstance(feature, feature_lib.FeaturesDict):
     if isinstance(hf_value, dict):
-      return hf_value
+      return {k: _convert_value(v, feature[k]) for k, v in hf_value.items()}
     raise ValueError(f"The feature is {feature}, but the value is: {hf_value}")
   elif isinstance(feature, feature_lib.Sequence):
-    if isinstance(hf_value, (dict, list)):
-      return hf_value
+    if isinstance(hf_value, dict):
+      # Should be a dict of lists:
+      return {
+          k: [_convert_value(el, feature.feature[k]) for el in v]
+          for k, v in hf_value.items()
+      }
+    if isinstance(hf_value, list):
+      return [_convert_value(v, feature.feature) for v in hf_value]
     else:
       return [hf_value]
   elif isinstance(feature, feature_lib.Audio):
@@ -214,6 +220,8 @@ def _convert_value(hf_value: Any, feature: feature_lib.FeatureConnector) -> Any:
     buffer = io.BytesIO()
     hf_value.save(fp=buffer, format=_IMAGE_ENCODING_FORMAT)
     return buffer.getvalue()
+  elif isinstance(feature, feature_lib.Tensor):
+    return hf_value
   raise ValueError(
       f"Type {type(hf_value)} of value {hf_value} "
       f"for feature {type(feature)} is not supported."

tensorflow_datasets/core/dataset_builders/huggingface_dataset_builder_test.py

@@ -149,6 +149,21 @@ def test_convert_value_sequence():
   )
 
 
+def test_convert_value_empty_sequence():
+  assert huggingface_dataset_builder._convert_value(
+      [None, "string"], feature_lib.Sequence(feature=np.str_)
+  ) == [b"", "string"]
+
+
+def test_convert_value_sequence_of_dict():
+  sequence_feature = feature_lib.Sequence(
+      {"someint": feature_lib.Scalar(dtype=np.str_)}
+  )
+  assert huggingface_dataset_builder._convert_value(
+      {"someint": [None, "string", None]}, sequence_feature
+  ) == {"someint": [b"", "string", b""]}
+
+
 def test_convert_value_image():
   image_feature = feature_lib.Image()
   image = lazy_imports_lib.lazy_imports.PIL_Image.new(mode="RGB", size=(4, 4))