Fix feature handling in Radon and NaN checks in LAION400M.

tensorflow_datasets/datasets/radon/radon_dataset_builder.py

@@ -133,9 +133,15 @@ def _generate_examples(self, file_path_srrs2, file_path_cty):
 
     for i, (_, row) in enumerate(df.iterrows()):
       radon_val = row.pop('activity')
+      features_dict = {}
+      for name, (_, convert_fn) in features().items():
+        try:
+          features_dict[name] = convert_fn(row[name])
+        except ValueError as e:
+          raise ValueError(
+              f'Failed to convert {name} with value {row[name]!r}'
+          ) from e
       yield i, {
           'activity': float(radon_val),
-          'features': {
-              name: features()[name][1](value) for name, value in row.items()
-          },
+          'features': features_dict,
       }

tensorflow_datasets/vision_language/laion400m/laion400m.py

@@ -14,12 +14,12 @@
 # limitations under the License.
 
 """LAION-400M image dataset."""
+
 import functools
 from typing import Dict, Tuple
 
 from etils import epath
 import numpy as np
-
 from tensorflow_datasets.core.utils.lazy_imports_utils import tensorflow as tf
 import tensorflow_datasets.public_api as tfds
 
@@ -78,15 +78,21 @@
 
 def _get_example_metadata(metadata_df_row):
   """Returns example metadata."""
+  pd = tfds.core.lazy_imports.pandas
   nsfw_tag = metadata_df_row['NSFW']
   if nsfw_tag not in _NSFW_TAGS:
     nsfw_tag = _NSFW_MISSING_TAG
 
+  similarity = metadata_df_row['similarity']
+  license_ = metadata_df_row['LICENSE']
+
   return {
       'caption': metadata_df_row['caption'],
       'nsfw': nsfw_tag,
-      'similarity': metadata_df_row['similarity'] or _MISSING_SIMILARITY_VALUE,
-      'license': metadata_df_row['LICENSE'] or '',
+      'similarity': (
+          _MISSING_SIMILARITY_VALUE if pd.isna(similarity) else similarity
+      ),
+      'license': '' if pd.isna(license_) else license_,
       'url': metadata_df_row['url'],
       'original_width': metadata_df_row['original_width'],
       'original_height': metadata_df_row['original_height'],
@@ -174,11 +180,9 @@ def _info(self) -> tfds.core.DatasetInfo:
     }
 
     if self.builder_config.name == LAION400M_IMAGES_CONFIG.name:
-      features.update(
-          {
-              'image': tfds.features.Image(doc='image'),
-          }
-      )
+      features.update({
+          'image': tfds.features.Image(doc='image'),
+      })
     else:
       features.update({
           'image_embedding': tfds.features.Tensor(