Make VideoMAEImageProcessor much faster

src/transformers/feature_extraction_utils.py

@@ -135,8 +135,13 @@ def _get_is_as_tensor_fns(self, tensor_type: Optional[Union[str, TensorType]] =
                 raise ImportError("Unable to convert output to PyTorch tensors format, PyTorch is not installed.")
             import torch  # noqa
 
+            def recursive_ndarray_check(value):
+                if isinstance(value, (list, tuple)) and len(value) > 0:
+                    return recursive_ndarray_check(value[0])
+                return isinstance(value, np.ndarray)
+
             def as_tensor(value):
-                if isinstance(value, (list, tuple)) and len(value) > 0 and isinstance(value[0], np.ndarray):
+                if recursive_ndarray_check(value):
                     value = np.array(value)
                 return torch.tensor(value)
 

src/transformers/models/videomae/image_processing_videomae.py

@@ -339,5 +339,7 @@ def preprocess(
             for video in videos
         ]
 
-        data = {"pixel_values": videos}
+        # Speeds up tensor conversion - see: https://github.com/huggingface/transformers/pull/28221/files
+        data = {"pixel_values": np.asarray(videos) if return_tensors == "pt" else videos}
+
         return BatchFeature(data=data, tensor_type=return_tensors)