Improve BatchFeature: stack list and lists of torch tensors

src/transformers/feature_extraction_utils.py

@@ -67,11 +67,18 @@ class BatchFeature(UserDict):
         tensor_type (`Union[None, str, TensorType]`, *optional*):
             You can give a tensor_type here to convert the lists of integers in PyTorch/Numpy Tensors at
             initialization.
+        skip_tensor_conversion (`list[str]` or `set[str]`, *optional*):
+            List or set of keys that should NOT be converted to tensors, even when `tensor_type` is specified.
     """
 
-    def __init__(self, data: Optional[dict[str, Any]] = None, tensor_type: Union[None, str, TensorType] = None):
+    def __init__(
+        self,
+        data: Optional[dict[str, Any]] = None,
+        tensor_type: Union[None, str, TensorType] = None,
+        skip_tensor_conversion: Optional[Union[list[str], set[str]]] = None,
+    ):
         super().__init__(data)
-        self.convert_to_tensors(tensor_type=tensor_type)
+        self.convert_to_tensors(tensor_type=tensor_type, skip_tensor_conversion=skip_tensor_conversion)
 
     def __getitem__(self, item: str) -> Any:
         """
@@ -110,6 +117,14 @@ def _get_is_as_tensor_fns(self, tensor_type: Optional[Union[str, TensorType]] =
             import torch
 
             def as_tensor(value):
+                if torch.is_tensor(value):
+                    return value
+
+                # stack list of tensors if tensor_type is PyTorch (# torch.tensor() does not support list of tensors)
+                if isinstance(value, (list, tuple)) and len(value) > 0 and torch.is_tensor(value[0]):
+                    return torch.stack(value)
+
+                # convert list of numpy arrays to numpy array (stack) if tensor_type is Numpy
                 if isinstance(value, (list, tuple)) and len(value) > 0:
                     if isinstance(value[0], np.ndarray):
                         value = np.array(value)
@@ -138,14 +153,20 @@ def as_tensor(value, dtype=None):
             is_tensor = is_numpy_array
         return is_tensor, as_tensor
 
-    def convert_to_tensors(self, tensor_type: Optional[Union[str, TensorType]] = None):
+    def convert_to_tensors(
+        self,
+        tensor_type: Optional[Union[str, TensorType]] = None,
+        skip_tensor_conversion: Optional[Union[list[str], set[str]]] = None,
+    ):
         """
         Convert the inner content to tensors.
 
         Args:
             tensor_type (`str` or [`~utils.TensorType`], *optional*):
                 The type of tensors to use. If `str`, should be one of the values of the enum [`~utils.TensorType`]. If
                 `None`, no modification is done.
+            skip_tensor_conversion (`list[str]` or `set[str]`, *optional*):
+                List or set of keys that should NOT be converted to tensors, even when `tensor_type` is specified.
         """
         if tensor_type is None:
             return self
@@ -154,18 +175,26 @@ def convert_to_tensors(self, tensor_type: Optional[Union[str, TensorType]] = Non
 
         # Do the tensor conversion in batch
         for key, value in self.items():
+            # Skip keys explicitly marked for no conversion
+            if skip_tensor_conversion and key in skip_tensor_conversion:
+                continue
+
             try:
                 if not is_tensor(value):
                     tensor = as_tensor(value)
-
                     self[key] = tensor
-            except:  # noqa E722
+            except Exception as e:
                 if key == "overflowing_values":
-                    raise ValueError("Unable to create tensor returning overflowing values of different lengths. ")
+                    raise ValueError(
+                        f"Unable to create tensor for '{key}' with overflowing values of different lengths. "
+                        f"Original error: {str(e)}"
+                    ) from e
                 raise ValueError(
-                    "Unable to create tensor, you should probably activate padding "
-                    "with 'padding=True' to have batched tensors with the same length."
-                )
+                    f"Unable to convert output '{key}' (type: {type(value).__name__}) to tensor: {str(e)}\n"
+                    f"You can try:\n"
+                    f"  1. Use padding=True to ensure all outputs have the same shape\n"
+                    f"  2. Set return_tensors=None to return Python objects instead of tensors"
+                ) from e
 
         return self
 

src/transformers/image_processing_utils_fast.py

@@ -932,7 +932,6 @@ def _preprocess(
         if do_pad:
             processed_images = self.pad(processed_images, pad_size=pad_size, disable_grouping=disable_grouping)
 
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 
     def to_dict(self):

src/transformers/models/beit/image_processing_beit_fast.py

@@ -163,7 +163,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/bridgetower/image_processing_bridgetower_fast.py

@@ -251,10 +251,8 @@ def _preprocess(
             processed_images, processed_masks = self.pad(
                 processed_images, return_mask=True, disable_grouping=disable_grouping
             )
-            processed_masks = torch.stack(processed_masks, dim=0) if return_tensors else processed_masks
             data["pixel_mask"] = processed_masks
 
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
         data["pixel_values"] = processed_images
 
         return BatchFeature(data=data, tensor_type=return_tensors)

src/transformers/models/cohere2_vision/image_processing_cohere2_vision_fast.py

@@ -262,7 +262,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(
             data={"pixel_values": processed_images, "num_patches": num_patches}, tensor_type=return_tensors

src/transformers/models/convnext/image_processing_convnext_fast.py

@@ -162,7 +162,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/deepseek_vl/image_processing_deepseek_vl_fast.py

@@ -171,7 +171,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/deepseek_vl_hybrid/image_processing_deepseek_vl_hybrid_fast.py

@@ -207,9 +207,6 @@ def _preprocess(
             )
             high_res_processed_images_grouped[shape] = stacked_high_res_images
         high_res_processed_images = reorder_images(high_res_processed_images_grouped, grouped_high_res_images_index)
-        high_res_processed_images = (
-            torch.stack(high_res_processed_images, dim=0) if return_tensors else high_res_processed_images
-        )
 
         resized_images_grouped = {}
         for shape, stacked_high_res_padded_images in high_res_padded_images.items():
@@ -233,7 +230,6 @@ def _preprocess(
             )
             processed_images_grouped[shape] = stacked_images
         processed_images = reorder_images(processed_images_grouped, grouped_resized_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(
             data={"pixel_values": processed_images, "high_res_pixel_values": high_res_processed_images},

src/transformers/models/deepseek_vl_hybrid/modular_deepseek_vl_hybrid.py

@@ -888,9 +888,6 @@ def _preprocess(
             )
             high_res_processed_images_grouped[shape] = stacked_high_res_images
         high_res_processed_images = reorder_images(high_res_processed_images_grouped, grouped_high_res_images_index)
-        high_res_processed_images = (
-            torch.stack(high_res_processed_images, dim=0) if return_tensors else high_res_processed_images
-        )
 
         resized_images_grouped = {}
         for shape, stacked_high_res_padded_images in high_res_padded_images.items():
@@ -914,7 +911,6 @@ def _preprocess(
             )
             processed_images_grouped[shape] = stacked_images
         processed_images = reorder_images(processed_images_grouped, grouped_resized_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(
             data={"pixel_values": processed_images, "high_res_pixel_values": high_res_processed_images},

src/transformers/models/depth_pro/image_processing_depth_pro_fast.py

@@ -94,7 +94,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/dinov3_vit/image_processing_dinov3_vit_fast.py

@@ -88,7 +88,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/donut/image_processing_donut_fast.py

@@ -231,7 +231,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/dpt/image_processing_dpt_fast.py

@@ -225,8 +225,7 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
-        return BatchFeature(data={"pixel_values": processed_images})
+        return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 
     def post_process_semantic_segmentation(self, outputs, target_sizes: Optional[list[tuple]] = None):
         """

src/transformers/models/dpt/modular_dpt.py

@@ -228,8 +228,7 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
-        return BatchFeature(data={"pixel_values": processed_images})
+        return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 
     def post_process_depth_estimation(
         self,

src/transformers/models/efficientloftr/image_processing_efficientloftr_fast.py

@@ -153,9 +153,8 @@ def _preprocess(
         stacked_pairs = [torch.stack(pair, dim=0) for pair in image_pairs]
 
         # Return in same format as slow processor
-        image_pairs = torch.stack(stacked_pairs, dim=0) if return_tensors else stacked_pairs
 
-        return BatchFeature(data={"pixel_values": image_pairs})
+        return BatchFeature(data={"pixel_values": stacked_pairs}, tensor_type=return_tensors)
 
     def post_process_keypoint_matching(
         self,

src/transformers/models/efficientnet/image_processing_efficientnet_fast.py

@@ -178,7 +178,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/eomt/image_processing_eomt_fast.py

@@ -162,8 +162,7 @@ def _preprocess_image_like_inputs(
         )
         ignore_index = kwargs.pop("ignore_index", None)
         images_kwargs = kwargs.copy()
-        processed_images, patch_offsets = self._preprocess(images, **images_kwargs)
-        outputs = BatchFeature({"pixel_values": processed_images})
+        outputs = self._preprocess(images, **images_kwargs)
 
         if segmentation_maps is not None:
             processed_segmentation_maps = self._prepare_image_like_inputs(
@@ -183,9 +182,9 @@ def _preprocess_image_like_inputs(
                 }
             )
 
-            processed_segmentation_maps, _ = self._preprocess(
+            processed_segmentation_maps = self._preprocess(
                 images=processed_segmentation_maps, **segmentation_maps_kwargs
-            )
+            ).pixel_values
             processed_segmentation_maps = processed_segmentation_maps.squeeze(1).to(torch.int64)
             # Convert to list of binary masks and labels
             mask_labels, class_labels = [], []
@@ -208,8 +207,8 @@ def _preprocess_image_like_inputs(
             outputs["mask_labels"] = mask_labels
             outputs["class_labels"] = class_labels
 
-        if patch_offsets:
-            outputs["patch_offsets"] = [torch.tensor(offsets) for offsets in patch_offsets]
+        if outputs.patch_offsets:
+            outputs["patch_offsets"] = [torch.tensor(offsets) for offsets in outputs.patch_offsets]
 
         return outputs
 
@@ -274,11 +273,13 @@ def _preprocess(
                 stacked_images, do_rescale, rescale_factor, do_normalize, image_mean, image_std
             )
             processed_images_grouped[shape] = stacked_images
-        images = reorder_images(processed_images_grouped, grouped_images_index)
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
 
-        processed_images = torch.stack(images, dim=0) if return_tensors else images
-
-        return processed_images, patch_offsets
+        return BatchFeature(
+            data={"pixel_values": processed_images, "patch_offsets": patch_offsets},
+            tensor_type=return_tensors,
+            skip_tensor_conversion=["patch_offsets"],
+        )
 
     def merge_image_patches(
         self,

src/transformers/models/flava/image_processing_flava_fast.py

@@ -306,7 +306,6 @@ def _preprocess_image(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return processed_images
 
@@ -397,7 +396,6 @@ def _preprocess(
                 mask_group_max_aspect_ratio=mask_group_max_aspect_ratio,
             )
             masks = [mask_generator() for _ in range(len(images))]
-            masks = torch.stack(masks, dim=0) if return_tensors else masks
             data["bool_masked_pos"] = masks
 
         return BatchFeature(data=data, tensor_type=return_tensors)

src/transformers/models/fuyu/image_processing_fuyu.py

@@ -94,7 +94,7 @@ class FuyuBatchFeature(BatchFeature):
     The outputs dictionary from the processors contains a mix of tensors and lists of tensors.
     """
 
-    def convert_to_tensors(self, tensor_type: Optional[Union[str, TensorType]] = None):
+    def convert_to_tensors(self, tensor_type: Optional[Union[str, TensorType]] = None, **kwargs):
         """
         Convert the inner content to tensors.
 

src/transformers/models/gemma3/image_processing_gemma3_fast.py

@@ -231,7 +231,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
         return BatchFeature(
             data={"pixel_values": processed_images, "num_crops": num_crops}, tensor_type=return_tensors
         )

src/transformers/models/glpn/image_processing_glpn_fast.py

@@ -107,7 +107,6 @@ def _preprocess(
             processed_groups[shape] = stacked_images
 
         processed_images = reorder_images(processed_groups, grouped_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 
     def post_process_depth_estimation(self, outputs, target_sizes=None):

src/transformers/models/got_ocr2/image_processing_got_ocr2_fast.py

@@ -189,7 +189,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(
             data={"pixel_values": processed_images, "num_patches": num_patches}, tensor_type=return_tensors

src/transformers/models/imagegpt/image_processing_imagegpt_fast.py

@@ -164,12 +164,8 @@ def _preprocess(
 
             input_ids = reorder_images(input_ids_grouped, grouped_images_index)
 
-            return BatchFeature(
-                data={"input_ids": torch.stack(input_ids, dim=0) if return_tensors else input_ids},
-                tensor_type=return_tensors,
-            )
+            return BatchFeature(data={"input_ids": input_ids}, tensor_type=return_tensors)
 
-        pixel_values = torch.stack(pixel_values, dim=0) if return_tensors else pixel_values
         return BatchFeature(data={"pixel_values": pixel_values}, tensor_type=return_tensors)
 
     def to_dict(self):

src/transformers/models/instructblipvideo/video_processing_instructblipvideo.py

@@ -84,7 +84,6 @@ def _preprocess(
             processed_videos_grouped[shape] = stacked_videos
 
         processed_videos = reorder_videos(processed_videos_grouped, grouped_videos_index)
-        processed_videos = torch.stack(processed_videos, dim=0) if return_tensors else processed_videos
 
         return BatchFeature(data={"pixel_values": processed_videos}, tensor_type=return_tensors)
 

src/transformers/models/internvl/video_processing_internvl.py

@@ -140,7 +140,6 @@ def _preprocess(
             processed_videos_grouped[shape] = stacked_videos
 
         processed_videos = reorder_videos(processed_videos_grouped, grouped_videos_index)
-        processed_videos = torch.stack(processed_videos, dim=0) if return_tensors else processed_videos
 
         return BatchFeature(data={"pixel_values_videos": processed_videos}, tensor_type=return_tensors)
 

src/transformers/models/janus/image_processing_janus_fast.py

@@ -180,7 +180,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
 

src/transformers/models/kosmos2_5/image_processing_kosmos2_5_fast.py

@@ -264,8 +264,8 @@ def _preprocess(
 
         encoded_outputs = BatchFeature(
             data={
-                "flattened_patches": torch.stack(flattened_patches, dim=0) if return_tensors else flattened_patches,
-                "attention_mask": torch.stack(attention_masks, dim=0) if return_tensors else attention_masks,
+                "flattened_patches": flattened_patches,
+                "attention_mask": attention_masks,
                 "width": width,
                 "height": height,
                 "rows": rows,

src/transformers/models/layoutlmv2/image_processing_layoutlmv2_fast.py

@@ -101,7 +101,6 @@ def _preprocess(
             processed_images_grouped[shape] = stacked_images
 
         processed_images = reorder_images(processed_images_grouped, grouped_images_index)
-        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
 
         data = BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)