[llava] one pixel is missing from padding when length is odd

setup.py

@@ -466,7 +466,12 @@ def run(self):
     package_data={"": ["**/*.cu", "**/*.cpp", "**/*.cuh", "**/*.h", "**/*.pyx", "py.typed"]},
     zip_safe=False,
     extras_require=extras,
-    entry_points={"console_scripts": ["transformers=transformers.commands.transformers_cli:main", "transformers-cli=transformers.commands.transformers_cli:main_cli"]},
+    entry_points={
+        "console_scripts": [
+            "transformers=transformers.commands.transformers_cli:main",
+            "transformers-cli=transformers.commands.transformers_cli:main_cli",
+        ]
+    },
     python_requires=">=3.9.0",
     install_requires=list(install_requires),
     classifiers=[

src/transformers/models/aria/image_processing_aria.py

@@ -18,12 +18,11 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import math
 from typing import Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 
-from ...image_processing_utils import BaseImageProcessor, BatchFeature, select_best_resolution
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_patch_output_size, select_best_resolution
 from ...image_transforms import PaddingMode, convert_to_rgb, pad, resize, to_channel_dimension_format
 from ...image_utils import (
     ChannelDimension,
@@ -71,23 +70,6 @@ def divide_to_patches(image: np.array, patch_size: int, input_data_format) -> Li
     return patches
 
 
-def _get_patch_output_size(image, target_resolution, input_data_format):
-    original_height, original_width = get_image_size(image, channel_dim=input_data_format)
-    target_height, target_width = target_resolution
-
-    scale_w = target_width / original_width
-    scale_h = target_height / original_height
-
-    if scale_w < scale_h:
-        new_width = target_width
-        new_height = min(math.ceil(original_height * scale_w), target_height)
-    else:
-        new_height = target_height
-        new_width = min(math.ceil(original_width * scale_h), target_width)
-
-    return new_height, new_width
-
-
 class AriaImageProcessor(BaseImageProcessor):
     """
     A vision processor for the Aria model that handles image preprocessing.
@@ -375,7 +357,7 @@ def _resize_for_patching(
         Returns:
             np.array: The resized and padded image.
         """
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
         # Resize the image
         resized_image = resize(image, (new_height, new_width), resample=resample, input_data_format=input_data_format)
@@ -389,12 +371,12 @@ def _pad_for_patching(
         Pad an image to a target resolution while maintaining aspect ratio.
         """
         target_height, target_width = target_resolution
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
-        paste_x = (target_width - new_width) // 2
-        paste_y = (target_height - new_height) // 2
+        paste_x, r_x = divmod(target_width - new_width, 2)
+        paste_y, r_y = divmod(target_height - new_height, 2)
 
-        padded_image = self.pad(image, padding=((paste_y, paste_y), (paste_x, paste_x)))
+        padded_image = self.pad(image, padding=((paste_y, paste_y + r_y), (paste_x, paste_x + r_x)))
 
         return padded_image
 

src/transformers/models/aria/modular_aria.py

@@ -12,15 +12,14 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import math
 from typing import Dict, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 
 from ...activations import ACT2FN
 from ...configuration_utils import PretrainedConfig
 from ...generation import GenerationMixin
-from ...image_processing_utils import BaseImageProcessor, BatchFeature, select_best_resolution
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_patch_output_size, select_best_resolution
 from ...image_transforms import PaddingMode, convert_to_rgb, pad, resize, to_channel_dimension_format
 from ...image_utils import (
     ChannelDimension,
@@ -461,23 +460,6 @@ def forward(self, key_value_states: torch.Tensor, attn_mask: Optional[torch.Tens
         return out
 
 
-def _get_patch_output_size(image, target_resolution, input_data_format):
-    original_height, original_width = get_image_size(image, channel_dim=input_data_format)
-    target_height, target_width = target_resolution
-
-    scale_w = target_width / original_width
-    scale_h = target_height / original_height
-
-    if scale_w < scale_h:
-        new_width = target_width
-        new_height = min(math.ceil(original_height * scale_w), target_height)
-    else:
-        new_height = target_height
-        new_width = min(math.ceil(original_width * scale_h), target_width)
-
-    return new_height, new_width
-
-
 class AriaImageProcessor(BaseImageProcessor):
     """
     A vision processor for the Aria model that handles image preprocessing.
@@ -765,7 +747,7 @@ def _resize_for_patching(
         Returns:
             np.array: The resized and padded image.
         """
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
         # Resize the image
         resized_image = resize(image, (new_height, new_width), resample=resample, input_data_format=input_data_format)
@@ -779,12 +761,12 @@ def _pad_for_patching(
         Pad an image to a target resolution while maintaining aspect ratio.
         """
         target_height, target_width = target_resolution
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
-        paste_x = (target_width - new_width) // 2
-        paste_y = (target_height - new_height) // 2
+        paste_x, r_x = divmod(target_width - new_width, 2)
+        paste_y, r_y = divmod(target_height - new_height, 2)
 
-        padded_image = self.pad(image, padding=((paste_y, paste_y), (paste_x, paste_x)))
+        padded_image = self.pad(image, padding=((paste_y, paste_y + r_y), (paste_x, paste_x + r_x)))
 
         return padded_image
 

src/transformers/models/llava_next/image_processing_llava_next.py

@@ -14,12 +14,17 @@
 # limitations under the License.
 """Image processor class for LLaVa-NeXT."""
 
-import math
 from typing import Dict, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 
-from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict, select_best_resolution
+from ...image_processing_utils import (
+    BaseImageProcessor,
+    BatchFeature,
+    get_patch_output_size,
+    get_size_dict,
+    select_best_resolution,
+)
 from ...image_transforms import (
     PaddingMode,
     convert_to_rgb,
@@ -99,23 +104,6 @@ def expand_to_square(image: np.array, background_color, input_data_format) -> np
         return result
 
 
-def _get_patch_output_size(image, target_resolution, input_data_format):
-    original_height, original_width = get_image_size(image, channel_dim=input_data_format)
-    target_height, target_width = target_resolution
-
-    scale_w = target_width / original_width
-    scale_h = target_height / original_height
-
-    if scale_w < scale_h:
-        new_width = target_width
-        new_height = min(math.ceil(original_height * scale_w), target_height)
-    else:
-        new_height = target_height
-        new_width = min(math.ceil(original_width * scale_h), target_width)
-
-    return new_height, new_width
-
-
 class LlavaNextImageProcessor(BaseImageProcessor):
     r"""
     Constructs a LLaVa-NeXT image processor. Based on [`CLIPImageProcessor`] with incorporation of additional techniques
@@ -429,7 +417,7 @@ def _resize_for_patching(
         Returns:
             np.array: The resized and padded image.
         """
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
         # Resize the image
         resized_image = resize(image, (new_height, new_width), resample=resample, input_data_format=input_data_format)
@@ -443,12 +431,12 @@ def _pad_for_patching(
         Pad an image to a target resolution while maintaining aspect ratio.
         """
         target_height, target_width = target_resolution
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
-        paste_x = (target_width - new_width) // 2
-        paste_y = (target_height - new_height) // 2
+        paste_x, r_x = divmod(target_width - new_width, 2)
+        paste_y, r_y = divmod(target_height - new_height, 2)
 
-        padded_image = self.pad(image, padding=((paste_y, paste_y), (paste_x, paste_x)))
+        padded_image = self.pad(image, padding=((paste_y, paste_y + r_y), (paste_x, paste_x + r_x)))
 
         return padded_image
 

src/transformers/models/llava_next/image_processing_llava_next_fast.py

@@ -102,8 +102,8 @@ def __init__(self, **kwargs: Unpack[LlavaNextFastImageProcessorKwargs]):
                 A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list
                 of the form `(height, width)`.
             do_pad (`bool`, *optional*):
-                    Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
-                    number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+                Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+                number of patches in the batch. Padding will be applied to the bottom and right with zeros.
         """,
     )
     def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaNextFastImageProcessorKwargs]) -> BatchFeature:
@@ -164,10 +164,10 @@ def _pad_for_patching(
         target_height, target_width = target_resolution
         new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
-        paste_x = (target_width - new_width) // 2
-        paste_y = (target_height - new_height) // 2
+        paste_x, r_x = divmod(target_width - new_width, 2)
+        paste_y, r_y = divmod(target_height - new_height, 2)
 
-        padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x, paste_y])
+        padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x + r_x, paste_y + r_y])
 
         return padded_image
 

src/transformers/models/llava_next/modeling_llava_next.py

@@ -139,14 +139,14 @@ def unpad_image(tensor, original_size):
 
     if original_aspect_ratio > current_aspect_ratio:
         scale_factor = current_width / original_width
-        new_height = int(round(original_height * scale_factor, 7))
-        padding = (current_height - new_height) // 2
-        unpadded_tensor = tensor[:, padding : current_height - padding, :]
+        new_height = min(math.ceil(original_height * scale_factor), current_height)
+        padding, r = divmod(current_height - new_height, 2)
+        unpadded_tensor = tensor[:, padding : current_height - (padding + r), :]
     else:
         scale_factor = current_height / original_height
-        new_width = int(round(original_width * scale_factor, 7))
-        padding = (current_width - new_width) // 2
-        unpadded_tensor = tensor[:, :, padding : current_width - padding]
+        new_width = min(math.ceil(original_width * scale_factor), current_width)
+        padding, r = divmod(current_width - new_width, 2)
+        unpadded_tensor = tensor[:, :, padding : current_width - (padding + r)]
 
     return unpadded_tensor
 

src/transformers/models/llava_next_video/modeling_llava_next_video.py

@@ -262,14 +262,14 @@ def unpad_image(tensor, original_size):
 
     if original_aspect_ratio > current_aspect_ratio:
         scale_factor = current_width / original_width
-        new_height = int(round(original_height * scale_factor, 7))
-        padding = (current_height - new_height) // 2
-        unpadded_tensor = tensor[:, padding : current_height - padding, :]
+        new_height = min(math.ceil(original_height * scale_factor), current_height)
+        padding, r = divmod(current_height - new_height, 2)
+        unpadded_tensor = tensor[:, padding : current_height - (padding + r), :]
     else:
         scale_factor = current_height / original_height
-        new_width = int(round(original_width * scale_factor, 7))
-        padding = (current_width - new_width) // 2
-        unpadded_tensor = tensor[:, :, padding : current_width - padding]
+        new_width = min(math.ceil(original_width * scale_factor), current_width)
+        padding, r = divmod(current_width - new_width, 2)
+        unpadded_tensor = tensor[:, :, padding : current_width - (padding + r)]
 
     return unpadded_tensor
 

src/transformers/models/llava_onevision/image_processing_llava_onevision.py

@@ -14,12 +14,17 @@
 # limitations under the License.
 """Image processor class for LLaVa-Onevision."""
 
-import math
 from typing import Dict, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 
-from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict, select_best_resolution
+from ...image_processing_utils import (
+    BaseImageProcessor,
+    BatchFeature,
+    get_patch_output_size,
+    get_size_dict,
+    select_best_resolution,
+)
 from ...image_transforms import (
     PaddingMode,
     convert_to_rgb,
@@ -99,24 +104,6 @@ def expand_to_square(image: np.array, background_color, input_data_format) -> np
         return result
 
 
-# Copied from transformers.models.llava_next.image_processing_llava_next._get_patch_output_size
-def _get_patch_output_size(image, target_resolution, input_data_format):
-    original_height, original_width = get_image_size(image, channel_dim=input_data_format)
-    target_height, target_width = target_resolution
-
-    scale_w = target_width / original_width
-    scale_h = target_height / original_height
-
-    if scale_w < scale_h:
-        new_width = target_width
-        new_height = min(math.ceil(original_height * scale_w), target_height)
-    else:
-        new_height = target_height
-        new_width = min(math.ceil(original_width * scale_h), target_width)
-
-    return new_height, new_width
-
-
 class LlavaOnevisionImageProcessor(BaseImageProcessor):
     r"""
     Constructs a LLaVa-Onevision image processor. Based on [`SiglipImageProcessor`] with incorporation of processing each video frame.
@@ -151,8 +138,8 @@ class LlavaOnevisionImageProcessor(BaseImageProcessor):
             number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
             Can be overridden by the `image_std` parameter in the `preprocess` method.
         do_pad (`bool`, *optional*, defaults to `True`):
-                Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
-                number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+            Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+            number of patches in the batch. Padding will be applied to the bottom and right with zeros.
         do_convert_rgb (`bool`, *optional*, defaults to `True`):
             Whether to convert the image to RGB.
     """
@@ -321,7 +308,7 @@ def _resize_for_patching(
         Returns:
             np.array: The resized and padded image.
         """
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
         # Resize the image
         resized_image = resize(image, (new_height, new_width), resample=resample, input_data_format=input_data_format)
@@ -336,12 +323,12 @@ def _pad_for_patching(
         Pad an image to a target resolution while maintaining aspect ratio.
         """
         target_height, target_width = target_resolution
-        new_height, new_width = _get_patch_output_size(image, target_resolution, input_data_format)
+        new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
-        paste_x = (target_width - new_width) // 2
-        paste_y = (target_height - new_height) // 2
+        paste_x, r_x = divmod(target_width - new_width, 2)
+        paste_y, r_y = divmod(target_height - new_height, 2)
 
-        padded_image = self.pad(image, padding=((paste_y, paste_y), (paste_x, paste_x)))
+        padded_image = self.pad(image, padding=((paste_y, paste_y + r_y), (paste_x, paste_x + r_x)))
 
         return padded_image
 

src/transformers/models/llava_onevision/image_processing_llava_onevision_fast.py

@@ -84,8 +84,8 @@ def __init__(self, **kwargs: Unpack[LlavaOnevisionFastImageProcessorKwargs]):
                 A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list
                 of the form `(height, width)`.
             do_pad (`bool`, *optional*):
-                    Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
-                    number of patches in the batch. Padding will be applied to the bottom and right with zeros.
+                Whether to pad the image. If `True`, will pad the patch dimension of the images in the batch to the largest
+                number of patches in the batch. Padding will be applied to the bottom and right with zeros.
         """,
     )
     def preprocess(self, images: ImageInput, **kwargs: Unpack[LlavaOnevisionFastImageProcessorKwargs]) -> BatchFeature:
@@ -146,10 +146,10 @@ def _pad_for_patching(
         target_height, target_width = target_resolution
         new_height, new_width = get_patch_output_size(image, target_resolution, input_data_format)
 
-        paste_x = (target_width - new_width) // 2
-        paste_y = (target_height - new_height) // 2
+        paste_x, r_x = divmod(target_width - new_width, 2)
+        paste_y, r_y = divmod(target_height - new_height, 2)
 
-        padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x, paste_y])
+        padded_image = F.pad(image, padding=[paste_x, paste_y, paste_x + r_x, paste_y + r_y])
 
         return padded_image