[llava] one pixel is missing from padding when length is odd (#37819)

* [fix] one pixel should be added when length is odd

* [fix] add vision_aspect_ratio args & typo

* [fix] style

* [fix] do not fix fast file directly

* [fix] convert using modular

* remove duplicate codes

* match unpad logic with pad logic

* test odd-sized images for llava & aria

* test unpad odd-sized padding for llava family

* fix style

* add kwarg to onvision modular

* move vision_aspect_ratio from image_processor to processor
(llava_onevision)

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 @@ class AriaImageProcessor(BaseImageProcessor):
         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 @@ class AriaImageProcessor(BaseImageProcessor):
         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

@@ -12,7 +12,6 @@
 # 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
@@ -20,7 +19,7 @@ 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 @@ class AriaProjector(nn.Module):
         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 @@ class AriaImageProcessor(BaseImageProcessor):
         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 @@ class AriaImageProcessor(BaseImageProcessor):
         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