[`ImageProcessor`] Refactor default `mean` & `std` to `OPENAI_CLIP_MEAN` &...

[`ImageProcessor`] Refactor default `mean` & `std` to `OPENAI_CLIP_MEAN` & `OPENAI_CLIP_STD` (#21425)

* fix default value

* add the fix on other models

src/transformers/image_utils.py

@@ -35,6 +35,8 @@ from .utils.constants import (  # noqa: F401
     IMAGENET_DEFAULT_STD,
     IMAGENET_STANDARD_MEAN,
     IMAGENET_STANDARD_STD,
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
 )
 
 

src/transformers/models/bit/image_processing_bit.py

@@ -31,6 +31,8 @@ from ...image_transforms import (
     to_channel_dimension_format,
 )
 from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -118,8 +120,8 @@ class BitImageProcessor(BaseImageProcessor):
         self.do_rescale = do_rescale
         self.rescale_factor = rescale_factor
         self.do_normalize = do_normalize
-        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
-        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
         self.do_convert_rgb = do_convert_rgb
 
     def resize(

src/transformers/models/blip/image_processing_blip.py

@@ -24,8 +24,8 @@ from transformers.utils.generic import TensorType
 from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
 from ...image_transforms import convert_to_rgb, normalize, rescale, resize, to_channel_dimension_format
 from ...image_utils import (
-    IMAGENET_STANDARD_MEAN,
-    IMAGENET_STANDARD_STD,
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -103,8 +103,8 @@ class BlipImageProcessor(BaseImageProcessor):
         self.do_rescale = do_rescale
         self.rescale_factor = rescale_factor
         self.do_normalize = do_normalize
-        self.image_mean = image_mean if image_mean is not None else IMAGENET_STANDARD_MEAN
-        self.image_std = image_std if image_std is not None else IMAGENET_STANDARD_STD
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
         self.do_convert_rgb = do_convert_rgb
 
     def resize(

src/transformers/models/bridgetower/image_processing_bridgetower.py

@@ -25,6 +25,8 @@ from transformers.utils.generic import TensorType
 from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
 from ...image_transforms import PaddingMode, center_crop, normalize, pad, rescale, resize, to_channel_dimension_format
 from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -186,8 +188,8 @@ class BridgeTowerImageProcessor(BaseImageProcessor):
         self.do_rescale = do_rescale
         self.rescale_factor = rescale_factor
         self.do_normalize = do_normalize
-        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
-        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
         self.do_pad = do_pad
         self.do_center_crop = do_center_crop
 

src/transformers/models/chinese_clip/image_processing_chinese_clip.py

@@ -31,6 +31,8 @@ from ...image_transforms import (
     to_channel_dimension_format,
 )
 from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -118,8 +120,8 @@ class ChineseCLIPImageProcessor(BaseImageProcessor):
         self.do_rescale = do_rescale
         self.rescale_factor = rescale_factor
         self.do_normalize = do_normalize
-        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
-        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
         self.do_convert_rgb = do_convert_rgb
 
     def resize(

src/transformers/models/clip/image_processing_clip.py

@@ -31,6 +31,8 @@ from ...image_transforms import (
     to_channel_dimension_format,
 )
 from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -118,8 +120,8 @@ class CLIPImageProcessor(BaseImageProcessor):
         self.do_rescale = do_rescale
         self.rescale_factor = rescale_factor
         self.do_normalize = do_normalize
-        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
-        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
         self.do_convert_rgb = do_convert_rgb
 
     def resize(

src/transformers/models/flava/image_processing_flava.py

@@ -27,6 +27,8 @@ from transformers.utils.generic import TensorType
 from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
 from ...image_transforms import center_crop, normalize, rescale, resize, to_channel_dimension_format
 from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -45,8 +47,8 @@ logger = logging.get_logger(__name__)
 
 
 # These values are taken from CLIP
-FLAVA_IMAGE_MEAN = [0.48145466, 0.4578275, 0.40821073]
-FLAVA_IMAGE_STD = [0.26862954, 0.26130258, 0.27577711]
+FLAVA_IMAGE_MEAN = OPENAI_CLIP_MEAN
+FLAVA_IMAGE_STD = OPENAI_CLIP_STD
 FLAVA_CODEBOOK_MEAN = [0.0, 0.0, 0.0]
 FLAVA_CODEBOOK_STD = [1.0, 1.0, 1.0]
 LOGIT_LAPLACE_EPS: float = 0.1

src/transformers/models/owlvit/image_processing_owlvit.py

@@ -30,6 +30,8 @@ from transformers.image_transforms import (
     to_numpy_array,
 )
 from transformers.image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -160,8 +162,8 @@ class OwlViTImageProcessor(BaseImageProcessor):
         self.do_rescale = do_rescale
         self.rescale_factor = rescale_factor
         self.do_normalize = do_normalize
-        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
-        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
 
     def resize(
         self,

src/transformers/models/vit_hybrid/image_processing_vit_hybrid.py

@@ -31,6 +31,8 @@ from ...image_transforms import (
     to_channel_dimension_format,
 )
 from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
     ChannelDimension,
     ImageInput,
     PILImageResampling,
@@ -118,8 +120,8 @@ class ViTHybridImageProcessor(BaseImageProcessor):
         self.do_rescale = do_rescale
         self.rescale_factor = rescale_factor
         self.do_normalize = do_normalize
-        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
-        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
         self.do_convert_rgb = do_convert_rgb
 
     def resize(