🚨🚨🚨 Fix rescale ViVit Efficientnet (#25174)

* Fix rescaling bug

* Add tests

* Update integration tests

* Fix up

* Update src/transformers/image_transforms.py

* Update test - new possible order in list

src/transformers/image_transforms.py

@@ -110,10 +110,11 @@ def rescale(
     if not isinstance(image, np.ndarray):
         raise ValueError(f"Input image must be of type np.ndarray, got {type(image)}")
 
+    image = image.astype(dtype)
+
     rescaled_image = image * scale
     if data_format is not None:
         rescaled_image = to_channel_dimension_format(rescaled_image, data_format)
-    rescaled_image = rescaled_image.astype(dtype)
     return rescaled_image
 
 

src/transformers/models/efficientnet/image_processing_efficientnet.py

@@ -153,7 +153,13 @@ class EfficientNetImageProcessor(BaseImageProcessor):
         **kwargs,
     ):
         """
-        Rescale an image by a scale factor. image = image * scale.
+        Rescale an image by a scale factor.
+
+        If offset is True, the image is rescaled between [-1, 1].
+            image = image * scale * 2 - 1
+
+        If offset is False, the image is rescaled between [0, 1].
+            image = image * scale
 
         Args:
             image (`np.ndarray`):
@@ -165,13 +171,12 @@ class EfficientNetImageProcessor(BaseImageProcessor):
             data_format (`str` or `ChannelDimension`, *optional*):
                 The channel dimension format of the image. If not provided, it will be the same as the input image.
         """
-        if offset:
-            rescaled_image = (image - 127.5) * scale
-            if data_format is not None:
-                rescaled_image = to_channel_dimension_format(rescaled_image, data_format)
-            rescaled_image = rescaled_image.astype(np.float32)
-        else:
+        scale = scale * 2 if offset else scale
         rescaled_image = rescale(image, scale=scale, data_format=data_format, **kwargs)
+
+        if offset:
+            rescaled_image = rescaled_image - 1
+
         return rescaled_image
 
     def preprocess(

src/transformers/models/vivit/image_processing_vivit.py

@@ -167,6 +167,7 @@ class VivitImageProcessor(BaseImageProcessor):
             raise ValueError(f"Size must have 'height' and 'width' or 'shortest_edge' as keys. Got {size.keys()}")
         return resize(image, size=output_size, resample=resample, data_format=data_format, **kwargs)
 
+    # Copied from transformers.models.efficientnet.image_processing_efficientnet.EfficientNetImageProcessor.rescale
     def rescale(
         self,
         image: np.ndarray,
@@ -178,8 +179,11 @@ class VivitImageProcessor(BaseImageProcessor):
         """
         Rescale an image by a scale factor.
 
-        If offset is `True`, image scaled between [-1, 1]: image = (image - 127.5) * scale. If offset is `False`, image
-        scaled between [0, 1]: image = image * scale
+        If offset is True, the image is rescaled between [-1, 1].
+            image = image * scale * 2 - 1
+
+        If offset is False, the image is rescaled between [0, 1].
+            image = image * scale
 
         Args:
             image (`np.ndarray`):
@@ -191,10 +195,13 @@ class VivitImageProcessor(BaseImageProcessor):
             data_format (`str` or `ChannelDimension`, *optional*):
                 The channel dimension format of the image. If not provided, it will be the same as the input image.
         """
-        image = image.astype(np.float32)
+        scale = scale * 2 if offset else scale
+        rescaled_image = rescale(image, scale=scale, data_format=data_format, **kwargs)
+
         if offset:
-            image = image - (scale / 2)
-        return rescale(image, scale=scale, data_format=data_format, **kwargs)
+            rescaled_image = rescaled_image - 1
+
+        return rescaled_image
 
     def _preprocess_image(
         self,

tests/models/efficientnet/test_image_processing_efficientnet.py

@@ -193,3 +193,17 @@ class EfficientNetImageProcessorTest(ImageProcessingSavingTestMixin, unittest.Te
                 self.image_processor_tester.size["width"],
             ),
         )
+
+    def test_rescale(self):
+        # EfficientNet optionally rescales between -1 and 1 instead of the usual 0 and 1
+        image = np.arange(0, 256, 1, dtype=np.uint8).reshape(1, 8, 32)
+
+        image_processor = self.image_processing_class(**self.image_processor_dict)
+
+        rescaled_image = image_processor.rescale(image, scale=1 / 255)
+        expected_image = image.astype(np.float32) * (2 / 255.0) - 1
+        self.assertTrue(np.allclose(rescaled_image, expected_image))
+
+        rescaled_image = image_processor.rescale(image, scale=1 / 255, offset=False)
+        expected_image = image.astype(np.float32) / 255.0
+        self.assertTrue(np.allclose(rescaled_image, expected_image))

tests/models/vivit/test_image_processing_vivit.py

@@ -212,3 +212,17 @@ class VivitImageProcessingTest(ImageProcessingSavingTestMixin, unittest.TestCase
                 self.image_processor_tester.crop_size["width"],
             ),
         )
+
+    def test_rescale(self):
+        # ViVit optionally rescales between -1 and 1 instead of the usual 0 and 1
+        image = np.arange(0, 256, 1, dtype=np.uint8).reshape(1, 8, 32)
+
+        image_processor = self.image_processing_class(**self.image_processor_dict)
+
+        rescaled_image = image_processor.rescale(image, scale=1 / 255)
+        expected_image = image.astype(np.float32) * (2 / 255.0) - 1
+        self.assertTrue(np.allclose(rescaled_image, expected_image))
+
+        rescaled_image = image_processor.rescale(image, scale=1 / 255, offset=False)
+        expected_image = image.astype(np.float32) / 255.0
+        self.assertTrue(np.allclose(rescaled_image, expected_image))

tests/models/vivit/test_modeling_vivit.py

@@ -345,6 +345,6 @@ class VivitModelIntegrationTest(unittest.TestCase):
         self.assertEqual(outputs.logits.shape, expected_shape)
 
         # taken from original model
-        expected_slice = torch.tensor([-1.0543, 2.0764, -0.2104, 0.4439, -0.9658]).to(torch_device)
+        expected_slice = torch.tensor([-0.9498, 2.7971, -1.4049, 0.1024, -1.8353]).to(torch_device)
 
         self.assertTrue(torch.allclose(outputs.logits[0, :5], expected_slice, atol=1e-4))

tests/pipelines/test_pipelines_zero_shot_image_classification.py

@@ -85,6 +85,7 @@ class ZeroShotImageClassificationPipelineTests(unittest.TestCase):
             [
                 [{"score": 0.333, "label": "a"}, {"score": 0.333, "label": "b"}, {"score": 0.333, "label": "c"}],
                 [{"score": 0.333, "label": "a"}, {"score": 0.333, "label": "c"}, {"score": 0.333, "label": "b"}],
+                [{"score": 0.333, "label": "b"}, {"score": 0.333, "label": "a"}, {"score": 0.333, "label": "c"}],
             ],
         )