Merge pull request #16 from amyeroberts/type-cast-before-normalize-vit

Type cast before normalize vit

src/transformers/models/vit/feature_extraction_vit.py

@@ -98,13 +98,15 @@ def __call__(
                 tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
                 number of channels, H and W are image height and width.
 
-            return_tensors (`str` or [`~utils.TensorType`], *optional*, defaults to `'np'`):
-                If set, will return tensors of a particular framework. Acceptable values are:
+            return_tensors (`str` or [`~utils.TensorType`], *optional*, defaults to `None`):
+                If set, will return a tensor of a particular framework.
 
-                - `'tf'`: Return TensorFlow `tf.constant` objects.
-                - `'pt'`: Return PyTorch `torch.Tensor` objects.
-                - `'np'`: Return NumPy `np.ndarray` objects.
-                - `'jax'`: Return JAX `jnp.ndarray` objects.
+                Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` object.
+                - `'pt'`: Return PyTorch `torch.Tensor` object.
+                - `'np'`: Return NumPy `np.ndarray` object.
+                - `'jax'`: Return JAX `jnp.ndarray` object.
+                - None: Return list of `np.ndarray` objects.
 
         Returns:
             [`BatchFeature`]: A [`BatchFeature`] with the following fields:
@@ -139,8 +141,15 @@ def __call__(
         # transformations (resizing + normalization)
         if self.do_resize and self.size is not None:
             images = [self.resize(image=image, size=self.size, resample=self.resample) for image in images]
+
+        # if do_normalize=False, the casting to a numpy array won't happen, so we need to do it here
+        make_channel_first = True if isinstance(images[0], Image.Image) else images[0].shape[-1] in (1, 3)
+        images = [self.to_numpy_array(image, rescale=False, channel_first=make_channel_first) for image in images]
+
         if self.do_normalize:
-            images = [self.normalize(image=image, mean=self.image_mean, std=self.image_std) for image in images]
+            images = [
+                self.normalize(image=image, mean=self.image_mean, std=self.image_std, rescale=True) for image in images
+            ]
 
         # return as BatchFeature
         data = {"pixel_values": images}

tests/models/vit/test_feature_extraction_vit.py

@@ -18,6 +18,7 @@
 
 import numpy as np
 
+from parameterized import parameterized
 from transformers.testing_utils import require_torch, require_vision
 from transformers.utils import is_torch_available, is_vision_available
 
@@ -189,3 +190,37 @@ def test_call_pytorch(self):
                 self.feature_extract_tester.size,
             ),
         )
+
+    @parameterized.expand(
+        [
+            ("do_resize_True_do_normalize_True", True, True),
+            ("do_resize_True_do_normalize_False", True, False),
+            ("do_resize_False_do_normalize_True", False, True),
+            ("do_resize_False_do_normalize_False", False, False),
+        ]
+    )
+    def test_call_flags(self, _, do_resize, do_normalize):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        feature_extractor.do_resize = do_resize
+        feature_extractor.do_normalize = do_normalize
+        # create random PIL images
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False)
+
+        if do_resize:
+            expected_shapes = [
+                (
+                    self.feature_extract_tester.num_channels,
+                    self.feature_extract_tester.size,
+                    self.feature_extract_tester.size,
+                )
+                for _ in range(self.feature_extract_tester.batch_size)
+            ]
+        else:
+            expected_shapes = [(3, *x.size[::-1]) for x in image_inputs]
+
+        pixel_values = feature_extractor(image_inputs, return_tensors=None)["pixel_values"]
+        self.assertEqual(len(pixel_values), self.feature_extract_tester.batch_size)
+        for idx, image in enumerate(pixel_values):
+            self.assertEqual(image.shape, expected_shapes[idx])
+            self.assertIsInstance(image, np.ndarray)