Fall back to slow image processor in ImageProcessingAuto when no fast processor available (#34785)

* refactor image_processing_auto logic

* fix fast image processor tests

* Fix tests fast vit image processor

* Add safeguard when use_fast True and torchvision not available

* change default use_fast back to None, add warnings

* remove debugging print

* call get_image_processor_class_from_name once

tests/models/auto/test_image_processing_auto.py

@@ -140,6 +140,7 @@ class AutoImageProcessorTest(unittest.TestCase):
     def test_use_fast_selection(self):
         checkpoint = "hf-internal-testing/tiny-random-vit"
 
+        # TODO: @yoni, change in v4.48 (when use_fast set to True by default)
         # Slow image processor is selected by default
         image_processor = AutoImageProcessor.from_pretrained(checkpoint)
         self.assertIsInstance(image_processor, ViTImageProcessor)

tests/models/detr/test_image_processing_detr.py

@@ -19,7 +19,7 @@ import unittest
 
 import numpy as np
 
-from transformers.testing_utils import require_torch, require_torch_gpu, require_vision, slow
+from transformers.testing_utils import require_torch, require_torch_gpu, require_torchvision, require_vision, slow
 from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
 
 from ...test_image_processing_common import AnnotationFormatTestMixin, ImageProcessingTestMixin, prepare_image_inputs
@@ -669,6 +669,7 @@ class DetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcessingTestMixi
 
     @slow
     @require_torch_gpu
+    @require_torchvision
     def test_fast_processor_equivalence_cpu_gpu_coco_detection_annotations(self):
         # prepare image and target
         image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
@@ -724,6 +725,7 @@ class DetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcessingTestMixi
 
     @slow
     @require_torch_gpu
+    @require_torchvision
     def test_fast_processor_equivalence_cpu_gpu_coco_panoptic_annotations(self):
         # prepare image, target and masks_path
         image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")

tests/models/rt_detr/test_image_processing_rt_detr.py

@@ -16,7 +16,7 @@ import unittest
 
 import requests
 
-from transformers.testing_utils import require_torch, require_torch_gpu, require_vision, slow
+from transformers.testing_utils import require_torch, require_torch_gpu, require_torchvision, require_vision, slow
 from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
 
 from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
@@ -374,6 +374,7 @@ class RtDetrImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
 
     @slow
     @require_torch_gpu
+    @require_torchvision
     # Copied from tests.models.detr.test_image_processing_detr.DetrImageProcessingTest.test_fast_processor_equivalence_cpu_gpu_coco_detection_annotations
     def test_fast_processor_equivalence_cpu_gpu_coco_detection_annotations(self):
         # prepare image and target

tests/models/vision_text_dual_encoder/test_processor_vision_text_dual_encoder.py

@@ -21,13 +21,13 @@ import unittest
 from transformers import BertTokenizerFast
 from transformers.models.bert.tokenization_bert import VOCAB_FILES_NAMES, BertTokenizer
 from transformers.testing_utils import require_tokenizers, require_vision
-from transformers.utils import IMAGE_PROCESSOR_NAME, is_vision_available
+from transformers.utils import IMAGE_PROCESSOR_NAME, is_torchvision_available, is_vision_available
 
 from ...test_processing_common import ProcessorTesterMixin
 
 
 if is_vision_available():
-    from transformers import VisionTextDualEncoderProcessor, ViTImageProcessor
+    from transformers import VisionTextDualEncoderProcessor, ViTImageProcessor, ViTImageProcessorFast
 
 
 @require_tokenizers
@@ -63,6 +63,8 @@ class VisionTextDualEncoderProcessorTest(ProcessorTesterMixin, unittest.TestCase
         return BertTokenizer.from_pretrained(self.tmpdirname, **kwargs)
 
     def get_image_processor(self, **kwargs):
+        if is_torchvision_available():
+            return ViTImageProcessorFast.from_pretrained(self.tmpdirname, **kwargs)
         return ViTImageProcessor.from_pretrained(self.tmpdirname, **kwargs)
 
     def tearDown(self):
@@ -81,7 +83,7 @@ class VisionTextDualEncoderProcessorTest(ProcessorTesterMixin, unittest.TestCase
         self.assertIsInstance(processor.tokenizer, (BertTokenizer, BertTokenizerFast))
 
         self.assertEqual(processor.image_processor.to_json_string(), image_processor.to_json_string())
-        self.assertIsInstance(processor.image_processor, ViTImageProcessor)
+        self.assertIsInstance(processor.image_processor, (ViTImageProcessor, ViTImageProcessorFast))
 
     def test_save_load_pretrained_additional_features(self):
         processor = VisionTextDualEncoderProcessor(
@@ -100,7 +102,7 @@ class VisionTextDualEncoderProcessorTest(ProcessorTesterMixin, unittest.TestCase
         self.assertIsInstance(processor.tokenizer, (BertTokenizer, BertTokenizerFast))
 
         self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
-        self.assertIsInstance(processor.image_processor, ViTImageProcessor)
+        self.assertIsInstance(processor.image_processor, (ViTImageProcessor, ViTImageProcessorFast))
 
     def test_image_processor(self):
         image_processor = self.get_image_processor()
@@ -110,8 +112,8 @@ class VisionTextDualEncoderProcessorTest(ProcessorTesterMixin, unittest.TestCase
 
         image_input = self.prepare_image_inputs()
 
-        input_feat_extract = image_processor(image_input, return_tensors="np")
-        input_processor = processor(images=image_input, return_tensors="np")
+        input_feat_extract = image_processor(image_input, return_tensors="pt")
+        input_processor = processor(images=image_input, return_tensors="pt")
 
         for key in input_feat_extract.keys():
             self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)