Image Feature Extraction pipeline (#28216)

* Draft pipeline

* Fixup

* Fix docstrings

* Update doctest

* Update pipeline_model_mapping

* Update docstring

* Update tests

* Update src/transformers/pipelines/image_feature_extraction.py
Co-authored-by: Omar Sanseviero <osanseviero@gmail.com>

* Fix docstrings - review comments

* Remove pipeline mapping for composite vision models

* Add to pipeline tests

* Remove for flava (multimodal)

* safe pil import

* Add requirements for pipeline run

* Account for super slow efficientnet

* Review comments

* Fix tests

* Swap order of kwargs

* Use build_pipeline_init_args

* Add back FE pipeline for Vilt

* Include image_processor_kwargs in docstring

* Mark test as flaky

* Update TODO

* Update tests/pipelines/test_pipelines_image_feature_extraction.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Add license header

---------
Co-authored-by: Omar Sanseviero <osanseviero@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

tests/models/owlvit/test_modeling_owlvit.py

@@ -428,7 +428,10 @@ class OwlViTModelTester:
 class OwlViTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (OwlViTModel,) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": OwlViTModel, "zero-shot-object-detection": OwlViTForObjectDetection}
+        {
+            "feature-extraction": OwlViTModel,
+            "zero-shot-object-detection": OwlViTForObjectDetection,
+        }
         if is_torch_available()
         else {}
     )

tests/models/poolformer/test_modeling_poolformer.py

@@ -124,7 +124,7 @@ class PoolFormerModelTester:
 class PoolFormerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (PoolFormerModel, PoolFormerForImageClassification) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": PoolFormerModel, "image-classification": PoolFormerForImageClassification}
+        {"image-feature-extraction": PoolFormerModel, "image-classification": PoolFormerForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/pvt/test_modeling_pvt.py

@@ -158,7 +158,7 @@ def prepare_img():
 class PvtModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (PvtModel, PvtForImageClassification) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": PvtModel, "image-classification": PvtForImageClassification}
+        {"image-feature-extraction": PvtModel, "image-classification": PvtForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/regnet/test_modeling_regnet.py

@@ -126,7 +126,7 @@ class RegNetModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
 
     all_model_classes = (RegNetModel, RegNetForImageClassification) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": RegNetModel, "image-classification": RegNetForImageClassification}
+        {"image-feature-extraction": RegNetModel, "image-classification": RegNetForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/resnet/test_modeling_resnet.py

@@ -170,7 +170,7 @@ class ResNetModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": ResNetModel, "image-classification": ResNetForImageClassification}
+        {"image-feature-extraction": ResNetModel, "image-classification": ResNetForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/segformer/test_modeling_segformer.py

@@ -171,7 +171,7 @@ class SegformerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCas
     )
     pipeline_model_mapping = (
         {
-            "feature-extraction": SegformerModel,
+            "image-feature-extraction": SegformerModel,
             "image-classification": SegformerForImageClassification,
             "image-segmentation": SegformerForSemanticSegmentation,
         }

tests/models/swiftformer/test_modeling_swiftformer.py

@@ -139,7 +139,7 @@ class SwiftFormerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestC
 
     all_model_classes = (SwiftFormerModel, SwiftFormerForImageClassification) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": SwiftFormerModel, "image-classification": SwiftFormerForImageClassification}
+        {"image-feature-extraction": SwiftFormerModel, "image-classification": SwiftFormerForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/swin/test_modeling_swin.py

@@ -232,7 +232,7 @@ class SwinModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": SwinModel, "image-classification": SwinForImageClassification}
+        {"image-feature-extraction": SwinModel, "image-classification": SwinForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/swin2sr/test_modeling_swin2sr.py

@@ -162,7 +162,7 @@ class Swin2SRModelTester:
 class Swin2SRModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (Swin2SRModel, Swin2SRForImageSuperResolution) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": Swin2SRModel, "image-to-image": Swin2SRForImageSuperResolution}
+        {"image-feature-extraction": Swin2SRModel, "image-to-image": Swin2SRForImageSuperResolution}
         if is_torch_available()
         else {}
     )

tests/models/swinv2/test_modeling_swinv2.py

@@ -217,7 +217,7 @@ class Swinv2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": Swinv2Model, "image-classification": Swinv2ForImageClassification}
+        {"image-feature-extraction": Swinv2Model, "image-classification": Swinv2ForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/table_transformer/test_modeling_table_transformer.py

@@ -200,7 +200,7 @@ class TableTransformerModelTest(ModelTesterMixin, GenerationTesterMixin, Pipelin
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": TableTransformerModel, "object-detection": TableTransformerForObjectDetection}
+        {"image-feature-extraction": TableTransformerModel, "object-detection": TableTransformerForObjectDetection}
         if is_torch_available()
         else {}
     )

tests/models/vilt/test_modeling_vilt.py

@@ -228,7 +228,7 @@ class ViltModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": ViltModel, "visual-question-answering": ViltForQuestionAnswering}
+        {"image-feature-extraction": ViltModel, "visual-question-answering": ViltForQuestionAnswering}
         if is_torch_available()
         else {}
     )

tests/models/vit/test_modeling_vit.py

@@ -193,7 +193,7 @@ class ViTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": ViTModel, "image-classification": ViTForImageClassification}
+        {"image-feature-extraction": ViTModel, "image-classification": ViTForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/vit_hybrid/test_modeling_vit_hybrid.py

@@ -156,7 +156,7 @@ class ViTHybridModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCas
 
     all_model_classes = (ViTHybridModel, ViTHybridForImageClassification) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": ViTHybridModel, "image-classification": ViTHybridForImageClassification}
+        {"image-feature-extraction": ViTHybridModel, "image-classification": ViTHybridForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/vit_mae/test_modeling_vit_mae.py

@@ -164,7 +164,7 @@ class ViTMAEModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     """
 
     all_model_classes = (ViTMAEModel, ViTMAEForPreTraining) if is_torch_available() else ()
-    pipeline_model_mapping = {"feature-extraction": ViTMAEModel} if is_torch_available() else {}
+    pipeline_model_mapping = {"image-feature-extraction": ViTMAEModel} if is_torch_available() else {}
 
     test_pruning = False
     test_torchscript = False

tests/models/vit_msn/test_modeling_vit_msn.py

@@ -152,7 +152,7 @@ class ViTMSNModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
 
     all_model_classes = (ViTMSNModel, ViTMSNForImageClassification) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": ViTMSNModel, "image-classification": ViTMSNForImageClassification}
+        {"image-feature-extraction": ViTMSNModel, "image-classification": ViTMSNForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/yolos/test_modeling_yolos.py

@@ -168,7 +168,9 @@ class YolosModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
 
     all_model_classes = (YolosModel, YolosForObjectDetection) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": YolosModel, "object-detection": YolosForObjectDetection} if is_torch_available() else {}
+        {"image-feature-extraction": YolosModel, "object-detection": YolosForObjectDetection}
+        if is_torch_available()
+        else {}
     )
 
     test_pruning = False

tests/pipelines/test_pipelines_image_feature_extraction.py

+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 unittest
+
+import numpy as np
+import pytest
+
+from transformers import (
+    MODEL_MAPPING,
+    TF_MODEL_MAPPING,
+    TOKENIZER_MAPPING,
+    ImageFeatureExtractionPipeline,
+    is_tf_available,
+    is_torch_available,
+    is_vision_available,
+    pipeline,
+)
+from transformers.testing_utils import is_pipeline_test, nested_simplify, require_tf, require_torch
+
+
+if is_torch_available():
+    import torch
+
+if is_tf_available():
+    import tensorflow as tf
+
+if is_vision_available():
+    from PIL import Image
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+    return image
+
+
+@is_pipeline_test
+class ImageFeatureExtractionPipelineTests(unittest.TestCase):
+    model_mapping = MODEL_MAPPING
+    tf_model_mapping = TF_MODEL_MAPPING
+
+    @require_torch
+    def test_small_model_pt(self):
+        feature_extractor = pipeline(
+            task="image-feature-extraction", model="hf-internal-testing/tiny-random-vit", framework="pt"
+        )
+        img = prepare_img()
+        outputs = feature_extractor(img)
+        self.assertEqual(
+            nested_simplify(outputs[0][0]),
+            [-1.417, -0.392, -1.264, -1.196, 1.648, 0.885, 0.56, -0.606, -1.175, 0.823, 1.912, 0.081, -0.053, 1.119, -0.062, -1.757, -0.571, 0.075, 0.959, 0.118, 1.201, -0.672, -0.498, 0.364, 0.937, -1.623, 0.228, 0.19, 1.697, -1.115, 0.583, -0.981])  # fmt: skip
+
+    @require_tf
+    def test_small_model_tf(self):
+        feature_extractor = pipeline(
+            task="image-feature-extraction", model="hf-internal-testing/tiny-random-vit", framework="tf"
+        )
+        img = prepare_img()
+        outputs = feature_extractor(img)
+        self.assertEqual(
+            nested_simplify(outputs[0][0]),
+            [-1.417, -0.392, -1.264, -1.196, 1.648, 0.885, 0.56, -0.606, -1.175, 0.823, 1.912, 0.081, -0.053, 1.119, -0.062, -1.757, -0.571, 0.075, 0.959, 0.118, 1.201, -0.672, -0.498, 0.364, 0.937, -1.623, 0.228, 0.19, 1.697, -1.115, 0.583, -0.981])  # fmt: skip
+
+    @require_torch
+    def test_image_processing_small_model_pt(self):
+        feature_extractor = pipeline(
+            task="image-feature-extraction", model="hf-internal-testing/tiny-random-vit", framework="pt"
+        )
+
+        # test with image processor parameters
+        image_processor_kwargs = {"size": {"height": 300, "width": 300}}
+        img = prepare_img()
+        with pytest.raises(ValueError):
+            # Image doesn't match model input size
+            feature_extractor(img, image_processor_kwargs=image_processor_kwargs)
+
+        image_processor_kwargs = {"image_mean": [0, 0, 0], "image_std": [1, 1, 1]}
+        img = prepare_img()
+        outputs = feature_extractor(img, image_processor_kwargs=image_processor_kwargs)
+        self.assertEqual(np.squeeze(outputs).shape, (226, 32))
+
+    @require_tf
+    def test_image_processing_small_model_tf(self):
+        feature_extractor = pipeline(
+            task="image-feature-extraction", model="hf-internal-testing/tiny-random-vit", framework="tf"
+        )
+
+        # test with image processor parameters
+        image_processor_kwargs = {"size": {"height": 300, "width": 300}}
+        img = prepare_img()
+        with pytest.raises(ValueError):
+            # Image doesn't match model input size
+            feature_extractor(img, image_processor_kwargs=image_processor_kwargs)
+
+        image_processor_kwargs = {"image_mean": [0, 0, 0], "image_std": [1, 1, 1]}
+        img = prepare_img()
+        outputs = feature_extractor(img, image_processor_kwargs=image_processor_kwargs)
+        self.assertEqual(np.squeeze(outputs).shape, (226, 32))
+
+    @require_torch
+    def test_return_tensors_pt(self):
+        feature_extractor = pipeline(
+            task="image-feature-extraction", model="hf-internal-testing/tiny-random-vit", framework="pt"
+        )
+        img = prepare_img()
+        outputs = feature_extractor(img, return_tensors=True)
+        self.assertTrue(torch.is_tensor(outputs))
+
+    @require_tf
+    def test_return_tensors_tf(self):
+        feature_extractor = pipeline(
+            task="image-feature-extraction", model="hf-internal-testing/tiny-random-vit", framework="tf"
+        )
+        img = prepare_img()
+        outputs = feature_extractor(img, return_tensors=True)
+        self.assertTrue(tf.is_tensor(outputs))
+
+    def get_test_pipeline(self, model, tokenizer, processor):
+        if processor is None:
+            self.skipTest("No image processor")
+
+        elif type(model.config) in TOKENIZER_MAPPING:
+            self.skipTest("This is a bimodal model, we need to find a more consistent way to switch on those models.")
+
+        elif model.config.is_encoder_decoder:
+            self.skipTest(
+                """encoder_decoder models are trickier for this pipeline.
+                Do we want encoder + decoder inputs to get some featues?
+                Do we want encoder only features ?
+                For now ignore those.
+                """
+            )
+
+        feature_extractor = ImageFeatureExtractionPipeline(model=model, image_processor=processor)
+        img = prepare_img()
+        return feature_extractor, [img, img]
+
+    def run_pipeline_test(self, feature_extractor, examples):
+        imgs = examples
+        outputs = feature_extractor(imgs[0])
+
+        self.assertEqual(len(outputs), 1)
+
+        outputs = feature_extractor(imgs)
+        self.assertEqual(len(outputs), 2)

tests/test_pipeline_mixin.py

@@ -39,6 +39,7 @@ from .pipelines.test_pipelines_document_question_answering import DocumentQuesti
 from .pipelines.test_pipelines_feature_extraction import FeatureExtractionPipelineTests
 from .pipelines.test_pipelines_fill_mask import FillMaskPipelineTests
 from .pipelines.test_pipelines_image_classification import ImageClassificationPipelineTests
+from .pipelines.test_pipelines_image_feature_extraction import ImageFeatureExtractionPipelineTests
 from .pipelines.test_pipelines_image_segmentation import ImageSegmentationPipelineTests
 from .pipelines.test_pipelines_image_to_image import ImageToImagePipelineTests
 from .pipelines.test_pipelines_image_to_text import ImageToTextPipelineTests
@@ -70,6 +71,7 @@ pipeline_test_mapping = {
     "feature-extraction": {"test": FeatureExtractionPipelineTests},
     "fill-mask": {"test": FillMaskPipelineTests},
     "image-classification": {"test": ImageClassificationPipelineTests},
+    "image-feature-extraction": {"test": ImageFeatureExtractionPipelineTests},
     "image-segmentation": {"test": ImageSegmentationPipelineTests},
     "image-to-image": {"test": ImageToImagePipelineTests},
     "image-to-text": {"test": ImageToTextPipelineTests},
@@ -374,6 +376,13 @@ class PipelineTesterMixin:
     def test_pipeline_image_to_text(self):
         self.run_task_tests(task="image-to-text")
 
+    @is_pipeline_test
+    @require_timm
+    @require_vision
+    @require_torch
+    def test_pipeline_image_feature_extraction(self):
+        self.run_task_tests(task="image-feature-extraction")
+
     @unittest.skip(reason="`run_pipeline_test` is currently not implemented.")
     @is_pipeline_test
     @require_vision

utils/check_docstrings.py

@@ -324,6 +324,7 @@ OBJECTS_TO_IGNORE = [
     "IdeficsConfig",
     "IdeficsProcessor",
     "ImageClassificationPipeline",
+    "ImageFeatureExtractionPipeline",
     "ImageGPTConfig",
     "ImageSegmentationPipeline",
     "ImageToImagePipeline",