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>

docs/source/en/main_classes/pipelines.md

@@ -469,6 +469,12 @@ Pipelines available for multimodal tasks include the following.
     - __call__
     - all
 
+### ImageFeatureExtractionPipeline
+
+[[autodoc]] ImageFeatureExtractionPipeline
+    - __call__
+    - all
+
 ### ImageToTextPipeline
 
 [[autodoc]] ImageToTextPipeline

docs/source/ja/main_classes/pipelines.md

@@ -25,7 +25,7 @@ Recognition、Masked Language Modeling、Sentiment Analysis、Feature Extraction
 パイプラインの抽象化には2つのカテゴリーがある：
 
 - [`pipeline`] は、他のすべてのパイプラインをカプセル化する最も強力なオブジェクトです。
-- タスク固有のパイプラインは、[オーディオ](#audio)、[コンピューター ビジョン](#computer-vision)、[自然言語処理](#natural-language-processing)、および [マルチモーダル](#multimodal) タスクで使用できます。 
+- タスク固有のパイプラインは、[オーディオ](#audio)、[コンピューター ビジョン](#computer-vision)、[自然言語処理](#natural-language-processing)、および [マルチモーダル](#multimodal) タスクで使用できます。
 
 ## The pipeline abstraction
 
@@ -477,6 +477,12 @@ my_pipeline = pipeline(model="xxxx", pipeline_class=MyPipeline)
     - __call__
     - all
 
+### ImageFeatureExtractionPipeline
+
+[[autodoc]] ImageFeatureExtractionPipeline
+    - __call__
+    - all
+
 ### ImageToTextPipeline
 
 [[autodoc]] ImageToTextPipeline

docs/source/zh/main_classes/pipelines.md

@@ -435,7 +435,7 @@ See [`TokenClassificationPipeline`] for all details.
     - __call__
     - all
 
-## 多模态 
+## 多模态
 
 可用于多模态任务的pipeline包括以下几种。
 
@@ -451,6 +451,12 @@ See [`TokenClassificationPipeline`] for all details.
     - __call__
     - all
 
+### ImageFeatureExtractionPipeline
+
+[[autodoc]] ImageFeatureExtractionPipeline
+    - __call__
+    - all
+
 ### ImageToTextPipeline
 
 [[autodoc]] ImageToTextPipeline

src/transformers/__init__.py

@@ -973,6 +973,7 @@ _import_structure = {
         "FeatureExtractionPipeline",
         "FillMaskPipeline",
         "ImageClassificationPipeline",
+        "ImageFeatureExtractionPipeline",
         "ImageSegmentationPipeline",
         "ImageToImagePipeline",
         "ImageToTextPipeline",
@@ -5709,6 +5710,7 @@ if TYPE_CHECKING:
         FeatureExtractionPipeline,
         FillMaskPipeline,
         ImageClassificationPipeline,
+        ImageFeatureExtractionPipeline,
         ImageSegmentationPipeline,
         ImageToImagePipeline,
         ImageToTextPipeline,

src/transformers/pipelines/__init__.py

@@ -66,6 +66,7 @@ from .document_question_answering import DocumentQuestionAnsweringPipeline
 from .feature_extraction import FeatureExtractionPipeline
 from .fill_mask import FillMaskPipeline
 from .image_classification import ImageClassificationPipeline
+from .image_feature_extraction import ImageFeatureExtractionPipeline
 from .image_segmentation import ImageSegmentationPipeline
 from .image_to_image import ImageToImagePipeline
 from .image_to_text import ImageToTextPipeline
@@ -362,6 +363,18 @@ SUPPORTED_TASKS = {
         },
         "type": "image",
     },
+    "image-feature-extraction": {
+        "impl": ImageFeatureExtractionPipeline,
+        "tf": (TFAutoModel,) if is_tf_available() else (),
+        "pt": (AutoModel,) if is_torch_available() else (),
+        "default": {
+            "model": {
+                "pt": ("google/vit-base-patch16-224", "29e7a1e183"),
+                "tf": ("google/vit-base-patch16-224", "29e7a1e183"),
+            }
+        },
+        "type": "image",
+    },
     "image-segmentation": {
         "impl": ImageSegmentationPipeline,
         "tf": (),
@@ -500,6 +513,7 @@ def check_task(task: str) -> Tuple[str, Dict, Any]:
             - `"feature-extraction"`
             - `"fill-mask"`
             - `"image-classification"`
+            - `"image-feature-extraction"`
             - `"image-segmentation"`
             - `"image-to-text"`
             - `"image-to-image"`
@@ -586,6 +600,7 @@ def pipeline(
             - `"feature-extraction"`: will return a [`FeatureExtractionPipeline`].
             - `"fill-mask"`: will return a [`FillMaskPipeline`]:.
             - `"image-classification"`: will return a [`ImageClassificationPipeline`].
+            - `"image-feature-extraction"`: will return an [`ImageFeatureExtractionPipeline`].
             - `"image-segmentation"`: will return a [`ImageSegmentationPipeline`].
             - `"image-to-image"`: will return a [`ImageToImagePipeline`].
             - `"image-to-text"`: will return a [`ImageToTextPipeline`].

src/transformers/pipelines/feature_extraction.py

@@ -14,7 +14,7 @@ from .base import GenericTensor, Pipeline, build_pipeline_init_args
 )
 class FeatureExtractionPipeline(Pipeline):
     """
-    Feature extraction pipeline using no model head. This pipeline extracts the hidden states from the base
+    Feature extraction pipeline uses no model head. This pipeline extracts the hidden states from the base
     transformer, which can be used as features in downstream tasks.
 
     Example:

src/transformers/pipelines/image_feature_extraction.py

+from typing import Dict
+
+from ..utils import add_end_docstrings, is_vision_available
+from .base import GenericTensor, Pipeline, build_pipeline_init_args
+
+
+if is_vision_available():
+    from ..image_utils import load_image
+
+
+@add_end_docstrings(
+    build_pipeline_init_args(has_image_processor=True),
+    """
+        image_processor_kwargs (`dict`, *optional*):
+                Additional dictionary of keyword arguments passed along to the image processor e.g.
+                {"size": {"height": 100, "width": 100}}
+    """,
+)
+class ImageFeatureExtractionPipeline(Pipeline):
+    """
+    Image feature extraction pipeline uses no model head. This pipeline extracts the hidden states from the base
+    transformer, which can be used as features in downstream tasks.
+
+    Example:
+
+    ```python
+    >>> from transformers import pipeline
+
+    >>> extractor = pipeline(model="google/vit-base-patch16-224", task="image-feature-extraction")
+    >>> result = extractor("https://huggingface.co/datasets/Narsil/image_dummy/raw/main/parrots.png", return_tensors=True)
+    >>> result.shape  # This is a tensor of shape [1, sequence_lenth, hidden_dimension] representing the input image.
+    torch.Size([1, 197, 768])
+    ```
+
+    Learn more about the basics of using a pipeline in the [pipeline tutorial](../pipeline_tutorial)
+
+    This image feature extraction pipeline can currently be loaded from [`pipeline`] using the task identifier:
+    `"image-feature-extraction"`.
+
+    All vision models may be used for this pipeline. See a list of all models, including community-contributed models on
+    [huggingface.co/models](https://huggingface.co/models).
+    """
+
+    def _sanitize_parameters(self, image_processor_kwargs=None, return_tensors=None, **kwargs):
+        preprocess_params = {} if image_processor_kwargs is None else image_processor_kwargs
+        postprocess_params = {"return_tensors": return_tensors} if return_tensors is not None else {}
+
+        if "timeout" in kwargs:
+            preprocess_params["timeout"] = kwargs["timeout"]
+
+        return preprocess_params, {}, postprocess_params
+
+    def preprocess(self, image, timeout=None, **image_processor_kwargs) -> Dict[str, GenericTensor]:
+        image = load_image(image, timeout=timeout)
+        model_inputs = self.image_processor(image, return_tensors=self.framework, **image_processor_kwargs)
+        return model_inputs
+
+    def _forward(self, model_inputs):
+        model_outputs = self.model(**model_inputs)
+        return model_outputs
+
+    def postprocess(self, model_outputs, return_tensors=False):
+        # [0] is the first available tensor, logits or last_hidden_state.
+        if return_tensors:
+            return model_outputs[0]
+        if self.framework == "pt":
+            return model_outputs[0].tolist()
+        elif self.framework == "tf":
+            return model_outputs[0].numpy().tolist()
+
+    def __call__(self, *args, **kwargs):
+        """
+        Extract the features of the input(s).
+
+        Args:
+            images (`str`, `List[str]`, `PIL.Image` or `List[PIL.Image]`):
+                The pipeline handles three types of images:
+
+                - A string containing a http link pointing to an image
+                - A string containing a local path to an image
+                - An image loaded in PIL directly
+
+                The pipeline accepts either a single image or a batch of images, which must then be passed as a string.
+                Images in a batch must all be in the same format: all as http links, all as local paths, or all as PIL
+                images.
+            timeout (`float`, *optional*, defaults to None):
+                The maximum time in seconds to wait for fetching images from the web. If None, no timeout is used and
+                the call may block forever.
+        Return:
+            A nested list of `float`: The features computed by the model.
+        """
+        return super().__call__(*args, **kwargs)

src/transformers/pipelines/image_to_text.py

+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. 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.
+
 from typing import List, Union
 
 from ..utils import (

tests/models/beit/test_modeling_beit.py

@@ -242,7 +242,7 @@ class BeitModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     )
     pipeline_model_mapping = (
         {
-            "feature-extraction": BeitModel,
+            "image-feature-extraction": BeitModel,
             "image-classification": BeitForImageClassification,
             "image-segmentation": BeitForSemanticSegmentation,
         }

tests/models/bit/test_modeling_bit.py

@@ -162,7 +162,7 @@ class BitModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
 
     all_model_classes = (BitModel, BitForImageClassification, BitBackbone) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": BitModel, "image-classification": BitForImageClassification}
+        {"image-feature-extraction": BitModel, "image-classification": BitForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/blip/test_modeling_blip.py

@@ -429,7 +429,10 @@ class BlipModelTester:
 class BlipModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (BlipModel,) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": BlipModel, "image-to-text": BlipForConditionalGeneration}
+        {
+            "feature-extraction": BlipModel,
+            "image-to-text": BlipForConditionalGeneration,
+        }
         if is_torch_available()
         else {}
     )

tests/models/clip/test_modeling_clip.py

@@ -477,7 +477,9 @@ class CLIPModelTester:
 @require_torch
 class CLIPModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (CLIPModel,) if is_torch_available() else ()
-    pipeline_model_mapping = {"feature-extraction": CLIPModel} if is_torch_available() else {}
+    pipeline_model_mapping = (
+        {"feature-extraction": CLIPModel, "image-feature-extraction": CLIPVisionModel} if is_torch_available() else {}
+    )
     fx_compatible = True
     test_head_masking = False
     test_pruning = False

tests/models/conditional_detr/test_modeling_conditional_detr.py

@@ -185,7 +185,7 @@ class ConditionalDetrModelTest(ModelTesterMixin, GenerationTesterMixin, Pipeline
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": ConditionalDetrModel, "object-detection": ConditionalDetrForObjectDetection}
+        {"image-feature-extraction": ConditionalDetrModel, "object-detection": ConditionalDetrForObjectDetection}
         if is_torch_available()
         else {}
     )

tests/models/convnext/test_modeling_convnext.py

@@ -172,7 +172,7 @@ class ConvNextModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": ConvNextModel, "image-classification": ConvNextForImageClassification}
+        {"image-feature-extraction": ConvNextModel, "image-classification": ConvNextForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/convnextv2/test_modeling_convnextv2.py

@@ -180,7 +180,7 @@ class ConvNextV2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCa
         else ()
     )
     pipeline_model_mapping = (
-        {"feature-extraction": ConvNextV2Model, "image-classification": ConvNextV2ForImageClassification}
+        {"image-feature-extraction": ConvNextV2Model, "image-classification": ConvNextV2ForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/cvt/test_modeling_cvt.py

@@ -151,7 +151,7 @@ class CvtModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
 
     all_model_classes = (CvtModel, CvtForImageClassification) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": CvtModel, "image-classification": CvtForImageClassification}
+        {"image-feature-extraction": CvtModel, "image-classification": CvtForImageClassification}
         if is_torch_available()
         else {}
     )

tests/models/data2vec/test_modeling_data2vec_vision.py

@@ -178,7 +178,7 @@ class Data2VecVisionModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.Te
     )
     pipeline_model_mapping = (
         {
-            "feature-extraction": Data2VecVisionModel,
+            "image-feature-extraction": Data2VecVisionModel,
             "image-classification": Data2VecVisionForImageClassification,
             "image-segmentation": Data2VecVisionForSemanticSegmentation,
         }

tests/models/deformable_detr/test_modeling_deformable_detr.py

@@ -191,7 +191,7 @@ class DeformableDetrModelTester:
 class DeformableDetrModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (DeformableDetrModel, DeformableDetrForObjectDetection) if is_torch_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": DeformableDetrModel, "object-detection": DeformableDetrForObjectDetection}
+        {"image-feature-extraction": DeformableDetrModel, "object-detection": DeformableDetrForObjectDetection}
         if is_torch_available()
         else {}
     )

tests/models/deit/test_modeling_deit.py

@@ -206,7 +206,7 @@ class DeiTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     )
     pipeline_model_mapping = (
         {
-            "feature-extraction": DeiTModel,
+            "image-feature-extraction": DeiTModel,
             "image-classification": (DeiTForImageClassification, DeiTForImageClassificationWithTeacher),
         }
         if is_torch_available()

tests/models/deta/test_modeling_deta.py

@@ -217,7 +217,7 @@ class DetaModelTester:
 class DetaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (DetaModel, DetaForObjectDetection) if is_torchvision_available() else ()
     pipeline_model_mapping = (
-        {"feature-extraction": DetaModel, "object-detection": DetaForObjectDetection}
+        {"image-feature-extraction": DetaModel, "object-detection": DetaForObjectDetection}
         if is_torchvision_available()
         else {}
     )