Add I-JEPA

docs/source/en/_toctree.yml

@@ -653,6 +653,8 @@
         title: GLPN
       - local: model_doc/hiera
         title: Hiera
+      - local: model_doc/ijepa
+        title: I-JEPA
       - local: model_doc/imagegpt
         title: ImageGPT
       - local: model_doc/levit

docs/source/en/index.md

@@ -168,6 +168,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                         [Hiera](model_doc/hiera)                         |       ✅        |         ❌         |      ❌      |
 |                        [Hubert](model_doc/hubert)                        |       ✅        |         ✅         |      ❌      |
 |                        [I-BERT](model_doc/ibert)                         |       ✅        |         ❌         |      ❌      |
+|                        [I-JEPA](model_doc/ijepa)                         |       ✅        |         ❌         |      ❌      |
 |                       [IDEFICS](model_doc/idefics)                       |       ✅        |         ✅         |      ❌      |
 |                      [Idefics2](model_doc/idefics2)                      |       ✅        |         ❌         |      ❌      |
 |                      [Idefics3](model_doc/idefics3)                      |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/ijepa.md

@@ -0,0 +1,78 @@
+<!--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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# I-JEPA
+
+## Overview
+
+The I-JEPA model was proposed in [Image-based Joint-Embedding Predictive Architecture](https://arxiv.org/pdf/2301.08243.pdf) by Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, Nicolas Ballas.
+I-JEPA is a self-supervised learning method that predicts the representations of one part of an image based on other parts of the same image. This approach focuses on learning semantic features without relying on pre-defined invariances from hand-crafted data transformations, which can bias specific tasks, or on filling in pixel-level details, which often leads to less meaningful representations.
+
+The abstract from the paper is the following:
+
+This paper demonstrates an approach for learning highly semantic image representations without relying on hand-crafted data-augmentations. We introduce the Image- based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. A core design choice to guide I-JEPA towards producing semantic representations is the masking strategy; specifically, it is crucial to (a) sample tar- get blocks with sufficiently large scale (semantic), and to (b) use a sufficiently informative (spatially distributed) context block. Empirically, when combined with Vision Transform- ers, we find I-JEPA to be highly scalable. For instance, we train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks, from linear classification to object counting and depth prediction.
+
+This model was contributed by [jmtzt](https://huggingface.co/jmtzt).
+The original code can be found [here](https://github.com/facebookresearch/ijepa).
+
+## How to use
+
+Here is how to use this model for image feature extraction:
+
+```python
+import requests
+import torch
+from PIL import Image
+from torch.nn.functional import cosine_similarity
+
+from transformers import AutoModel, AutoProcessor
+
+url_1 = "http://images.cocodataset.org/val2017/000000039769.jpg"
+url_2 = "http://images.cocodataset.org/val2017/000000219578.jpg"
+image_1 = Image.open(requests.get(url_1, stream=True).raw)
+image_2 = Image.open(requests.get(url_2, stream=True).raw)
+
+model_id = "jmtzt/ijepa_vith14_1k"
+processor = AutoProcessor.from_pretrained(model_id)
+model = AutoModel.from_pretrained(model_id)
+
+@torch.no_grad()
+def infer(image):
+    inputs = processor(image, return_tensors="pt")
+    outputs = model(**inputs)
+    return outputs.last_hidden_state.mean(dim=1)
+
+
+embed_1 = infer(image_1)
+embed_2 = infer(image_2)
+
+similarity = cosine_similarity(embed_1, embed_2)
+print(similarity)
+```
+
+## IJepaConfig
+
+[[autodoc]] IJepaConfig
+
+## IJepaModel
+
+[[autodoc]] IJepaModel
+    - forward
+
+## IJepaForImageClassification
+
+[[autodoc]] IJepaForImageClassification
+    - forward

docs/source/en/perf_infer_gpu_one.md

@@ -234,14 +234,15 @@ For now, Transformers supports SDPA inference and training for the following arc
 * [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
 * [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
 * [Gemma2](https://huggingface.co/docs/transformers/model_doc/gemma2#transformers.Gemma2Model)
+* [Granite](https://huggingface.co/docs/transformers/model_doc/granite#transformers.GraniteModel)
 * [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2)
 * [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
 * [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel)
 * [Hubert](https://huggingface.co/docs/transformers/model_doc/hubert#transformers.HubertModel)
 * [Idefics](https://huggingface.co/docs/transformers/model_doc/idefics#transformers.IdeficsModel)
 * [Idefics2](https://huggingface.co/docs/transformers/model_doc/idefics2#transformers.Idefics2Model)
 * [Idefics3](https://huggingface.co/docs/transformers/model_doc/idefics3#transformers.Idefics3Model)
-* [Granite](https://huggingface.co/docs/transformers/model_doc/granite#transformers.GraniteModel)
+* [I-JEPA](https://huggingface.co/docs/transformers/model_doc/ijepa#transformers.IJepaModel)
 * [GraniteMoe](https://huggingface.co/docs/transformers/model_doc/granitemoe#transformers.GraniteMoeModel)
 * [JetMoe](https://huggingface.co/docs/transformers/model_doc/jetmoe#transformers.JetMoeModel)
 * [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)

src/transformers/__init__.py

@@ -484,6 +484,7 @@
     "models.idefics": ["IdeficsConfig"],
     "models.idefics2": ["Idefics2Config"],
     "models.idefics3": ["Idefics3Config"],
+    "models.ijepa": ["IJepaConfig"],
     "models.imagegpt": ["ImageGPTConfig"],
     "models.informer": ["InformerConfig"],
     "models.instructblip": [
@@ -2456,6 +2457,13 @@
             "Idefics3Processor",
         ]
     )
+    _import_structure["models.ijepa"].extend(
+        [
+            "IJepaForImageClassification",
+            "IJepaModel",
+            "IJepaPreTrainedModel",
+        ]
+    )
     _import_structure["models.imagegpt"].extend(
         [
             "ImageGPTForCausalImageModeling",
@@ -5355,6 +5363,7 @@
     )
     from .models.idefics2 import Idefics2Config
     from .models.idefics3 import Idefics3Config
+    from .models.ijepa import IJepaConfig
     from .models.imagegpt import ImageGPTConfig
     from .models.informer import InformerConfig
     from .models.instructblip import (
@@ -7166,6 +7175,11 @@
             Idefics3PreTrainedModel,
             Idefics3Processor,
         )
+        from .models.ijepa import (
+            IJepaForImageClassification,
+            IJepaModel,
+            IJepaPreTrainedModel,
+        )
         from .models.imagegpt import (
             ImageGPTForCausalImageModeling,
             ImageGPTForImageClassification,

src/transformers/models/__init__.py

@@ -117,6 +117,7 @@
     idefics,
     idefics2,
     idefics3,
+    ijepa,
     imagegpt,
     informer,
     instructblip,

src/transformers/models/auto/configuration_auto.py

@@ -22,7 +22,10 @@
 from typing import List, Union
 
 from ...configuration_utils import PretrainedConfig
-from ...dynamic_module_utils import get_class_from_dynamic_module, resolve_trust_remote_code
+from ...dynamic_module_utils import (
+    get_class_from_dynamic_module,
+    resolve_trust_remote_code,
+)
 from ...utils import CONFIG_NAME, logging
 
 
@@ -135,6 +138,7 @@
         ("idefics", "IdeficsConfig"),
         ("idefics2", "Idefics2Config"),
         ("idefics3", "Idefics3Config"),
+        ("ijepa", "IJepaConfig"),
         ("imagegpt", "ImageGPTConfig"),
         ("informer", "InformerConfig"),
         ("instructblip", "InstructBlipConfig"),
@@ -439,6 +443,7 @@
         ("idefics", "IDEFICS"),
         ("idefics2", "Idefics2"),
         ("idefics3", "Idefics3"),
+        ("ijepa", "I-JEPA"),
         ("imagegpt", "ImageGPT"),
         ("informer", "Informer"),
         ("instructblip", "InstructBLIP"),
@@ -877,7 +882,11 @@ def docstring_decorator(fn):
             indent = re.search(r"^(\s*)List options\s*$", lines[i]).groups()[0]
             if use_model_types:
                 indent = f"{indent}    "
-            lines[i] = _list_model_options(indent, config_to_class=config_to_class, use_model_types=use_model_types)
+            lines[i] = _list_model_options(
+                indent,
+                config_to_class=config_to_class,
+                use_model_types=use_model_types,
+            )
             docstrings = "\n".join(lines)
         else:
             raise ValueError(
@@ -1018,13 +1027,19 @@ def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
         has_remote_code = "auto_map" in config_dict and "AutoConfig" in config_dict["auto_map"]
         has_local_code = "model_type" in config_dict and config_dict["model_type"] in CONFIG_MAPPING
         trust_remote_code = resolve_trust_remote_code(
-            trust_remote_code, pretrained_model_name_or_path, has_local_code, has_remote_code
+            trust_remote_code,
+            pretrained_model_name_or_path,
+            has_local_code,
+            has_remote_code,
         )
 
         if has_remote_code and trust_remote_code:
             class_ref = config_dict["auto_map"]["AutoConfig"]
             config_class = get_class_from_dynamic_module(
-                class_ref, pretrained_model_name_or_path, code_revision=code_revision, **kwargs
+                class_ref,
+                pretrained_model_name_or_path,
+                code_revision=code_revision,
+                **kwargs,
             )
             if os.path.isdir(pretrained_model_name_or_path):
                 config_class.register_for_auto_class()

src/transformers/models/auto/image_processing_auto.py

@@ -23,7 +23,10 @@
 
 # Build the list of all image processors
 from ...configuration_utils import PretrainedConfig
-from ...dynamic_module_utils import get_class_from_dynamic_module, resolve_trust_remote_code
+from ...dynamic_module_utils import (
+    get_class_from_dynamic_module,
+    resolve_trust_remote_code,
+)
 from ...image_processing_utils import BaseImageProcessor, ImageProcessingMixin
 from ...image_processing_utils_fast import BaseImageProcessorFast
 from ...utils import (
@@ -90,6 +93,7 @@
             ("idefics", ("IdeficsImageProcessor",)),
             ("idefics2", ("Idefics2ImageProcessor",)),
             ("idefics3", ("Idefics3ImageProcessor",)),
+            ("ijepa", ("ViTImageProcessor", "ViTImageProcessorFast")),
             ("imagegpt", ("ImageGPTImageProcessor",)),
             ("instructblip", ("BlipImageProcessor",)),
             ("instructblipvideo", ("InstructBlipVideoImageProcessor",)),
@@ -165,7 +169,10 @@
     else:
         fast_image_processor_class = fast_image_processor_class[0]
 
-    IMAGE_PROCESSOR_MAPPING_NAMES[model_type] = (slow_image_processor_class, fast_image_processor_class)
+    IMAGE_PROCESSOR_MAPPING_NAMES[model_type] = (
+        slow_image_processor_class,
+        fast_image_processor_class,
+    )
 
 IMAGE_PROCESSOR_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, IMAGE_PROCESSOR_MAPPING_NAMES)
 
@@ -433,7 +440,9 @@ def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
         if image_processor_class is None and image_processor_auto_map is None:
             if not isinstance(config, PretrainedConfig):
                 config = AutoConfig.from_pretrained(
-                    pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
+                    pretrained_model_name_or_path,
+                    trust_remote_code=trust_remote_code,
+                    **kwargs,
                 )
             # It could be in `config.image_processor_type``
             image_processor_class = getattr(config, "image_processor_type", None)
@@ -452,7 +461,10 @@ def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
         has_remote_code = image_processor_auto_map is not None
         has_local_code = image_processor_class is not None or type(config) in IMAGE_PROCESSOR_MAPPING
         trust_remote_code = resolve_trust_remote_code(
-            trust_remote_code, pretrained_model_name_or_path, has_local_code, has_remote_code
+            trust_remote_code,
+            pretrained_model_name_or_path,
+            has_local_code,
+            has_remote_code,
         )
 
         if image_processor_auto_map is not None and not isinstance(image_processor_auto_map, tuple):
@@ -553,5 +565,7 @@ def register(
                 fast_image_processor_class = existing_fast
 
         IMAGE_PROCESSOR_MAPPING.register(
-            config_class, (slow_image_processor_class, fast_image_processor_class), exist_ok=exist_ok
+            config_class,
+            (slow_image_processor_class, fast_image_processor_class),
+            exist_ok=exist_ok,
         )

src/transformers/models/auto/modeling_auto.py

@@ -132,6 +132,7 @@
         ("idefics", "IdeficsModel"),
         ("idefics2", "Idefics2Model"),
         ("idefics3", "Idefics3Model"),
+        ("ijepa", "IJepaModel"),
         ("imagegpt", "ImageGPTModel"),
         ("informer", "InformerModel"),
         ("jamba", "JambaModel"),
@@ -576,6 +577,7 @@
         ("focalnet", "FocalNetModel"),
         ("glpn", "GLPNModel"),
         ("hiera", "HieraModel"),
+        ("ijepa", "IJepaModel"),
         ("imagegpt", "ImageGPTModel"),
         ("levit", "LevitModel"),
         ("mllama", "MllamaVisionModel"),
@@ -639,7 +641,10 @@
         ("data2vec-vision", "Data2VecVisionForImageClassification"),
         (
             "deit",
-            ("DeiTForImageClassification", "DeiTForImageClassificationWithTeacher"),
+            (
+                "DeiTForImageClassification",
+                "DeiTForImageClassificationWithTeacher",
+            ),
         ),
         ("dinat", "DinatForImageClassification"),
         ("dinov2", "Dinov2ForImageClassification"),
@@ -653,10 +658,14 @@
         ("efficientnet", "EfficientNetForImageClassification"),
         ("focalnet", "FocalNetForImageClassification"),
         ("hiera", "HieraForImageClassification"),
+        ("ijepa", "IJepaForImageClassification"),
         ("imagegpt", "ImageGPTForImageClassification"),
         (
             "levit",
-            ("LevitForImageClassification", "LevitForImageClassificationWithTeacher"),
+            (
+                "LevitForImageClassification",
+                "LevitForImageClassificationWithTeacher",
+            ),
         ),
         ("mobilenet_v1", "MobileNetV1ForImageClassification"),
         ("mobilenet_v2", "MobileNetV2ForImageClassification"),
@@ -991,7 +1000,10 @@
         ("reformer", "ReformerForSequenceClassification"),
         ("rembert", "RemBertForSequenceClassification"),
         ("roberta", "RobertaForSequenceClassification"),
-        ("roberta-prelayernorm", "RobertaPreLayerNormForSequenceClassification"),
+        (
+            "roberta-prelayernorm",
+            "RobertaPreLayerNormForSequenceClassification",
+        ),
         ("roc_bert", "RoCBertForSequenceClassification"),
         ("roformer", "RoFormerForSequenceClassification"),
         ("squeezebert", "SqueezeBertForSequenceClassification"),
@@ -1437,7 +1449,8 @@
     CONFIG_MAPPING_NAMES, MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES
 )
 MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING = _LazyAutoMapping(
-    CONFIG_MAPPING_NAMES, MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES
+    CONFIG_MAPPING_NAMES,
+    MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES,
 )
 MODEL_FOR_IMAGE_SEGMENTATION_MAPPING = _LazyAutoMapping(
     CONFIG_MAPPING_NAMES, MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES
@@ -1677,7 +1690,8 @@ class AutoModelForZeroShotImageClassification(_BaseAutoModelClass):
 
 
 AutoModelForZeroShotImageClassification = auto_class_update(
-    AutoModelForZeroShotImageClassification, head_doc="zero-shot image classification"
+    AutoModelForZeroShotImageClassification,
+    head_doc="zero-shot image classification",
 )
 
 
@@ -1778,7 +1792,8 @@ class AutoModelForSpeechSeq2Seq(_BaseAutoModelClass):
 
 
 AutoModelForSpeechSeq2Seq = auto_class_update(
-    AutoModelForSpeechSeq2Seq, head_doc="sequence-to-sequence speech-to-text modeling"
+    AutoModelForSpeechSeq2Seq,
+    head_doc="sequence-to-sequence speech-to-text modeling",
 )
 
 
@@ -1787,7 +1802,8 @@ class AutoModelForAudioFrameClassification(_BaseAutoModelClass):
 
 
 AutoModelForAudioFrameClassification = auto_class_update(
-    AutoModelForAudioFrameClassification, head_doc="audio frame (token) classification"
+    AutoModelForAudioFrameClassification,
+    head_doc="audio frame (token) classification",
 )