Add CLIPSeg (#20066)

* Add first draft

* Update conversion script

* Improve conversion script

* Improve conversion script some more

* Add conditional embeddings

* Add initial decoder

* Fix activation function of decoder

* Make decoder outputs match original implementation

* Make decoder outputs match original implementation

* Add more copied from statements

* Improve model outputs

* Fix auto tokenizer file

* Fix more tests

* Add test

* Improve README and docs, improve conditional embeddings

* Fix more tests

* Remove print statements

* Remove initial embeddings

* Improve conversion script

* Add interpolation of position embeddings

* Finish addition of interpolation of position embeddings

* Add support for refined checkpoint

* Fix refined checkpoint

* Remove unused parameter

* Improve conversion script

* Add support for training

* Fix conversion script

* Add CLIPSegFeatureExtractor

* Fix processor

* Fix CLIPSegProcessor

* Fix conversion script

* Fix most tests

* Fix equivalence test

* Fix README

* Add model to doc tests

* Use better variable name

* Convert other checkpoint as well

* Update config, add link to paper

* Add docs

* Update organization

* Replace base_model_prefix with clip

* Fix base_model_prefix

* Fix checkpoint of config

* Fix config checkpoint

* Remove file

* Use logits for output

* Fix tests
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

README.md

@@ -279,6 +279,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
+1. **[CLIPSeg](https://huggingface.co/docs/transformers/main/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.

README_es.md

@@ -279,6 +279,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
+1. **[CLIPSeg](https://huggingface.co/docs/transformers/main/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.

README_ja.md

@@ -314,6 +314,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
+1. **[CLIPSeg](https://huggingface.co/docs/transformers/main/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.

README_ko.md

@@ -229,6 +229,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
+1. **[CLIPSeg](https://huggingface.co/docs/transformers/main/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.

README_zh-hans.md

@@ -253,6 +253,7 @@ conda install -c huggingface transformers
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (来自 Inria/Facebook/Sorbonne) 伴随论文 [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) 由 Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot 发布。
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (来自 Google Research) 伴随论文 [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) 由 Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting 发布。
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (来自 OpenAI) 伴随论文 [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) 由 Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever 发布。
+1. **[CLIPSeg](https://huggingface.co/docs/transformers/main/model_doc/clipseg)** (来自 University of Göttingen) 伴随论文 [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) 由 Timo Lüddecke and Alexander Ecker 发布。
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (来自 Salesforce) 伴随论文 [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) 由 Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong 发布。
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (来自 Microsoft Research Asia) 伴随论文 [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 由 Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang 发布。
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (来自 YituTech) 伴随论文 [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 由 Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan 发布。

README_zh-hant.md

@@ -265,6 +265,7 @@ conda install -c huggingface transformers
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
+1. **[CLIPSeg](https://huggingface.co/docs/transformers/main/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.

docs/source/en/_toctree.yml

@@ -466,6 +466,8 @@
       sections:
       - local: model_doc/clip
         title: CLIP
+      - local: model_doc/clipseg
+        title: CLIPSeg
       - local: model_doc/data2vec
         title: Data2Vec
       - local: model_doc/donut

docs/source/en/index.mdx

@@ -67,6 +67,7 @@ The documentation is organized into five sections:
 1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
 1. **[CLIP](model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
+1. **[CLIPSeg](model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
 1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
@@ -223,6 +224,7 @@ Flax), PyTorch, and/or TensorFlow.
 |          CamemBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |           CANINE            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            CLIP             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|           CLIPSeg           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           CodeGen           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |      Conditional DETR       |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          ConvBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |

docs/source/en/model_doc/clipseg.mdx

+<!--Copyright 2022 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.
+-->
+
+# CLIPSeg
+
+## Overview
+
+The CLIPSeg model was proposed in [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke
+and Alexander Ecker. CLIPSeg adds a minimal decoder on top of a frozen [CLIP](clip) model for zero- and one-shot image segmentation.
+
+The abstract from the paper is the following:
+
+*Image segmentation is usually addressed by training a
+model for a fixed set of object classes. Incorporating additional classes or more complex queries later is expensive
+as it requires re-training the model on a dataset that encompasses these expressions. Here we propose a system
+that can generate image segmentations based on arbitrary
+prompts at test time. A prompt can be either a text or an
+image. This approach enables us to create a unified model
+(trained once) for three common segmentation tasks, which
+come with distinct challenges: referring expression segmentation, zero-shot segmentation and one-shot segmentation.
+We build upon the CLIP model as a backbone which we extend with a transformer-based decoder that enables dense
+prediction. After training on an extended version of the
+PhraseCut dataset, our system generates a binary segmentation map for an image based on a free-text prompt or on
+an additional image expressing the query. We analyze different variants of the latter image-based prompts in detail.
+This novel hybrid input allows for dynamic adaptation not
+only to the three segmentation tasks mentioned above, but
+to any binary segmentation task where a text or image query
+can be formulated. Finally, we find our system to adapt well
+to generalized queries involving affordances or properties*
+
+Tips:
+
+- [`CLIPSegForImageSegmentation`] adds a decoder on top of [`CLIPSegModel`]. The latter is identical to [`CLIPModel`].
+- [`CLIPSegForImageSegmentation`] can generate image segmentations based on arbitrary prompts at test time. A prompt can be either a text
+(provided to the model as `input_ids`) or an image (provided to the model as `conditional_pixel_values`). One can also provide custom
+conditional embeddings (provided to the model as `conditional_embeddings`).
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/clipseg_architecture.png"
+alt="drawing" width="600"/> 
+
+<small> CLIPSeg overview. Taken from the <a href="https://arxiv.org/abs/2112.10003">original paper.</a> </small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/timojl/clipseg).
+
+
+## CLIPSegConfig
+
+[[autodoc]] CLIPSegConfig
+    - from_text_vision_configs
+
+## CLIPSegTextConfig
+
+[[autodoc]] CLIPSegTextConfig
+
+## CLIPSegVisionConfig
+
+[[autodoc]] CLIPSegVisionConfig
+
+## CLIPSegProcessor
+
+[[autodoc]] CLIPSegProcessor
+
+## CLIPSegModel
+
+[[autodoc]] CLIPSegModel
+    - forward
+    - get_text_features
+    - get_image_features
+
+## CLIPSegTextModel
+
+[[autodoc]] CLIPSegTextModel
+    - forward
+
+## CLIPSegVisionModel
+
+[[autodoc]] CLIPSegVisionModel
+    - forward
+
+## CLIPSegForImageSegmentation
+
+[[autodoc]] CLIPSegForImageSegmentation
+    - forward
\ No newline at end of file

src/transformers/__init__.py

@@ -171,6 +171,13 @@ _import_structure = {
         "CLIPTokenizer",
         "CLIPVisionConfig",
     ],
+    "models.clipseg": [
+        "CLIPSEG_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "CLIPSegConfig",
+        "CLIPSegProcessor",
+        "CLIPSegTextConfig",
+        "CLIPSegVisionConfig",
+    ],
     "models.codegen": ["CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP", "CodeGenConfig", "CodeGenTokenizer"],
     "models.conditional_detr": ["CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConditionalDetrConfig"],
     "models.convbert": ["CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvBertConfig", "ConvBertTokenizer"],
@@ -1074,6 +1081,16 @@ else:
             "CLIPVisionModel",
         ]
     )
+    _import_structure["models.clipseg"].extend(
+        [
+            "CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "CLIPSegModel",
+            "CLIPSegPreTrainedModel",
+            "CLIPSegTextModel",
+            "CLIPSegVisionModel",
+            "CLIPSegForImageSegmentation",
+        ]
+    )
     _import_structure["models.x_clip"].extend(
         [
             "XCLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3225,6 +3242,13 @@ if TYPE_CHECKING:
         CLIPTokenizer,
         CLIPVisionConfig,
     )
+    from .models.clipseg import (
+        CLIPSEG_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        CLIPSegConfig,
+        CLIPSegProcessor,
+        CLIPSegTextConfig,
+        CLIPSegVisionConfig,
+    )
     from .models.codegen import CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP, CodeGenConfig, CodeGenTokenizer
     from .models.conditional_detr import CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, ConditionalDetrConfig
     from .models.convbert import CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvBertConfig, ConvBertTokenizer
@@ -3993,6 +4017,14 @@ if TYPE_CHECKING:
             CLIPTextModel,
             CLIPVisionModel,
         )
+        from .models.clipseg import (
+            CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST,
+            CLIPSegForImageSegmentation,
+            CLIPSegModel,
+            CLIPSegPreTrainedModel,
+            CLIPSegTextModel,
+            CLIPSegVisionModel,
+        )
         from .models.codegen import (
             CODEGEN_PRETRAINED_MODEL_ARCHIVE_LIST,
             CodeGenForCausalLM,

src/transformers/models/__init__.py

@@ -37,6 +37,7 @@ from . import (
     camembert,
     canine,
     clip,
+    clipseg,
     codegen,
     conditional_detr,
     convbert,

src/transformers/models/auto/configuration_auto.py

@@ -42,6 +42,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("camembert", "CamembertConfig"),
         ("canine", "CanineConfig"),
         ("clip", "CLIPConfig"),
+        ("clipseg", "CLIPSegConfig"),
         ("codegen", "CodeGenConfig"),
         ("conditional_detr", "ConditionalDetrConfig"),
         ("convbert", "ConvBertConfig"),
@@ -182,6 +183,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("camembert", "CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("canine", "CANINE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("clip", "CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("clipseg", "CLIPSEG_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("codegen", "CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("conditional_detr", "CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("convbert", "CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -315,6 +317,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("camembert", "CamemBERT"),
         ("canine", "CANINE"),
         ("clip", "CLIP"),
+        ("clipseg", "CLIPSeg"),
         ("codegen", "CodeGen"),
         ("conditional_detr", "Conditional DETR"),
         ("convbert", "ConvBERT"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -39,6 +39,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
     [
         ("beit", "BeitFeatureExtractor"),
         ("clip", "CLIPFeatureExtractor"),
+        ("clipseg", "ViTFeatureExtractor"),
         ("conditional_detr", "ConditionalDetrFeatureExtractor"),
         ("convnext", "ConvNextFeatureExtractor"),
         ("cvt", "ConvNextFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -41,6 +41,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("camembert", "CamembertModel"),
         ("canine", "CanineModel"),
         ("clip", "CLIPModel"),
+        ("clipseg", "CLIPSegModel"),
         ("codegen", "CodeGenModel"),
         ("conditional_detr", "ConditionalDetrModel"),
         ("convbert", "ConvBertModel"),
@@ -813,6 +814,7 @@ _MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
     [
         # Model for Zero Shot Image Classification mapping
         ("clip", "CLIPModel"),
+        ("clipseg", "CLIPSegModel"),
     ]
 )
 

src/transformers/models/auto/processing_auto.py

@@ -40,6 +40,7 @@ logger = logging.get_logger(__name__)
 PROCESSOR_MAPPING_NAMES = OrderedDict(
     [
         ("clip", "CLIPProcessor"),
+        ("clipseg", "CLIPSegProcessor"),
         ("flava", "FlavaProcessor"),
         ("groupvit", "CLIPProcessor"),
         ("layoutlmv2", "LayoutLMv2Processor"),

src/transformers/models/auto/tokenization_auto.py

@@ -93,6 +93,13 @@ else:
                     "CLIPTokenizerFast" if is_tokenizers_available() else None,
                 ),
             ),
+            (
+                "clipseg",
+                (
+                    "CLIPTokenizer",
+                    "CLIPTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
             ("codegen", ("CodeGenTokenizer", "CodeGenTokenizerFast" if is_tokenizers_available() else None)),
             ("convbert", ("ConvBertTokenizer", "ConvBertTokenizerFast" if is_tokenizers_available() else None)),
             (

src/transformers/models/clipseg/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 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.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_clipseg": [
+        "CLIPSEG_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "CLIPSegConfig",
+        "CLIPSegTextConfig",
+        "CLIPSegVisionConfig",
+    ],
+    "processing_clipseg": ["CLIPSegProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_clipseg"] = [
+        "CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "CLIPSegModel",
+        "CLIPSegPreTrainedModel",
+        "CLIPSegTextModel",
+        "CLIPSegVisionModel",
+        "CLIPSegForImageSegmentation",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_clipseg import (
+        CLIPSEG_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        CLIPSegConfig,
+        CLIPSegTextConfig,
+        CLIPSegVisionConfig,
+    )
+    from .processing_clipseg import CLIPSegProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_clipseg import (
+            CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST,
+            CLIPSegForImageSegmentation,
+            CLIPSegModel,
+            CLIPSegPreTrainedModel,
+            CLIPSegTextModel,
+            CLIPSegVisionModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/clipseg/configuration_clipseg.py

+# coding=utf-8
+# Copyright 2022 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.
+""" CLIPSeg model configuration"""
+
+import copy
+import os
+from typing import Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+CLIPSEG_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "CIDAS/clipseg-rd64": "https://huggingface.co/CIDAS/clipseg-rd64/resolve/main/config.json",
+}
+
+
+class CLIPSegTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to instantiate an
+    CLIPSeg model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the CLIPSeg
+    [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 49408):
+            Vocabulary size of the CLIPSeg text model. Defines the number of different tokens that can be represented
+            by the `inputs_ids` passed when calling [`CLIPSegModel`].
+        hidden_size (`int`, *optional*, defaults to 512):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        max_position_embeddings (`int`, *optional*, defaults to 77):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported. layer_norm_eps (`float`, *optional*,
+            defaults to 1e-5): The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float``, *optional*, defaults to 1):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPSegTextConfig, CLIPSegTextModel
+
+    >>> # Initializing a CLIPSegTextConfig with CIDAS/clipseg-rd64 style configuration
+    >>> configuration = CLIPSegTextConfig()
+
+    >>> # Initializing a CLIPSegTextModel (with random weights) from the CIDAS/clipseg-rd64 style configuration
+    >>> model = CLIPSegTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "clipseg_text_model"
+
+    def __init__(
+        self,
+        vocab_size=49408,
+        hidden_size=512,
+        intermediate_size=2048,
+        num_hidden_layers=12,
+        num_attention_heads=8,
+        max_position_embeddings=77,
+        hidden_act="quick_gelu",
+        layer_norm_eps=0.00001,
+        dropout=0.0,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from CLIPSegConfig
+        if config_dict.get("model_type") == "clipseg":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class CLIPSegVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to instantiate an
+    CLIPSeg model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the CLIPSeg
+    [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 32):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported. layer_norm_eps (`float`, *optional*,
+            defaults to 1e-5): The epsilon used by the layer normalization layers.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float``, *optional*, defaults to 1):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPSegVisionConfig, CLIPSegVisionModel
+
+    >>> # Initializing a CLIPSegVisionConfig with CIDAS/clipseg-rd64 style configuration
+    >>> configuration = CLIPSegVisionConfig()
+
+    >>> # Initializing a CLIPSegVisionModel (with random weights) from the CIDAS/clipseg-rd64 style configuration
+    >>> model = CLIPSegVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "clipseg_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="quick_gelu",
+        layer_norm_eps=0.00001,
+        dropout=0.0,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from CLIPSegConfig
+        if config_dict.get("model_type") == "clipseg":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class CLIPSegConfig(PretrainedConfig):
+    r"""
+    [`CLIPSegConfig`] is the configuration class to store the configuration of a [`CLIPSegModel`]. It is used to
+    instantiate a CLIPSeg model according to the specified arguments, defining the text model and vision model configs.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the CLIPSeg
+    [CIDAS/clipseg-rd64](https://huggingface.co/CIDAS/clipseg-rd64) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPSegTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`CLIPSegVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimensionality of text and vision projection layers.
+        logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
+            The inital value of the *logit_scale* paramter. Default is used as per the original CLIPSeg implementation.
+        extract_layers (`List[int]`, *optional*, defaults to [3, 6, 9]):
+            Layers to extract when forwarding the query image through the frozen visual backbone of CLIP.
+        reduce_dim (`int`, *optional*, defaults to 64):
+            Dimensionality to reduce the CLIP vision embedding.
+        decoder_num_attention_heads (`int`, *optional*, defaults to 4):
+            Number of attention heads in the decoder of CLIPSeg.
+        decoder_attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        decoder_hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported. layer_norm_eps (`float`, *optional*,
+            defaults to 1e-5): The epsilon used by the layer normalization layers.
+        decoder_intermediate_size (`int`, *optional*, defaults to 2048):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layers in the Transformer decoder.
+        conditional_layer (`int`, *optional*, defaults to 0):
+            The layer to use of the Transformer encoder whose activations will be combined with the condition
+            embeddings using FiLM (Feature-wise Linear Modulation). If 0, the last layer is used.
+        use_complex_transposed_convolution (`bool`, *optional*, defaults to `False`):
+            Whether to use a more complex transposed convolution in the decoder, enabling more fine-grained
+            segmentation.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPSegConfig, CLIPSegModel
+
+    >>> # Initializing a CLIPSegConfig with CIDAS/clipseg-rd64 style configuration
+    >>> configuration = CLIPSegConfig()
+
+    >>> # Initializing a CLIPSegModel (with random weights) from the CIDAS/clipseg-rd64 style configuration
+    >>> model = CLIPSegModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a CLIPSegConfig from a CLIPSegTextConfig and a CLIPSegVisionConfig
+
+    >>> # Initializing a CLIPSegText and CLIPSegVision configuration
+    >>> config_text = CLIPSegTextConfig()
+    >>> config_vision = CLIPSegVisionConfig()
+
+    >>> config = CLIPSegConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "clipseg"
+    is_composition = True
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        projection_dim=512,
+        logit_scale_init_value=2.6592,
+        extract_layers=[3, 6, 9],
+        reduce_dim=64,
+        decoder_num_attention_heads=4,
+        decoder_attention_dropout=0.0,
+        decoder_hidden_act="quick_gelu",
+        decoder_intermediate_size=2048,
+        conditional_layer=0,
+        use_complex_transposed_convolution=False,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        text_config_dict = kwargs.pop("text_config_dict", None)
+        vision_config_dict = kwargs.pop("vision_config_dict", None)
+        if text_config_dict is not None:
+            text_config = text_config_dict
+        if vision_config_dict is not None:
+            vision_config = vision_config_dict
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the CLIPSegTextConfig with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("vision_config is None. initializing the CLIPSegVisionConfig with default values.")
+
+        self.text_config = CLIPSegTextConfig(**text_config)
+        self.vision_config = CLIPSegVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.logit_scale_init_value = logit_scale_init_value
+        self.extract_layers = extract_layers
+        self.reduce_dim = reduce_dim
+        self.decoder_num_attention_heads = decoder_num_attention_heads
+        self.decoder_attention_dropout = decoder_attention_dropout
+        self.decoder_hidden_act = decoder_hidden_act
+        self.decoder_intermediate_size = decoder_intermediate_size
+        self.conditional_layer = conditional_layer
+        self.initializer_factor = 1.0
+        self.use_complex_transposed_convolution = use_complex_transposed_convolution
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: CLIPSegTextConfig, vision_config: CLIPSegVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`CLIPSegConfig`] (or a derived class) from clipseg text model configuration and clipseg vision
+        model configuration.
+
+        Returns:
+            [`CLIPSegConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["text_config"] = self.text_config.to_dict()
+        output["vision_config"] = self.vision_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/clipseg/convert_clipseg_original_pytorch_to_hf.py

+# coding=utf-8
+# Copyright 2022 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.
+
+"""Convert CLIPSeg checkpoints from the original repository. URL: https://github.com/timojl/clipseg."""
+
+import argparse
+
+import torch
+from PIL import Image
+
+import requests
+from transformers import (
+    CLIPSegConfig,
+    CLIPSegForImageSegmentation,
+    CLIPSegProcessor,
+    CLIPSegTextConfig,
+    CLIPSegVisionConfig,
+    CLIPTokenizer,
+    ViTFeatureExtractor,
+)
+
+
+def get_clipseg_config(model_name):
+    text_config = CLIPSegTextConfig()
+    vision_config = CLIPSegVisionConfig(patch_size=16)
+
+    use_complex_transposed_convolution = True if "refined" in model_name else False
+    reduce_dim = 16 if "rd16" in model_name else 64
+
+    config = CLIPSegConfig.from_text_vision_configs(
+        text_config,
+        vision_config,
+        use_complex_transposed_convolution=use_complex_transposed_convolution,
+        reduce_dim=reduce_dim,
+    )
+    return config
+
+
+def rename_key(name):
+    # update prefixes
+    if "clip_model" in name:
+        name = name.replace("clip_model", "clip")
+    if "transformer" in name:
+        if "visual" in name:
+            name = name.replace("visual.transformer", "vision_model")
+        else:
+            name = name.replace("transformer", "text_model")
+    if "resblocks" in name:
+        name = name.replace("resblocks", "encoder.layers")
+    if "ln_1" in name:
+        name = name.replace("ln_1", "layer_norm1")
+    if "ln_2" in name:
+        name = name.replace("ln_2", "layer_norm2")
+    if "c_fc" in name:
+        name = name.replace("c_fc", "fc1")
+    if "c_proj" in name:
+        name = name.replace("c_proj", "fc2")
+    if "attn" in name and "self" not in name:
+        name = name.replace("attn", "self_attn")
+    # text encoder
+    if "token_embedding" in name:
+        name = name.replace("token_embedding", "text_model.embeddings.token_embedding")
+    if "positional_embedding" in name and "visual" not in name:
+        name = name.replace("positional_embedding", "text_model.embeddings.position_embedding.weight")
+    if "ln_final" in name:
+        name = name.replace("ln_final", "text_model.final_layer_norm")
+    # vision encoder
+    if "visual.class_embedding" in name:
+        name = name.replace("visual.class_embedding", "vision_model.embeddings.class_embedding")
+    if "visual.conv1" in name:
+        name = name.replace("visual.conv1", "vision_model.embeddings.patch_embedding")
+    if "visual.positional_embedding" in name:
+        name = name.replace("visual.positional_embedding", "vision_model.embeddings.position_embedding.weight")
+    if "visual.ln_pre" in name:
+        name = name.replace("visual.ln_pre", "vision_model.pre_layrnorm")
+    if "visual.ln_post" in name:
+        name = name.replace("visual.ln_post", "vision_model.post_layernorm")
+    # projection layers
+    if "visual.proj" in name:
+        name = name.replace("visual.proj", "visual_projection.weight")
+    if "text_projection" in name:
+        name = name.replace("text_projection", "text_projection.weight")
+    # decoder
+    if "trans_conv" in name:
+        name = name.replace("trans_conv", "transposed_convolution")
+    if "film_mul" in name or "film_add" in name or "reduce" in name or "transposed_convolution" in name:
+        name = "decoder." + name
+    if "blocks" in name:
+        name = name.replace("blocks", "decoder.layers")
+    if "linear1" in name:
+        name = name.replace("linear1", "mlp.fc1")
+    if "linear2" in name:
+        name = name.replace("linear2", "mlp.fc2")
+    if "norm1" in name and "layer_" not in name:
+        name = name.replace("norm1", "layer_norm1")
+    if "norm2" in name and "layer_" not in name:
+        name = name.replace("norm2", "layer_norm2")
+
+    return name
+
+
+def convert_state_dict(orig_state_dict, config):
+    for key in orig_state_dict.copy().keys():
+        val = orig_state_dict.pop(key)
+
+        if key.startswith("clip_model") and "attn.in_proj" in key:
+            key_split = key.split(".")
+            if "visual" in key:
+                layer_num = int(key_split[4])
+                dim = config.vision_config.hidden_size
+                prefix = "vision_model"
+            else:
+                layer_num = int(key_split[3])
+                dim = config.text_config.hidden_size
+                prefix = "text_model"
+
+            if "weight" in key:
+                orig_state_dict[f"clip.{prefix}.encoder.layers.{layer_num}.self_attn.q_proj.weight"] = val[:dim, :]
+                orig_state_dict[f"clip.{prefix}.encoder.layers.{layer_num}.self_attn.k_proj.weight"] = val[
+                    dim : dim * 2, :
+                ]
+                orig_state_dict[f"clip.{prefix}.encoder.layers.{layer_num}.self_attn.v_proj.weight"] = val[-dim:, :]
+            else:
+                orig_state_dict[f"clip.{prefix}.encoder.layers.{layer_num}.self_attn.q_proj.bias"] = val[:dim]
+                orig_state_dict[f"clip.{prefix}.encoder.layers.{layer_num}.self_attn.k_proj.bias"] = val[dim : dim * 2]
+                orig_state_dict[f"clip.{prefix}.encoder.layers.{layer_num}.self_attn.v_proj.bias"] = val[-dim:]
+        elif "self_attn" in key and "out_proj" not in key:
+            key_split = key.split(".")
+            layer_num = int(key_split[1])
+            dim = config.reduce_dim
+            if "weight" in key:
+                orig_state_dict[f"decoder.layers.{layer_num}.self_attn.q_proj.weight"] = val[:dim, :]
+                orig_state_dict[f"decoder.layers.{layer_num}.self_attn.k_proj.weight"] = val[dim : dim * 2, :]
+                orig_state_dict[f"decoder.layers.{layer_num}.self_attn.v_proj.weight"] = val[-dim:, :]
+            else:
+                orig_state_dict[f"decoder.layers.{layer_num}.self_attn.q_proj.bias"] = val[:dim]
+                orig_state_dict[f"decoder.layers.{layer_num}.self_attn.k_proj.bias"] = val[dim : dim * 2]
+                orig_state_dict[f"decoder.layers.{layer_num}.self_attn.v_proj.bias"] = val[-dim:]
+        else:
+            new_name = rename_key(key)
+            if "visual_projection" in new_name or "text_projection" in new_name:
+                val = val.T
+            orig_state_dict[new_name] = val
+
+    return orig_state_dict
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    image = Image.open(requests.get(url, stream=True).raw)
+    return image
+
+
+def convert_clipseg_checkpoint(model_name, checkpoint_path, pytorch_dump_folder_path, push_to_hub):
+    config = get_clipseg_config(model_name)
+    model = CLIPSegForImageSegmentation(config)
+    model.eval()
+
+    state_dict = torch.load(checkpoint_path, map_location="cpu")
+
+    # remove some keys
+    for key in state_dict.copy().keys():
+        if key.startswith("model"):
+            state_dict.pop(key, None)
+
+    # rename some keys
+    state_dict = convert_state_dict(state_dict, config)
+    missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False)
+
+    if missing_keys != ["clip.text_model.embeddings.position_ids", "clip.vision_model.embeddings.position_ids"]:
+        raise ValueError("Missing keys that are not expected: {}".format(missing_keys))
+    if unexpected_keys != ["decoder.reduce.weight", "decoder.reduce.bias"]:
+        raise ValueError(f"Unexpected keys: {unexpected_keys}")
+
+    feature_extractor = ViTFeatureExtractor(size=352)
+    tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+    processor = CLIPSegProcessor(feature_extractor=feature_extractor, tokenizer=tokenizer)
+
+    image = prepare_img()
+    text = ["a glass", "something to fill", "wood", "a jar"]
+
+    inputs = processor(text=text, images=[image] * len(text), padding="max_length", return_tensors="pt")
+
+    with torch.no_grad():
+        outputs = model(**inputs)
+
+    # verify values
+    expected_conditional = torch.tensor([0.1110, -0.1882, 0.1645])
+    expected_pooled_output = torch.tensor([0.2692, -0.7197, -0.1328])
+    if model_name == "clipseg-rd64-refined":
+        expected_masks_slice = torch.tensor(
+            [[-10.0407, -9.9431, -10.2646], [-9.9751, -9.7064, -9.9586], [-9.6891, -9.5645, -9.9618]]
+        )
+    elif model_name == "clipseg-rd64":
+        expected_masks_slice = torch.tensor(
+            [[-7.2877, -7.2711, -7.2463], [-7.2652, -7.2780, -7.2520], [-7.2239, -7.2204, -7.2001]]
+        )
+    elif model_name == "clipseg-rd16":
+        expected_masks_slice = torch.tensor(
+            [[-6.3955, -6.4055, -6.4151], [-6.3911, -6.4033, -6.4100], [-6.3474, -6.3702, -6.3762]]
+        )
+    else:
+        raise ValueError(f"Model name {model_name} not supported.")
+
+    assert torch.allclose(outputs.logits[0, :3, :3], expected_masks_slice, atol=1e-3)
+    assert torch.allclose(outputs.conditional_embeddings[0, :3], expected_conditional, atol=1e-3)
+    assert torch.allclose(outputs.pooled_output[0, :3], expected_pooled_output, atol=1e-3)
+    print("Looks ok!")
+
+    if pytorch_dump_folder_path is not None:
+        print(f"Saving model and processor to {pytorch_dump_folder_path}")
+        model.save_pretrained(pytorch_dump_folder_path)
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        print(f"Pushing model and processor for {model_name} to the hub")
+        model.push_to_hub(f"CIDAS/{model_name}")
+        processor.push_to_hub(f"CIDAS/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--model_name",
+        default="clipseg-rd64",
+        type=str,
+        choices=["clipseg-rd16", "clipseg-rd64", "clipseg-rd64-refined"],
+        help=(
+            "Name of the model. Supported models are: clipseg-rd64, clipseg-rd16 and clipseg-rd64-refined (rd meaning"
+            " reduce dimension)"
+        ),
+    )
+    parser.add_argument(
+        "--checkpoint_path",
+        default="/Users/nielsrogge/Documents/CLIPSeg/clip_plus_rd64-uni.pth",
+        type=str,
+        help=(
+            "Path to the original checkpoint. Note that the script assumes that the checkpoint includes both CLIP and"
+            " the decoder weights."
+        ),
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
+    )
+    parser.add_argument(
+        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
+    )
+
+    args = parser.parse_args()
+    convert_clipseg_checkpoint(args.model_name, args.checkpoint_path, args.pytorch_dump_folder_path, args.push_to_hub)