init commit

.flake8

@@ -0,0 +1,7 @@
+[flake8]
+ignore = W503, E203, E221, C901, C408, E741, C407, B017, F811, C101, EXE001, EXE002
+max-line-length = 100
+max-complexity = 18
+select = B,C,E,F,W,T4,B9
+per-file-ignores =
+  **/__init__.py:F401,F403,E402

.gitattributes

@@ -32,3 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,37 @@
+.nfs*
+
+# compilation and distribution
+__pycache__
+_ext
+*.pyc
+*.pyd
+*.so
+*.dll
+*.egg-info/
+build/
+dist/
+wheels/
+
+# pytorch/python/numpy formats
+*.pth
+*.pkl
+*.npy
+*.ts
+model_ts*.txt
+
+# onnx models
+*.onnx
+
+# ipython/jupyter notebooks
+**/.ipynb_checkpoints/
+
+# Editor temporaries
+*.swn
+*.swo
+*.swp
+*~
+
+# editor settings
+.idea
+.vscode
+_darcs

CODE_OF_CONDUCT.md

@@ -0,0 +1,80 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+  advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+  address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+This Code of Conduct also applies outside the project spaces when there is a
+reasonable belief that an individual's behavior may have a negative impact on
+the project or its community.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <[email protected]>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

CONTRIBUTING.md

@@ -0,0 +1,31 @@
+# Contributing to segment-anything
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints, using the `linter.sh` script in the project's root directory. Linting requires `black==23.*`, `isort==5.12.0`, `flake8`, and `mypy`.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+By contributing to segment-anything, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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.

README.md

@@ -1,13 +1,138 @@
----
-title: Expedit SAM
-emoji: 👁
-colorFrom: red
-colorTo: yellow
-sdk: gradio
-sdk_version: 3.24.1
-app_file: app.py
-pinned: false
-license: apache-2.0
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Expediting SAM without Fine-tuning
+
+<!-- **[Meta AI Research, FAIR](https://ai.facebook.com/research/)**
+
+[Alexander Kirillov](https://alexander-kirillov.github.io/), [Eric Mintun](https://ericmintun.github.io/), [Nikhila Ravi](https://nikhilaravi.com/), [Hanzi Mao](https://hanzimao.me/), Chloe Rolland, Laura Gustafson, [Tete Xiao](https://tetexiao.com), [Spencer Whitehead](https://www.spencerwhitehead.com/), Alex Berg, Wan-Yen Lo, [Piotr Dollar](https://pdollar.github.io/), [Ross Girshick](https://www.rossgirshick.info/)
+
+[[`Paper`](https://ai.facebook.com/research/publications/segment-anything/)] [[`Project`](https://segment-anything.com/)] [[`Demo`](https://segment-anything.com/demo)] [[`Dataset`](https://segment-anything.com/dataset/index.html)] [[`Blog`](https://ai.facebook.com/blog/segment-anything-foundation-model-image-segmentation/)] [[`BibTeX`](#citing-segment-anything)]
+
+![SAM design](assets/model_diagram.png?raw=true)
+
+The **Segment Anything Model (SAM)** produces high quality object masks from input prompts such as points or boxes, and it can be used to generate masks for all objects in an image. It has been trained on a [dataset](https://segment-anything.com/dataset/index.html) of 11 million images and 1.1 billion masks, and has strong zero-shot performance on a variety of segmentation tasks.
+
+<p float="left">
+  <img src="assets/masks1.png?raw=true" width="37.25%" />
+  <img src="assets/masks2.jpg?raw=true" width="61.5%" /> 
+</p> -->
+
+## Introduction
+
+This is the official implementation of the paper "[Expediting Large-Scale Vision Transformer for Dense Prediction without Fine-tuning](https://arxiv.org/abs/2210.01035)" on [Segment Anything Model (SAM)](https://segment-anything.com/).
+
+![framework](assets/Hourglass_transformer_framework.png)
+![framework](assets/TokenClusterReconstruct_Details.png)
+
+Our method can speed up SAM without any training. The bottleneck of SAM is image encoder. We implement our method on image encoder to signifficantly speed up the generation process. We test our method on different SAM models using a single 16G Tesla-V100. We set `--points-per-side=12` and `--points-per-batch=144` so that the generation process executes only one time.
+
+| Model            | clustering location | num of clusters | speed(image/s)     |
+| ---------------- | ------------------- | --------------- | ------------------ |
+| SAM-ViT-H        | -                   | -               | 1.27               |
+| SAM-ViT-H + ours | 18                   | 121             | 1.40(1.10x faster) |
+| SAM-ViT-H + ours | 14                   | 100             | 1.52(1.19x faster) |
+| SAM-ViT-H + ours | 8                   | 100             | 1.64(1.30x faster) |
+| SAM-ViT-H + ours | 8                   | 81              | 1.82(1.44x faster) |
+| SAM-ViT-H + ours | 6                   | 81              | 1.89(1.49x faster) |
+
+Here is the visualization of the setting above.
+
+![result of sam-vit-h + ours](assets/result_vit_h.png)
+
+We also try to implement our method on smaller model. Here are some examples generate by SAM w/ ViT-L + ours, with the setting of `--points-per-side=16` and `--points-per-batch=256`.  
+
+![result of sam-vit-l + ours](assets/result_vit_l.png)
+
+## Installation
+
+The code requires `python>=3.8`, as well as `pytorch>=1.7` and `torchvision>=0.8`. Please follow the instructions [here](https://pytorch.org/get-started/locally/) to install both PyTorch and TorchVision dependencies. Installing both PyTorch and TorchVision with CUDA support is strongly recommended.
+
+<!-- Install Segment Anything:
+
+```
+pip install git+https://github.com/facebookresearch/segment-anything.git
+```
+
+or clone the repository locally and install with -->
+
+To use Segment Anything with our method, please clone this repository locally and install with
+
+```
+pip install -e .
+```
+
+The following optional dependencies are necessary for mask post-processing, saving masks in COCO format, the example notebooks, and exporting the model in ONNX format. `jupyter` is also required to run the example notebooks.
+```
+pip install opencv-python pycocotools matplotlib onnxruntime onnx
+```
+
+
+## <a name="GettingStarted"></a>Getting Started
+
+You can run the code like using original Segment Anything Model. The only difference is that you need to add `use_hourglass=True` as parameter while calling `build_sam` function. Here is an example.
+
+First download a [model checkpoint](#model-checkpoints). Then the model can be used in just a few lines to get masks from a given prompt:
+
+```
+from segment_anything import build_sam, SamPredictor 
+predictor = SamPredictor(build_sam(checkpoint="</path/to/model.pth>", use_hourglass=True))
+predictor.set_image(<your_image>)
+masks, _, _ = predictor.predict(<input_prompts>)
+```
+
+or generate masks for an entire image:
+
+```
+from segment_anything import build_sam, SamAutomaticMaskGenerator
+mask_generator = SamAutomaticMaskGenerator(build_sam(checkpoint="</path/to/model.pth>", use_hourglass=True))
+masks = mask_generator.generate(<your_image>)
+```
+
+Additionally, masks can be generated for images from the command line:
+
+```
+python scripts/amg.py --checkpoint <path/to/sam/checkpoint> --input <image_or_folder> --output <output_directory> --use_hourglass
+```
+
+You need to add `--use_hourglass` if you want to use our method to accelerate the process.
+
+
+## <a name="Models"></a>Model Checkpoints
+
+<!-- Three model versions of the model are available with different backbone sizes. These models can be instantiated by running 
+```
+from segment_anything import sam_model_registry
+sam = sam_model_registry["<name>"](checkpoint="<path/to/checkpoint>")
+```
+Click the links below to download the checkpoint for the corresponding model name. The default model in bold can also be instantiated with `build_sam`, as in the examples in [Getting Started](#getting-started). -->
+
+Here are the official weight of SAM model.
+
+* **`default` or `vit_h`: [ViT-H SAM model.](https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth)**
+* `vit_l`: [ViT-L SAM model.](https://dl.fbaipublicfiles.com/segment_anything/sam_vit_l_0b3195.pth)
+* `vit_b`: [ViT-B SAM model.](https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth)
+
+## License
+The model is licensed under the [Apache 2.0 license](LICENSE).
+
+## Citation
+
+If you find this repo useful in your research, please consider citing:
+
+```latex
+@article{liang2022expediting,
+	author    = {Liang, Weicong and Yuan, Yuhui and Ding, Henghui and Luo, Xiao and Lin, Weihong and Jia, Ding and Zhang, Zheng and Zhang, Chao and Hu, Han},
+	title     = {Expediting large-scale vision transformer for dense prediction without fine-tuning},
+	journal   = {arXiv preprint arXiv:2210.01035},
+	year      = {2022},
+}
+```
+
+If you use SAM or SA-1B in your research, please use the following BibTeX entry. 
+
+```
+@article{kirillov2023segany,
+  title={Segment Anything}, 
+  author={Kirillov, Alexander and Mintun, Eric and Ravi, Nikhila and Mao, Hanzi and Rolland, Chloe and Gustafson, Laura and Xiao, Tete and Whitehead, Spencer and Berg, Alexander C. and Lo, Wan-Yen and Doll{\'a}r, Piotr and Girshick, Ross},
+  journal={arXiv:2304.02643},
+  year={2023}
+}
+```

app.py

@@ -0,0 +1,94 @@
+import os
+import torch
+import numpy as np
+
+import gradio as gr
+
+from segment_anything import build_sam, SamAutomaticMaskGenerator
+
+os.system(r'python -m wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth')
+
+hourglass_args = {
+    "baseline": {},
+    "1.2x faster": {
+        "use_hourglass": True,
+        "hourglass_clustering_location": 14,
+        "hourglass_num_cluster": 100,
+    },
+    "1.5x faster": {
+        "use_hourglass": True,
+        "hourglass_clustering_location": 6,
+        "hourglass_num_cluster": 81,
+    },
+}
+
+def predict(image, speed_mode):
+    mask_generator = SamAutomaticMaskGenerator(build_sam(checkpoint="sam_vit_h_4b8939.pth", hourglass_kwargs=hourglass_args[speed_mode]))
+    masks = mask_generator.generate(image)
+
+    if len(masks) == 0:
+        return image
+    sorted_masks = sorted(masks, key=(lambda x: x['area']), reverse=True)
+    img = np.ones(image.shape)
+    for mask in sorted_masks:
+        m = mask['segmentation']
+        color_mask = np.random.random((1, 1, 3))
+        img = img * (1 - m[..., None]) + color_mask * m[..., None]
+
+    image = ((image + img * 255) / 2).astype(np.uint8)
+    return image
+
+description = """
+#  <center>Expedit-SAM (Expedite Segment Anything Model without any training)</center>
+Github link: [Link](https://github.com/Expedit-LargeScale-Vision-Transformer/Expedit-SAM)
+You can select the speed mode you want to use from the "Speed Mode" dropdown menu and click "Run" to segment the image you uploaded to the "Input Image" box.
+"""
+if (SPACE_ID := os.getenv('SPACE_ID')) is not None:
+    description += f'\n<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. <a href="https://huggingface.co/spaces/{SPACE_ID}?duplicate=true"><img style="display: inline; margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space" /></a></p>'
+
+
+def main():
+    with gr.Blocks() as demo:
+        gr.Markdown(description)
+        with gr.Column():
+            with gr.Row():
+                with gr.Column():
+                    input_image = gr.Image(label="Input Image")
+                    speed_mode = gr.Dropdown(
+                        choices=list(hourglass_args.keys()),
+                        value="baseline", 
+                        label="Speed Mode",
+                        multiselect=False,
+                    )
+                    with gr.Row():
+                        run_btn = gr.Button(label="Run", id="run", value="Run")
+                        clear_btn = gr.Button(label="Clear", id="clear", value="Clear")
+                output_image = gr.Image(label="Output Image")
+            gr.Examples(
+                examples=[
+                    ["./notebooks/images/dog.jpg"],
+                    ["notebooks/images/groceries.jpg"],
+                    ["notebooks/images/truck.jpg"],
+                ],
+                inputs=[input_image],
+                outputs=[output_image],
+                fn=predict,
+            )
+        
+        run_btn.click(
+            fn=predict, 
+            inputs=[input_image, speed_mode], 
+            outputs=output_image
+        )
+        clear_btn.click(
+            fn=lambda: [None, None], 
+            inputs=None, 
+            outputs=[input_image, output_image], 
+            queue=False,
+        )
+
+    demo.queue()
+    demo.launch()
+
+if __name__ == "__main__":
+    main()

linter.sh

@@ -0,0 +1,32 @@
+#!/bin/bash -e
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+{
+  black --version | grep -E "23\." > /dev/null
+} || {
+  echo "Linter requires 'black==23.*' !"
+  exit 1
+}
+
+ISORT_VERSION=$(isort --version-number)
+if [[ "$ISORT_VERSION" != 5.12* ]]; then
+  echo "Linter requires isort==5.12.0 !"
+  exit 1
+fi
+
+echo "Running isort ..."
+isort . --atomic
+
+echo "Running black ..."
+black -l 100 .
+
+echo "Running flake8 ..."
+if [ -x "$(command -v flake8)" ]; then
+  flake8 .
+else
+  python3 -m flake8 .
+fi
+
+echo "Running mypy..."
+
+mypy --exclude 'setup.py|notebooks' .

notebooks/onnx_model_example.ipynb

@@ -0,0 +1,774 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "901c8ef3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Copyright (c) Meta Platforms, Inc. and affiliates."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "1662bb7c",
+   "metadata": {},
+   "source": [
+    "# Produces masks from prompts using an ONNX model"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7fcc21a0",
+   "metadata": {},
+   "source": [
+    "SAM's prompt encoder and mask decoder are very lightweight, which allows for efficient computation of a mask given user input. This notebook shows an example of how to export and use this lightweight component of the model in ONNX format, allowing it to run on a variety of platforms that support an ONNX runtime."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "86daff77",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "\n",
+       "<a target=\"_blank\" href=\"https://colab.research.google.com/github/facebookresearch/segment-anything/blob/main/notebooks/onnx_model_example.ipynb\">\n",
+       "  <img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/>\n",
+       "</a>\n"
+      ],
+      "text/plain": [
+       "<IPython.core.display.HTML object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from IPython.display import display, HTML\n",
+    "display(HTML(\n",
+    "\"\"\"\n",
+    "<a target=\"_blank\" href=\"https://colab.research.google.com/github/facebookresearch/segment-anything/blob/main/notebooks/onnx_model_example.ipynb\">\n",
+    "  <img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/>\n",
+    "</a>\n",
+    "\"\"\"\n",
+    "))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "55ae4e00",
+   "metadata": {},
+   "source": [
+    "## Environment Set-up"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "109a5cc2",
+   "metadata": {},
+   "source": [
+    "If running locally using jupyter, first install `segment_anything` in your environment using the [installation instructions](https://github.com/facebookresearch/segment-anything#installation) in the repository. The latest stable versions of PyTorch and ONNX are recommended for this notebook. If running from Google Colab, set `using_collab=True` below and run the cell. In Colab, be sure to select 'GPU' under 'Edit'->'Notebook Settings'->'Hardware accelerator'."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "39b99fc4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "using_colab = False"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "296a69be",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "if using_colab:\n",
+    "    import torch\n",
+    "    import torchvision\n",
+    "    print(\"PyTorch version:\", torch.__version__)\n",
+    "    print(\"Torchvision version:\", torchvision.__version__)\n",
+    "    print(\"CUDA is available:\", torch.cuda.is_available())\n",
+    "    import sys\n",
+    "    !{sys.executable} -m pip install opencv-python matplotlib onnx onnxruntime\n",
+    "    !{sys.executable} -m pip install 'git+https://github.com/facebookresearch/segment-anything.git'\n",
+    "    \n",
+    "    !mkdir images\n",
+    "    !wget -P images https://raw.githubusercontent.com/facebookresearch/segment-anything/main/notebooks/images/truck.jpg\n",
+    "        \n",
+    "    !wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "dc4a58be",
+   "metadata": {},
+   "source": [
+    "## Set-up"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "42396e8d",
+   "metadata": {},
+   "source": [
+    "Note that this notebook requires both the `onnx` and `onnxruntime` optional dependencies, in addition to `opencv-python` and `matplotlib` for visualization."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2c712610",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import numpy as np\n",
+    "import cv2\n",
+    "import matplotlib.pyplot as plt\n",
+    "from segment_anything import sam_model_registry, SamPredictor\n",
+    "from segment_anything.utils.onnx import SamOnnxModel\n",
+    "\n",
+    "import onnxruntime\n",
+    "from onnxruntime.quantization import QuantType\n",
+    "from onnxruntime.quantization.quantize import quantize_dynamic"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f29441b9",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def show_mask(mask, ax):\n",
+    "    color = np.array([30/255, 144/255, 255/255, 0.6])\n",
+    "    h, w = mask.shape[-2:]\n",
+    "    mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1)\n",
+    "    ax.imshow(mask_image)\n",
+    "    \n",
+    "def show_points(coords, labels, ax, marker_size=375):\n",
+    "    pos_points = coords[labels==1]\n",
+    "    neg_points = coords[labels==0]\n",
+    "    ax.scatter(pos_points[:, 0], pos_points[:, 1], color='green', marker='*', s=marker_size, edgecolor='white', linewidth=1.25)\n",
+    "    ax.scatter(neg_points[:, 0], neg_points[:, 1], color='red', marker='*', s=marker_size, edgecolor='white', linewidth=1.25)   \n",
+    "    \n",
+    "def show_box(box, ax):\n",
+    "    x0, y0 = box[0], box[1]\n",
+    "    w, h = box[2] - box[0], box[3] - box[1]\n",
+    "    ax.add_patch(plt.Rectangle((x0, y0), w, h, edgecolor='green', facecolor=(0,0,0,0), lw=2))   "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "bd0f6b2b",
+   "metadata": {},
+   "source": [
+    "## Export an ONNX model"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "1540f719",
+   "metadata": {},
+   "source": [
+    "Set the path below to a SAM model checkpoint, then load the model. This will be needed to both export the model and to calculate embeddings for the model."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "76fc53f4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "checkpoint = \"sam_vit_h_4b8939.pth\"\n",
+    "model_type = \"default\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "11bfc8aa",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "sam = sam_model_registry[model_type](checkpoint=checkpoint)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "450c089c",
+   "metadata": {},
+   "source": [
+    "The script `segment-anything/scripts/export_onnx_model.py` can be used to export the necessary portion of SAM. Alternatively, run the following code to export an ONNX model. If you have already exported a model, set the path below and skip to the next section. Assure that the exported ONNX model aligns with the checkpoint and model type set above. This notebook expects the model was exported with the parameter `return_single_mask=True`."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "38a8add8",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_model_path = None  # Set to use an already exported model, then skip to the next section."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7da638ba",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [],
+   "source": [
+    "import warnings\n",
+    "\n",
+    "onnx_model_path = \"sam_onnx_example.onnx\"\n",
+    "\n",
+    "onnx_model = SamOnnxModel(sam, return_single_mask=True)\n",
+    "\n",
+    "dynamic_axes = {\n",
+    "    \"point_coords\": {1: \"num_points\"},\n",
+    "    \"point_labels\": {1: \"num_points\"},\n",
+    "}\n",
+    "\n",
+    "embed_dim = sam.prompt_encoder.embed_dim\n",
+    "embed_size = sam.prompt_encoder.image_embedding_size\n",
+    "mask_input_size = [4 * x for x in embed_size]\n",
+    "dummy_inputs = {\n",
+    "    \"image_embeddings\": torch.randn(1, embed_dim, *embed_size, dtype=torch.float),\n",
+    "    \"point_coords\": torch.randint(low=0, high=1024, size=(1, 5, 2), dtype=torch.float),\n",
+    "    \"point_labels\": torch.randint(low=0, high=4, size=(1, 5), dtype=torch.float),\n",
+    "    \"mask_input\": torch.randn(1, 1, *mask_input_size, dtype=torch.float),\n",
+    "    \"has_mask_input\": torch.tensor([1], dtype=torch.float),\n",
+    "    \"orig_im_size\": torch.tensor([1500, 2250], dtype=torch.float),\n",
+    "}\n",
+    "output_names = [\"masks\", \"iou_predictions\", \"low_res_masks\"]\n",
+    "\n",
+    "with warnings.catch_warnings():\n",
+    "    warnings.filterwarnings(\"ignore\", category=torch.jit.TracerWarning)\n",
+    "    warnings.filterwarnings(\"ignore\", category=UserWarning)\n",
+    "    with open(onnx_model_path, \"wb\") as f:\n",
+    "        torch.onnx.export(\n",
+    "            onnx_model,\n",
+    "            tuple(dummy_inputs.values()),\n",
+    "            f,\n",
+    "            export_params=True,\n",
+    "            verbose=False,\n",
+    "            opset_version=17,\n",
+    "            do_constant_folding=True,\n",
+    "            input_names=list(dummy_inputs.keys()),\n",
+    "            output_names=output_names,\n",
+    "            dynamic_axes=dynamic_axes,\n",
+    "        )    "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "c450cf1a",
+   "metadata": {},
+   "source": [
+    "If desired, the model can additionally be quantized and optimized. We find this improves web runtime significantly for negligible change in qualitative performance. Run the next cell to quantize the model, or skip to the next section otherwise."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "235d39fe",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_model_quantized_path = \"sam_onnx_quantized_example.onnx\"\n",
+    "quantize_dynamic(\n",
+    "    model_input=onnx_model_path,\n",
+    "    model_output=onnx_model_quantized_path,\n",
+    "    optimize_model=True,\n",
+    "    per_channel=False,\n",
+    "    reduce_range=False,\n",
+    "    weight_type=QuantType.QUInt8,\n",
+    ")\n",
+    "onnx_model_path = onnx_model_quantized_path"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "927a928b",
+   "metadata": {},
+   "source": [
+    "## Example Image"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6be6eb55",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image = cv2.imread('images/truck.jpg')\n",
+    "image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b7e9a27a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "plt.figure(figsize=(10,10))\n",
+    "plt.imshow(image)\n",
+    "plt.axis('on')\n",
+    "plt.show()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "027b177b",
+   "metadata": {},
+   "source": [
+    "## Using an ONNX model"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "778d4593",
+   "metadata": {},
+   "source": [
+    "Here as an example, we use `onnxruntime` in python on CPU to execute the ONNX model. However, any platform that supports an ONNX runtime could be used in principle. Launch the runtime session below:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "9689b1bf",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ort_session = onnxruntime.InferenceSession(onnx_model_path)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7708ead6",
+   "metadata": {},
+   "source": [
+    "To use the ONNX model, the image must first be pre-processed using the SAM image encoder. This is a heavier weight process best performed on GPU. SamPredictor can be used as normal, then `.get_image_embedding()` will retreive the intermediate features."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "26e067b4",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "sam.to(device='cuda')\n",
+    "predictor = SamPredictor(sam)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7ad3f0d6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "predictor.set_image(image)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8a6f0f07",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image_embedding = predictor.get_image_embedding().cpu().numpy()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5e112f33",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image_embedding.shape"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6337b654",
+   "metadata": {},
+   "source": [
+    "The ONNX model has a different input signature than `SamPredictor.predict`. The following inputs must all be supplied. Note the special cases for both point and mask inputs. All inputs are `np.float32`.\n",
+    "* `image_embeddings`: The image embedding from `predictor.get_image_embedding()`. Has a batch index of length 1.\n",
+    "* `point_coords`: Coordinates of sparse input prompts, corresponding to both point inputs and box inputs. Boxes are encoded using two points, one for the top-left corner and one for the bottom-right corner. *Coordinates must already be transformed to long-side 1024.* Has a batch index of length 1.\n",
+    "* `point_labels`: Labels for the sparse input prompts. 0 is a negative input point, 1 is a positive input point, 2 is a top-left box corner, 3 is a bottom-right box corner, and -1 is a padding point. *If there is no box input, a single padding point with label -1 and coordinates (0.0, 0.0) should be concatenated.*\n",
+    "* `mask_input`: A mask input to the model with shape 1x1x256x256. This must be supplied even if there is no mask input. In this case, it can just be zeros.\n",
+    "* `has_mask_input`: An indicator for the mask input. 1 indicates a mask input, 0 indicates no mask input.\n",
+    "* `orig_im_size`: The size of the input image in (H,W) format, before any transformation. \n",
+    "\n",
+    "Additionally, the ONNX model does not threshold the output mask logits. To obtain a binary mask, threshold at `sam.mask_threshold` (equal to 0.0)."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "bf5a9f55",
+   "metadata": {},
+   "source": [
+    "### Example point input"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1c0deef0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "input_point = np.array([[500, 375]])\n",
+    "input_label = np.array([1])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7256394c",
+   "metadata": {},
+   "source": [
+    "Add a batch index, concatenate a padding point, and transform."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4f69903e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_coord = np.concatenate([input_point, np.array([[0.0, 0.0]])], axis=0)[None, :, :]\n",
+    "onnx_label = np.concatenate([input_label, np.array([-1])], axis=0)[None, :].astype(np.float32)\n",
+    "\n",
+    "onnx_coord = predictor.transform.apply_coords(onnx_coord, image.shape[:2]).astype(np.float32)\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b188dc53",
+   "metadata": {},
+   "source": [
+    "Create an empty mask input and an indicator for no mask."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "5cb52bcf",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_mask_input = np.zeros((1, 1, 256, 256), dtype=np.float32)\n",
+    "onnx_has_mask_input = np.zeros(1, dtype=np.float32)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a99c2cc5",
+   "metadata": {},
+   "source": [
+    "Package the inputs to run in the onnx model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b1d7ea11",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ort_inputs = {\n",
+    "    \"image_embeddings\": image_embedding,\n",
+    "    \"point_coords\": onnx_coord,\n",
+    "    \"point_labels\": onnx_label,\n",
+    "    \"mask_input\": onnx_mask_input,\n",
+    "    \"has_mask_input\": onnx_has_mask_input,\n",
+    "    \"orig_im_size\": np.array(image.shape[:2], dtype=np.float32)\n",
+    "}"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4b6409c9",
+   "metadata": {},
+   "source": [
+    "Predict a mask and threshold it."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "dc4cc082",
+   "metadata": {
+    "scrolled": false
+   },
+   "outputs": [],
+   "source": [
+    "masks, _, low_res_logits = ort_session.run(None, ort_inputs)\n",
+    "masks = masks > predictor.model.mask_threshold"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "d778a8fb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "masks.shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "badb1175",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "plt.figure(figsize=(10,10))\n",
+    "plt.imshow(image)\n",
+    "show_mask(masks, plt.gca())\n",
+    "show_points(input_point, input_label, plt.gca())\n",
+    "plt.axis('off')\n",
+    "plt.show() "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "1f1d4d15",
+   "metadata": {},
+   "source": [
+    "### Example mask input"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b319da82",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "input_point = np.array([[500, 375], [1125, 625]])\n",
+    "input_label = np.array([1, 1])\n",
+    "\n",
+    "# Use the mask output from the previous run. It is already in the correct form for input to the ONNX model.\n",
+    "onnx_mask_input = low_res_logits"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b1823b37",
+   "metadata": {},
+   "source": [
+    "Transform the points as in the previous example."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8885130f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_coord = np.concatenate([input_point, np.array([[0.0, 0.0]])], axis=0)[None, :, :]\n",
+    "onnx_label = np.concatenate([input_label, np.array([-1])], axis=0)[None, :].astype(np.float32)\n",
+    "\n",
+    "onnx_coord = predictor.transform.apply_coords(onnx_coord, image.shape[:2]).astype(np.float32)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "28e47b69",
+   "metadata": {},
+   "source": [
+    "The `has_mask_input` indicator is now 1."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "3ab4483a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_has_mask_input = np.ones(1, dtype=np.float32)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d3781955",
+   "metadata": {},
+   "source": [
+    "Package inputs, then predict and threshold the mask."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0c1ec096",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ort_inputs = {\n",
+    "    \"image_embeddings\": image_embedding,\n",
+    "    \"point_coords\": onnx_coord,\n",
+    "    \"point_labels\": onnx_label,\n",
+    "    \"mask_input\": onnx_mask_input,\n",
+    "    \"has_mask_input\": onnx_has_mask_input,\n",
+    "    \"orig_im_size\": np.array(image.shape[:2], dtype=np.float32)\n",
+    "}\n",
+    "\n",
+    "masks, _, _ = ort_session.run(None, ort_inputs)\n",
+    "masks = masks > predictor.model.mask_threshold"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1e36554b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "plt.figure(figsize=(10,10))\n",
+    "plt.imshow(image)\n",
+    "show_mask(masks, plt.gca())\n",
+    "show_points(input_point, input_label, plt.gca())\n",
+    "plt.axis('off')\n",
+    "plt.show() "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "2ef211d0",
+   "metadata": {},
+   "source": [
+    "### Example box and point input"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "51e58d2e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "input_box = np.array([425, 600, 700, 875])\n",
+    "input_point = np.array([[575, 750]])\n",
+    "input_label = np.array([0])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "6e119dcb",
+   "metadata": {},
+   "source": [
+    "Add a batch index, concatenate a box and point inputs, add the appropriate labels for the box corners, and transform. There is no padding point since the input includes a box input."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "bfbe4911",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_box_coords = input_box.reshape(2, 2)\n",
+    "onnx_box_labels = np.array([2,3])\n",
+    "\n",
+    "onnx_coord = np.concatenate([input_point, onnx_box_coords], axis=0)[None, :, :]\n",
+    "onnx_label = np.concatenate([input_label, onnx_box_labels], axis=0)[None, :].astype(np.float32)\n",
+    "\n",
+    "onnx_coord = predictor.transform.apply_coords(onnx_coord, image.shape[:2]).astype(np.float32)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "65edabd2",
+   "metadata": {},
+   "source": [
+    "Package inputs, then predict and threshold the mask."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2abfba56",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "onnx_mask_input = np.zeros((1, 1, 256, 256), dtype=np.float32)\n",
+    "onnx_has_mask_input = np.zeros(1, dtype=np.float32)\n",
+    "\n",
+    "ort_inputs = {\n",
+    "    \"image_embeddings\": image_embedding,\n",
+    "    \"point_coords\": onnx_coord,\n",
+    "    \"point_labels\": onnx_label,\n",
+    "    \"mask_input\": onnx_mask_input,\n",
+    "    \"has_mask_input\": onnx_has_mask_input,\n",
+    "    \"orig_im_size\": np.array(image.shape[:2], dtype=np.float32)\n",
+    "}\n",
+    "\n",
+    "masks, _, _ = ort_session.run(None, ort_inputs)\n",
+    "masks = masks > predictor.model.mask_threshold"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8301bf33",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "plt.figure(figsize=(10, 10))\n",
+    "plt.imshow(image)\n",
+    "show_mask(masks[0], plt.gca())\n",
+    "show_box(input_box, plt.gca())\n",
+    "show_points(input_point, input_label, plt.gca())\n",
+    "plt.axis('off')\n",
+    "plt.show()"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.10"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

scripts/amg.py

@@ -0,0 +1,335 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import cv2  # type: ignore
+
+from segment_anything import SamAutomaticMaskGenerator, sam_model_registry
+
+import argparse
+import json
+import os
+from typing import Any, Dict, List
+
+import numpy as np
+import matplotlib.pyplot as plt
+import time
+
+parser = argparse.ArgumentParser(
+    description=(
+        "Runs automatic mask generation on an input image or directory of images, "
+        "and outputs masks as either PNGs or COCO-style RLEs. Requires open-cv, "
+        "as well as pycocotools if saving in RLE format."
+    )
+)
+
+parser.add_argument(
+    "--input",
+    type=str,
+    required=True,
+    help="Path to either a single input image or folder of images.",
+)
+
+parser.add_argument(
+    "--output",
+    type=str,
+    required=True,
+    help=(
+        "Path to the directory where masks will be output. Output will be either a folder "
+        "of PNGs per image or a single json with COCO-style masks."
+    ),
+)
+
+parser.add_argument(
+    "--model-type",
+    type=str,
+    default="default",
+    help="The type of model to load, in ['default', 'vit_l', 'vit_b']",
+)
+
+parser.add_argument(
+    "--checkpoint",
+    type=str,
+    required=True,
+    help="The path to the SAM checkpoint to use for mask generation.",
+)
+
+parser.add_argument("--device", type=str, default="cuda", help="The device to run generation on.")
+
+parser.add_argument(
+    "--convert-to-rle",
+    action="store_true",
+    help=(
+        "Save masks as COCO RLEs in a single json instead of as a folder of PNGs. "
+        "Requires pycocotools."
+    ),
+)
+
+amg_settings = parser.add_argument_group("AMG Settings")
+
+amg_settings.add_argument(
+    "--points-per-side",
+    type=int,
+    default=None,
+    help="Generate masks by sampling a grid over the image with this many points to a side.",
+)
+
+amg_settings.add_argument(
+    "--points-per-batch",
+    type=int,
+    default=None,
+    help="How many input points to process simultaneously in one batch.",
+)
+
+amg_settings.add_argument(
+    "--pred-iou-thresh",
+    type=float,
+    default=None,
+    help="Exclude masks with a predicted score from the model that is lower than this threshold.",
+)
+
+amg_settings.add_argument(
+    "--stability-score-thresh",
+    type=float,
+    default=None,
+    help="Exclude masks with a stability score lower than this threshold.",
+)
+
+amg_settings.add_argument(
+    "--stability-score-offset",
+    type=float,
+    default=None,
+    help="Larger values perturb the mask more when measuring stability score.",
+)
+
+amg_settings.add_argument(
+    "--box-nms-thresh",
+    type=float,
+    default=None,
+    help="The overlap threshold for excluding a duplicate mask.",
+)
+
+amg_settings.add_argument(
+    "--crop-n-layers",
+    type=int,
+    default=None,
+    help=(
+        "If >0, mask generation is run on smaller crops of the image to generate more masks. "
+        "The value sets how many different scales to crop at."
+    ),
+)
+
+amg_settings.add_argument(
+    "--crop-nms-thresh",
+    type=float,
+    default=None,
+    help="The overlap threshold for excluding duplicate masks across different crops.",
+)
+
+amg_settings.add_argument(
+    "--crop-overlap-ratio",
+    type=int,
+    default=None,
+    help="Larger numbers mean image crops will overlap more.",
+)
+
+amg_settings.add_argument(
+    "--crop-n-points-downscale-factor",
+    type=int,
+    default=None,
+    help="The number of points-per-side in each layer of crop is reduced by this factor.",
+)
+
+amg_settings.add_argument(
+    "--min-mask-region-area",
+    type=int,
+    default=None,
+    help=(
+        "Disconnected mask regions or holes with area smaller than this value "
+        "in pixels are removed by postprocessing."
+    ),
+)
+
+# add hourglass settings
+amg_settings.add_argument(
+    "--use_hourglass",
+    action="store_true",
+    help="Use hourglass method to expedite mask generation.",
+)
+
+amg_settings.add_argument(
+    "--hourglass_clustering_location",
+    type=int,
+    default=6,
+    help="location of clustering, ranging from [0, num of layers of transformer)"
+)
+
+amg_settings.add_argument(
+    "--hourglass_num_cluster",
+    type=int,
+    default=100,
+    help="num of clusters, no more than total number of features"
+)
+
+amg_settings.add_argument(
+    "--hourglass_cluster_iters",
+    type=int,
+    default=5,
+    help="num of iterations in clustering"
+)
+
+amg_settings.add_argument(
+    "--hourglass_temperture",
+    type=float,
+    default=5e-3,
+    help="temperture in clustering and reconstruction"
+)
+
+amg_settings.add_argument(
+    "--hourglass_cluster_window_size",
+    type=int,
+    default=5,
+    help="window size in clustering"
+)
+
+amg_settings.add_argument(
+    "--hourglass_reconstruction_k",
+    type=int,
+    default=20,
+    help="k in token reconstruction layer of hourglass vit"
+)
+
+def write_masks_to_folder(masks: List[Dict[str, Any]], path: str) -> None:
+    header = "id,area,bbox_x0,bbox_y0,bbox_w,bbox_h,point_input_x,point_input_y,predicted_iou,stability_score,crop_box_x0,crop_box_y0,crop_box_w,crop_box_h"  # noqa
+    metadata = [header]
+    for i, mask_data in enumerate(masks):
+        mask = mask_data["segmentation"]
+        filename = f"{i}.png"
+        cv2.imwrite(os.path.join(path, filename), mask * 255)
+        mask_metadata = [
+            str(i),
+            str(mask_data["area"]),
+            *[str(x) for x in mask_data["bbox"]],
+            *[str(x) for x in mask_data["point_coords"][0]],
+            str(mask_data["predicted_iou"]),
+            str(mask_data["stability_score"]),
+            *[str(x) for x in mask_data["crop_box"]],
+        ]
+        row = ",".join(mask_metadata)
+        metadata.append(row)
+    metadata_path = os.path.join(path, "metadata.csv")
+    with open(metadata_path, "w") as f:
+        f.write("\n".join(metadata))
+
+    return
+
+
+def get_amg_kwargs(args):
+    amg_kwargs = {
+        "points_per_side": args.points_per_side,
+        "points_per_batch": args.points_per_batch,
+        "pred_iou_thresh": args.pred_iou_thresh,
+        "stability_score_thresh": args.stability_score_thresh,
+        "stability_score_offset": args.stability_score_offset,
+        "box_nms_thresh": args.box_nms_thresh,
+        "crop_n_layers": args.crop_n_layers,
+        "crop_nms_thresh": args.crop_nms_thresh,
+        "crop_overlap_ratio": args.crop_overlap_ratio,
+        "crop_n_points_downscale_factor": args.crop_n_points_downscale_factor,
+        "min_mask_region_area": args.min_mask_region_area,
+    }
+    amg_kwargs = {k: v for k, v in amg_kwargs.items() if v is not None}
+    return amg_kwargs
+
+
+def get_hourglass_kwargs(args):
+    hourglass_kwargs = {
+        "use_hourglass": args.use_hourglass,
+        "hourglass_clustering_location": args.hourglass_clustering_location,
+        "hourglass_num_cluster": args.hourglass_num_cluster,
+        "hourglass_cluster_iters": args.hourglass_cluster_iters,
+        "hourglass_temperture": args.hourglass_temperture,
+        "hourglass_cluster_window_size": args.hourglass_cluster_window_size,
+        "hourglass_reconstruction_k": args.hourglass_reconstruction_k,
+    }
+    hourglass_kwargs = {k: v for k, v in hourglass_kwargs.items() if v is not None}
+    return hourglass_kwargs
+
+
+def show_anns(anns):
+    if len(anns) == 0:
+        return
+    sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True)
+    ax = plt.gca()
+    ax.set_autoscale_on(False)
+    for ann in sorted_anns:
+        m = ann['segmentation']
+        img = np.ones((m.shape[0], m.shape[1], 3))
+        color_mask = np.random.random((1, 3)).tolist()[0]
+        for i in range(3):
+            img[:,:,i] = color_mask[i]
+        ax.imshow(np.dstack((img, m*0.35)))
+
+
+def main(args: argparse.Namespace) -> None:
+    print("Loading model...")
+    hourglass_kwargs = get_hourglass_kwargs(args)
+    sam = sam_model_registry[args.model_type](checkpoint=args.checkpoint, **hourglass_kwargs)
+    _ = sam.to(device=args.device)
+    output_mode = "coco_rle" if args.convert_to_rle else "binary_mask"
+    amg_kwargs = get_amg_kwargs(args)
+    generator = SamAutomaticMaskGenerator(sam, output_mode=output_mode, **amg_kwargs)
+
+    if not os.path.isdir(args.input):
+        targets = [args.input]
+    else:
+        targets = [
+            f for f in os.listdir(args.input) if not os.path.isdir(os.path.join(args.input, f))
+        ]
+        targets = [os.path.join(args.input, f) for f in targets]
+
+    os.makedirs(args.output, exist_ok=True)
+
+    plt.figure(figsize=(20,20))
+
+    total_time = 0
+    warmup = 0
+    for i, t in enumerate(targets):
+        print(f"Processing '{t}'...")
+        image = cv2.imread(t)
+        if image is None:
+            print(f"Could not load '{t}' as an image, skipping...")
+            continue
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        start = time.perf_counter()
+        masks = generator.generate(image)
+        eta = time.perf_counter() - start
+        if i > warmup:
+            total_time += eta
+
+        base = os.path.basename(t)
+        base = os.path.splitext(base)[0]
+        save_base = os.path.join(args.output, base)
+        if output_mode == "binary_mask":
+            os.makedirs(save_base, exist_ok=True)
+            write_masks_to_folder(masks, save_base)
+        else:
+            save_file = save_base + ".json"
+            with open(save_file, "w") as f:
+                json.dump(masks, f)
+
+        plt.clf()
+        plt.imshow(image)
+        show_anns(masks)
+        plt.axis('off')
+        plt.savefig(os.path.join(save_base, base + '.png'), bbox_inches='tight', pad_inches=0)
+    print("Done!")
+    print(f"Average time per image: {total_time / (len(targets) - warmup)} seconds")
+
+
+if __name__ == "__main__":
+    args = parser.parse_args()
+    main(args)

scripts/benchmark.py

@@ -0,0 +1,341 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import cv2  # type: ignore
+
+from segment_anything import SamAutomaticMaskGenerator, sam_model_registry
+from segment_anything.utils.amg import (
+    batch_iterator,
+    generate_crop_boxes,
+)
+
+import argparse
+import json
+import os
+from typing import Any, Dict, List
+
+import numpy as np
+import matplotlib.pyplot as plt
+import time
+
+import torch
+from tqdm import tqdm
+
+parser = argparse.ArgumentParser(
+    description=(
+        "Runs automatic mask generation on an input image or directory of images, "
+        "and outputs masks as either PNGs or COCO-style RLEs. Requires open-cv, "
+        "as well as pycocotools if saving in RLE format."
+    )
+)
+
+parser.add_argument(
+    "--input",
+    type=str,
+    required=True,
+    help="Path to either a single input image or folder of images.",
+)
+
+parser.add_argument(
+    "--output",
+    type=str,
+    required=True,
+    help=(
+        "Path to the directory where masks will be output. Output will be either a folder "
+        "of PNGs per image or a single json with COCO-style masks."
+    ),
+)
+
+parser.add_argument(
+    "--model-type",
+    type=str,
+    default="default",
+    help="The type of model to load, in ['default', 'vit_l', 'vit_b']",
+)
+
+parser.add_argument(
+    "--checkpoint",
+    type=str,
+    required=True,
+    help="The path to the SAM checkpoint to use for mask generation.",
+)
+
+parser.add_argument("--device", type=str, default="cuda", help="The device to run generation on.")
+
+parser.add_argument(
+    "--convert-to-rle",
+    action="store_true",
+    help=(
+        "Save masks as COCO RLEs in a single json instead of as a folder of PNGs. "
+        "Requires pycocotools."
+    ),
+)
+
+amg_settings = parser.add_argument_group("AMG Settings")
+
+amg_settings.add_argument(
+    "--points-per-side",
+    type=int,
+    default=None,
+    help="Generate masks by sampling a grid over the image with this many points to a side.",
+)
+
+amg_settings.add_argument(
+    "--points-per-batch",
+    type=int,
+    default=None,
+    help="How many input points to process simultaneously in one batch.",
+)
+
+amg_settings.add_argument(
+    "--pred-iou-thresh",
+    type=float,
+    default=None,
+    help="Exclude masks with a predicted score from the model that is lower than this threshold.",
+)
+
+amg_settings.add_argument(
+    "--stability-score-thresh",
+    type=float,
+    default=None,
+    help="Exclude masks with a stability score lower than this threshold.",
+)
+
+amg_settings.add_argument(
+    "--stability-score-offset",
+    type=float,
+    default=None,
+    help="Larger values perturb the mask more when measuring stability score.",
+)
+
+amg_settings.add_argument(
+    "--box-nms-thresh",
+    type=float,
+    default=None,
+    help="The overlap threshold for excluding a duplicate mask.",
+)
+
+amg_settings.add_argument(
+    "--crop-n-layers",
+    type=int,
+    default=None,
+    help=(
+        "If >0, mask generation is run on smaller crops of the image to generate more masks. "
+        "The value sets how many different scales to crop at."
+    ),
+)
+
+amg_settings.add_argument(
+    "--crop-nms-thresh",
+    type=float,
+    default=None,
+    help="The overlap threshold for excluding duplicate masks across different crops.",
+)
+
+amg_settings.add_argument(
+    "--crop-overlap-ratio",
+    type=int,
+    default=None,
+    help="Larger numbers mean image crops will overlap more.",
+)
+
+amg_settings.add_argument(
+    "--crop-n-points-downscale-factor",
+    type=int,
+    default=None,
+    help="The number of points-per-side in each layer of crop is reduced by this factor.",
+)
+
+amg_settings.add_argument(
+    "--min-mask-region-area",
+    type=int,
+    default=None,
+    help=(
+        "Disconnected mask regions or holes with area smaller than this value "
+        "in pixels are removed by postprocessing."
+    ),
+)
+
+# add hourglass settings
+amg_settings.add_argument(
+    "--use_hourglass",
+    action="store_true",
+    help="Use hourglass method to expedite mask generation.",
+)
+
+amg_settings.add_argument(
+    "--hourglass_clustering_location",
+    type=int,
+    default=6,
+    help="location of clustering, ranging from [0, num of layers of transformer)"
+)
+
+amg_settings.add_argument(
+    "--hourglass_num_cluster",
+    type=int,
+    default=100,
+    help="num of clusters, no more than total number of features"
+)
+
+amg_settings.add_argument(
+    "--hourglass_cluster_iters",
+    type=int,
+    default=5,
+    help="num of iterations in clustering"
+)
+
+amg_settings.add_argument(
+    "--hourglass_temperture",
+    type=float,
+    default=5e-3,
+    help="temperture in clustering and reconstruction"
+)
+
+amg_settings.add_argument(
+    "--hourglass_cluster_window_size",
+    type=int,
+    default=5,
+    help="window size in clustering"
+)
+
+amg_settings.add_argument(
+    "--hourglass_reconstruction_k",
+    type=int,
+    default=20,
+    help="k in token reconstruction layer of hourglass vit"
+)
+
+def write_masks_to_folder(masks: List[Dict[str, Any]], path: str) -> None:
+    header = "id,area,bbox_x0,bbox_y0,bbox_w,bbox_h,point_input_x,point_input_y,predicted_iou,stability_score,crop_box_x0,crop_box_y0,crop_box_w,crop_box_h"  # noqa
+    metadata = [header]
+    for i, mask_data in enumerate(masks):
+        mask = mask_data["segmentation"]
+        filename = f"{i}.png"
+        cv2.imwrite(os.path.join(path, filename), mask * 255)
+        mask_metadata = [
+            str(i),
+            str(mask_data["area"]),
+            *[str(x) for x in mask_data["bbox"]],
+            *[str(x) for x in mask_data["point_coords"][0]],
+            str(mask_data["predicted_iou"]),
+            str(mask_data["stability_score"]),
+            *[str(x) for x in mask_data["crop_box"]],
+        ]
+        row = ",".join(mask_metadata)
+        metadata.append(row)
+    metadata_path = os.path.join(path, "metadata.csv")
+    with open(metadata_path, "w") as f:
+        f.write("\n".join(metadata))
+
+    return
+
+
+def get_amg_kwargs(args):
+    amg_kwargs = {
+        "points_per_side": args.points_per_side,
+        "points_per_batch": args.points_per_batch,
+        "pred_iou_thresh": args.pred_iou_thresh,
+        "stability_score_thresh": args.stability_score_thresh,
+        "stability_score_offset": args.stability_score_offset,
+        "box_nms_thresh": args.box_nms_thresh,
+        "crop_n_layers": args.crop_n_layers,
+        "crop_nms_thresh": args.crop_nms_thresh,
+        "crop_overlap_ratio": args.crop_overlap_ratio,
+        "crop_n_points_downscale_factor": args.crop_n_points_downscale_factor,
+        "min_mask_region_area": args.min_mask_region_area,
+    }
+    amg_kwargs = {k: v for k, v in amg_kwargs.items() if v is not None}
+    return amg_kwargs
+
+
+def get_hourglass_kwargs(args):
+    hourglass_kwargs = {
+        "use_hourglass": args.use_hourglass,
+        "hourglass_clustering_location": args.hourglass_clustering_location,
+        "hourglass_num_cluster": args.hourglass_num_cluster,
+        "hourglass_cluster_iters": args.hourglass_cluster_iters,
+        "hourglass_temperture": args.hourglass_temperture,
+        "hourglass_cluster_window_size": args.hourglass_cluster_window_size,
+        "hourglass_reconstruction_k": args.hourglass_reconstruction_k,
+    }
+    hourglass_kwargs = {k: v for k, v in hourglass_kwargs.items() if v is not None}
+    return hourglass_kwargs
+
+
+def show_anns(anns):
+    if len(anns) == 0:
+        return
+    sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True)
+    ax = plt.gca()
+    ax.set_autoscale_on(False)
+    for ann in sorted_anns:
+        m = ann['segmentation']
+        img = np.ones((m.shape[0], m.shape[1], 3))
+        color_mask = np.random.random((1, 3)).tolist()[0]
+        for i in range(3):
+            img[:,:,i] = color_mask[i]
+        ax.imshow(np.dstack((img, m*0.35)))
+
+
+def main(args: argparse.Namespace) -> None:
+    print("Loading model...")
+    hourglass_kwargs = get_hourglass_kwargs(args)
+    sam = sam_model_registry[args.model_type](checkpoint=args.checkpoint, **hourglass_kwargs)
+    _ = sam.to(device=args.device)
+    output_mode = "coco_rle" if args.convert_to_rle else "binary_mask"
+    amg_kwargs = get_amg_kwargs(args)
+    generator = SamAutomaticMaskGenerator(sam, output_mode=output_mode, **amg_kwargs)
+
+    total_time = 0
+    warmup = 50
+    num_samples = 200
+    for i in tqdm(range(num_samples)):
+        image = np.random.randint(0, 255, size=(1024, 1024, 3), dtype=np.uint8)
+
+        start = time.perf_counter()
+        # masks = generator.generate(image)
+        with torch.no_grad():
+            # mask_data = generator._generate_masks(image)
+            orig_size = image.shape[:2]
+            crop_boxes, layer_idxs = generate_crop_boxes(
+                orig_size, generator.crop_n_layers, generator.crop_overlap_ratio
+            )
+
+            # Iterate over image crops
+            for crop_box, crop_layer_idx in zip(crop_boxes, layer_idxs):
+                # crop_data = generator._process_crop(image, crop_box, layer_idx, orig_size)
+                x0, y0, x1, y1 = crop_box
+                cropped_im = image[y0:y1, x0:x1, :]
+                cropped_im_size = cropped_im.shape[:2]
+                generator.predictor.set_image(cropped_im)
+
+                points_scale = np.array(cropped_im_size)[None, ::-1]
+                points_for_image = generator.point_grids[crop_layer_idx] * points_scale
+
+                for (points,) in batch_iterator(generator.points_per_batch, points_for_image):
+                    # batch_data = generator._process_batch(points, cropped_im_size, crop_box, orig_size)
+                    transformed_points = generator.predictor.transform.apply_coords(points, cropped_im_size)
+                    in_points = torch.as_tensor(transformed_points, device=generator.predictor.device)
+                    in_labels = torch.ones(in_points.shape[0], dtype=torch.int, device=in_points.device)
+                    masks, iou_preds, _ = generator.predictor.predict_torch(
+                        in_points[:, None, :],
+                        in_labels[:, None],
+                        multimask_output=True,
+                        return_logits=True,
+                    )
+                    del masks
+                    del iou_preds
+
+        eta = time.perf_counter() - start
+        if i >= warmup:
+            total_time += eta
+    print("Done!")
+    print(f"Average time per image: {total_time / (num_samples - warmup)} seconds")
+
+
+if __name__ == "__main__":
+    args = parser.parse_args()
+    main(args)

scripts/export_onnx_model.py

@@ -0,0 +1,204 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from segment_anything import build_sam, build_sam_vit_b, build_sam_vit_l
+from segment_anything.utils.onnx import SamOnnxModel
+
+import argparse
+import warnings
+
+try:
+    import onnxruntime  # type: ignore
+
+    onnxruntime_exists = True
+except ImportError:
+    onnxruntime_exists = False
+
+parser = argparse.ArgumentParser(
+    description="Export the SAM prompt encoder and mask decoder to an ONNX model."
+)
+
+parser.add_argument(
+    "--checkpoint", type=str, required=True, help="The path to the SAM model checkpoint."
+)
+
+parser.add_argument(
+    "--output", type=str, required=True, help="The filename to save the ONNX model to."
+)
+
+parser.add_argument(
+    "--model-type",
+    type=str,
+    default="default",
+    help="In ['default', 'vit_b', 'vit_l']. Which type of SAM model to export.",
+)
+
+parser.add_argument(
+    "--return-single-mask",
+    action="store_true",
+    help=(
+        "If true, the exported ONNX model will only return the best mask, "
+        "instead of returning multiple masks. For high resolution images "
+        "this can improve runtime when upscaling masks is expensive."
+    ),
+)
+
+parser.add_argument(
+    "--opset",
+    type=int,
+    default=17,
+    help="The ONNX opset version to use. Must be >=11",
+)
+
+parser.add_argument(
+    "--quantize-out",
+    type=str,
+    default=None,
+    help=(
+        "If set, will quantize the model and save it with this name. "
+        "Quantization is performed with quantize_dynamic from onnxruntime.quantization.quantize."
+    ),
+)
+
+parser.add_argument(
+    "--gelu-approximate",
+    action="store_true",
+    help=(
+        "Replace GELU operations with approximations using tanh. Useful "
+        "for some runtimes that have slow or unimplemented erf ops, used in GELU."
+    ),
+)
+
+parser.add_argument(
+    "--use-stability-score",
+    action="store_true",
+    help=(
+        "Replaces the model's predicted mask quality score with the stability "
+        "score calculated on the low resolution masks using an offset of 1.0. "
+    ),
+)
+
+parser.add_argument(
+    "--return-extra-metrics",
+    action="store_true",
+    help=(
+        "The model will return five results: (masks, scores, stability_scores, "
+        "areas, low_res_logits) instead of the usual three. This can be "
+        "significantly slower for high resolution outputs."
+    ),
+)
+
+
+def run_export(
+    model_type: str,
+    checkpoint: str,
+    output: str,
+    opset: int,
+    return_single_mask: bool,
+    gelu_approximate: bool = False,
+    use_stability_score: bool = False,
+    return_extra_metrics=False,
+):
+    print("Loading model...")
+    if model_type == "vit_b":
+        sam = build_sam_vit_b(checkpoint)
+    elif model_type == "vit_l":
+        sam = build_sam_vit_l(checkpoint)
+    else:
+        sam = build_sam(checkpoint)
+
+    onnx_model = SamOnnxModel(
+        model=sam,
+        return_single_mask=return_single_mask,
+        use_stability_score=use_stability_score,
+        return_extra_metrics=return_extra_metrics,
+    )
+
+    if gelu_approximate:
+        for n, m in onnx_model.named_modules():
+            if isinstance(m, torch.nn.GELU):
+                m.approximate = "tanh"
+
+    dynamic_axes = {
+        "point_coords": {1: "num_points"},
+        "point_labels": {1: "num_points"},
+    }
+
+    embed_dim = sam.prompt_encoder.embed_dim
+    embed_size = sam.prompt_encoder.image_embedding_size
+    mask_input_size = [4 * x for x in embed_size]
+    dummy_inputs = {
+        "image_embeddings": torch.randn(1, embed_dim, *embed_size, dtype=torch.float),
+        "point_coords": torch.randint(low=0, high=1024, size=(1, 5, 2), dtype=torch.float),
+        "point_labels": torch.randint(low=0, high=4, size=(1, 5), dtype=torch.float),
+        "mask_input": torch.randn(1, 1, *mask_input_size, dtype=torch.float),
+        "has_mask_input": torch.tensor([1], dtype=torch.float),
+        "orig_im_size": torch.tensor([1500, 2250], dtype=torch.float),
+    }
+
+    _ = onnx_model(**dummy_inputs)
+
+    output_names = ["masks", "iou_predictions", "low_res_masks"]
+
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=torch.jit.TracerWarning)
+        warnings.filterwarnings("ignore", category=UserWarning)
+        with open(output, "wb") as f:
+            print(f"Exporing onnx model to {output}...")
+            torch.onnx.export(
+                onnx_model,
+                tuple(dummy_inputs.values()),
+                f,
+                export_params=True,
+                verbose=False,
+                opset_version=opset,
+                do_constant_folding=True,
+                input_names=list(dummy_inputs.keys()),
+                output_names=output_names,
+                dynamic_axes=dynamic_axes,
+            )
+
+    if onnxruntime_exists:
+        ort_inputs = {k: to_numpy(v) for k, v in dummy_inputs.items()}
+        ort_session = onnxruntime.InferenceSession(output)
+        _ = ort_session.run(None, ort_inputs)
+        print("Model has successfully been run with ONNXRuntime.")
+
+
+def to_numpy(tensor):
+    return tensor.cpu().numpy()
+
+
+if __name__ == "__main__":
+    args = parser.parse_args()
+    run_export(
+        model_type=args.model_type,
+        checkpoint=args.checkpoint,
+        output=args.output,
+        opset=args.opset,
+        return_single_mask=args.return_single_mask,
+        gelu_approximate=args.gelu_approximate,
+        use_stability_score=args.use_stability_score,
+        return_extra_metrics=args.return_extra_metrics,
+    )
+
+    if args.quantize_out is not None:
+        assert onnxruntime_exists, "onnxruntime is required to quantize the model."
+        from onnxruntime.quantization import QuantType  # type: ignore
+        from onnxruntime.quantization.quantize import quantize_dynamic  # type: ignore
+
+        print(f"Quantizing model and writing to {args.quantize_out}...")
+        quantize_dynamic(
+            model_input=args.output,
+            model_output=args.quantize_out,
+            optimize_model=True,
+            per_channel=False,
+            reduce_range=False,
+            weight_type=QuantType.QUInt8,
+        )
+        print("Done!")

segment_anything/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .build_sam import (
+    build_sam,
+    build_sam_vit_h,
+    build_sam_vit_l,
+    build_sam_vit_b,
+    sam_model_registry,
+)
+from .predictor import SamPredictor
+from .automatic_mask_generator import SamAutomaticMaskGenerator

segment_anything/automatic_mask_generator.py

@@ -0,0 +1,372 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+from torchvision.ops.boxes import batched_nms, box_area  # type: ignore
+
+from typing import Any, Dict, List, Optional, Tuple
+
+from .modeling import Sam
+from .predictor import SamPredictor
+from .utils.amg import (
+    MaskData,
+    area_from_rle,
+    batch_iterator,
+    batched_mask_to_box,
+    box_xyxy_to_xywh,
+    build_all_layer_point_grids,
+    calculate_stability_score,
+    coco_encode_rle,
+    generate_crop_boxes,
+    is_box_near_crop_edge,
+    mask_to_rle_pytorch,
+    remove_small_regions,
+    rle_to_mask,
+    uncrop_boxes_xyxy,
+    uncrop_masks,
+    uncrop_points,
+)
+
+
+class SamAutomaticMaskGenerator:
+    def __init__(
+        self,
+        model: Sam,
+        points_per_side: Optional[int] = 32,
+        points_per_batch: int = 64,
+        pred_iou_thresh: float = 0.88,
+        stability_score_thresh: float = 0.95,
+        stability_score_offset: float = 1.0,
+        box_nms_thresh: float = 0.7,
+        crop_n_layers: int = 0,
+        crop_nms_thresh: float = 0.7,
+        crop_overlap_ratio: float = 512 / 1500,
+        crop_n_points_downscale_factor: int = 1,
+        point_grids: Optional[List[np.ndarray]] = None,
+        min_mask_region_area: int = 0,
+        output_mode: str = "binary_mask",
+    ) -> None:
+        """
+        Using a SAM model, generates masks for the entire image.
+        Generates a grid of point prompts over the image, then filters
+        low quality and duplicate masks. The default settings are chosen
+        for SAM with a ViT-H backbone.
+
+        Arguments:
+          model (Sam): The SAM model to use for mask prediction.
+          points_per_side (int or None): The number of points to be sampled
+            along one side of the image. The total number of points is
+            points_per_side**2. If None, 'point_grids' must provide explicit
+            point sampling.
+          points_per_batch (int): Sets the number of points run simultaneously
+            by the model. Higher numbers may be faster but use more GPU memory.
+          pred_iou_thresh (float): A filtering threshold in [0,1], using the
+            model's predicted mask quality.
+          stability_score_thresh (float): A filtering threshold in [0,1], using
+            the stability of the mask under changes to the cutoff used to binarize
+            the model's mask predictions.
+          stability_score_offset (float): The amount to shift the cutoff when
+            calculated the stability score.
+          box_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks.
+          crops_n_layers (int): If >0, mask prediction will be run again on
+            crops of the image. Sets the number of layers to run, where each
+            layer has 2**i_layer number of image crops.
+          crops_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks between different crops.
+          crop_overlap_ratio (float): Sets the degree to which crops overlap.
+            In the first crop layer, crops will overlap by this fraction of
+            the image length. Later layers with more crops scale down this overlap.
+          crop_n_points_downscale_factor (int): The number of points-per-side
+            sampled in layer n is scaled down by crop_n_points_downscale_factor**n.
+          point_grids (list(np.ndarray) or None): A list over explicit grids
+            of points used for sampling, normalized to [0,1]. The nth grid in the
+            list is used in the nth crop layer. Exclusive with points_per_side.
+          min_mask_region_area (int): If >0, postprocessing will be applied
+            to remove disconnected regions and holes in masks with area smaller
+            than min_mask_region_area. Requires opencv.
+          output_mode (str): The form masks are returned in. Can be 'binary_mask',
+            'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.
+            For large resolutions, 'binary_mask' may consume large amounts of
+            memory.
+        """
+
+        assert (points_per_side is None) != (
+            point_grids is None
+        ), "Exactly one of points_per_side or point_grid must be provided."
+        if points_per_side is not None:
+            self.point_grids = build_all_layer_point_grids(
+                points_per_side,
+                crop_n_layers,
+                crop_n_points_downscale_factor,
+            )
+        elif point_grids is not None:
+            self.point_grids = point_grids
+        else:
+            raise ValueError("Can't have both points_per_side and point_grid be None.")
+
+        assert output_mode in [
+            "binary_mask",
+            "uncompressed_rle",
+            "coco_rle",
+        ], f"Unknown output_mode {output_mode}."
+        if output_mode == "coco_rle":
+            from pycocotools import mask as mask_utils  # type: ignore # noqa: F401
+
+        if min_mask_region_area > 0:
+            import cv2  # type: ignore # noqa: F401
+
+        self.predictor = SamPredictor(model)
+        self.points_per_batch = points_per_batch
+        self.pred_iou_thresh = pred_iou_thresh
+        self.stability_score_thresh = stability_score_thresh
+        self.stability_score_offset = stability_score_offset
+        self.box_nms_thresh = box_nms_thresh
+        self.crop_n_layers = crop_n_layers
+        self.crop_nms_thresh = crop_nms_thresh
+        self.crop_overlap_ratio = crop_overlap_ratio
+        self.crop_n_points_downscale_factor = crop_n_points_downscale_factor
+        self.min_mask_region_area = min_mask_region_area
+        self.output_mode = output_mode
+
+    @torch.no_grad()
+    def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:
+        """
+        Generates masks for the given image.
+
+        Arguments:
+          image (np.ndarray): The image to generate masks for, in HWC uint8 format.
+
+        Returns:
+           list(dict(str, any)): A list over records for masks. Each record is
+             a dict containing the following keys:
+               segmentation (dict(str, any) or np.ndarray): The mask. If
+                 output_mode='binary_mask', is an array of shape HW. Otherwise,
+                 is a dictionary containing the RLE.
+               bbox (list(float)): The box around the mask, in XYWH format.
+               area (int): The area in pixels of the mask.
+               predicted_iou (float): The model's own prediction of the mask's
+                 quality. This is filtered by the pred_iou_thresh parameter.
+               point_coords (list(list(float))): The point coordinates input
+                 to the model to generate this mask.
+               stability_score (float): A measure of the mask's quality. This
+                 is filtered on using the stability_score_thresh parameter.
+               crop_box (list(float)): The crop of the image used to generate
+                 the mask, given in XYWH format.
+        """
+
+        # Generate masks
+        mask_data = self._generate_masks(image)
+
+        # Filter small disconnected regions and holes in masks
+        if self.min_mask_region_area > 0:
+            mask_data = self.postprocess_small_regions(
+                mask_data,
+                self.min_mask_region_area,
+                max(self.box_nms_thresh, self.crop_nms_thresh),
+            )
+
+        # Encode masks
+        if self.output_mode == "coco_rle":
+            mask_data["segmentations"] = [coco_encode_rle(rle) for rle in mask_data["rles"]]
+        elif self.output_mode == "binary_mask":
+            mask_data["segmentations"] = [rle_to_mask(rle) for rle in mask_data["rles"]]
+        else:
+            mask_data["segmentations"] = mask_data["rles"]
+
+        # Write mask records
+        curr_anns = []
+        for idx in range(len(mask_data["segmentations"])):
+            ann = {
+                "segmentation": mask_data["segmentations"][idx],
+                "area": area_from_rle(mask_data["rles"][idx]),
+                "bbox": box_xyxy_to_xywh(mask_data["boxes"][idx]).tolist(),
+                "predicted_iou": mask_data["iou_preds"][idx].item(),
+                "point_coords": [mask_data["points"][idx].tolist()],
+                "stability_score": mask_data["stability_score"][idx].item(),
+                "crop_box": box_xyxy_to_xywh(mask_data["crop_boxes"][idx]).tolist(),
+            }
+            curr_anns.append(ann)
+
+        return curr_anns
+
+    def _generate_masks(self, image: np.ndarray) -> MaskData:
+        orig_size = image.shape[:2]
+        crop_boxes, layer_idxs = generate_crop_boxes(
+            orig_size, self.crop_n_layers, self.crop_overlap_ratio
+        )
+
+        # Iterate over image crops
+        data = MaskData()
+        for crop_box, layer_idx in zip(crop_boxes, layer_idxs):
+            crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)
+            data.cat(crop_data)
+
+        # Remove duplicate masks between crops
+        if len(crop_boxes) > 1:
+            # Prefer masks from smaller crops
+            scores = 1 / box_area(data["crop_boxes"])
+            scores = scores.to(data["boxes"].device)
+            keep_by_nms = batched_nms(
+                data["boxes"].float(),
+                scores,
+                torch.zeros(len(data["boxes"])),  # categories
+                iou_threshold=self.crop_nms_thresh,
+            )
+            data.filter(keep_by_nms)
+
+        data.to_numpy()
+        return data
+
+    def _process_crop(
+        self,
+        image: np.ndarray,
+        crop_box: List[int],
+        crop_layer_idx: int,
+        orig_size: Tuple[int, ...],
+    ) -> MaskData:
+        # Crop the image and calculate embeddings
+        x0, y0, x1, y1 = crop_box
+        cropped_im = image[y0:y1, x0:x1, :]
+        cropped_im_size = cropped_im.shape[:2]
+        self.predictor.set_image(cropped_im)
+
+        # Get points for this crop
+        points_scale = np.array(cropped_im_size)[None, ::-1]
+        points_for_image = self.point_grids[crop_layer_idx] * points_scale
+
+        # Generate masks for this crop in batches
+        data = MaskData()
+        for (points,) in batch_iterator(self.points_per_batch, points_for_image):
+            batch_data = self._process_batch(points, cropped_im_size, crop_box, orig_size)
+            data.cat(batch_data)
+            del batch_data
+        self.predictor.reset_image()
+
+        # Remove duplicates within this crop.
+        keep_by_nms = batched_nms(
+            data["boxes"].float(),
+            data["iou_preds"],
+            torch.zeros(len(data["boxes"])),  # categories
+            iou_threshold=self.box_nms_thresh,
+        )
+        data.filter(keep_by_nms)
+
+        # Return to the original image frame
+        data["boxes"] = uncrop_boxes_xyxy(data["boxes"], crop_box)
+        data["points"] = uncrop_points(data["points"], crop_box)
+        data["crop_boxes"] = torch.tensor([crop_box for _ in range(len(data["rles"]))])
+
+        return data
+
+    def _process_batch(
+        self,
+        points: np.ndarray,
+        im_size: Tuple[int, ...],
+        crop_box: List[int],
+        orig_size: Tuple[int, ...],
+    ) -> MaskData:
+        orig_h, orig_w = orig_size
+
+        # Run model on this batch
+        transformed_points = self.predictor.transform.apply_coords(points, im_size)
+        in_points = torch.as_tensor(transformed_points, device=self.predictor.device)
+        in_labels = torch.ones(in_points.shape[0], dtype=torch.int, device=in_points.device)
+        masks, iou_preds, _ = self.predictor.predict_torch(
+            in_points[:, None, :],
+            in_labels[:, None],
+            multimask_output=True,
+            return_logits=True,
+        )
+
+        # Serialize predictions and store in MaskData
+        data = MaskData(
+            masks=masks.flatten(0, 1),
+            iou_preds=iou_preds.flatten(0, 1),
+            points=torch.as_tensor(points.repeat(masks.shape[1], axis=0)),
+        )
+        del masks
+
+        # Filter by predicted IoU
+        if self.pred_iou_thresh > 0.0:
+            keep_mask = data["iou_preds"] > self.pred_iou_thresh
+            data.filter(keep_mask)
+
+        # Calculate stability score
+        data["stability_score"] = calculate_stability_score(
+            data["masks"], self.predictor.model.mask_threshold, self.stability_score_offset
+        )
+        if self.stability_score_thresh > 0.0:
+            keep_mask = data["stability_score"] >= self.stability_score_thresh
+            data.filter(keep_mask)
+
+        # Threshold masks and calculate boxes
+        data["masks"] = data["masks"] > self.predictor.model.mask_threshold
+        data["boxes"] = batched_mask_to_box(data["masks"])
+
+        # Filter boxes that touch crop boundaries
+        keep_mask = ~is_box_near_crop_edge(data["boxes"], crop_box, [0, 0, orig_w, orig_h])
+        if not torch.all(keep_mask):
+            data.filter(keep_mask)
+
+        # Compress to RLE
+        data["masks"] = uncrop_masks(data["masks"], crop_box, orig_h, orig_w)
+        data["rles"] = mask_to_rle_pytorch(data["masks"])
+        del data["masks"]
+
+        return data
+
+    @staticmethod
+    def postprocess_small_regions(
+        mask_data: MaskData, min_area: int, nms_thresh: float
+    ) -> MaskData:
+        """
+        Removes small disconnected regions and holes in masks, then reruns
+        box NMS to remove any new duplicates.
+
+        Edits mask_data in place.
+
+        Requires open-cv as a dependency.
+        """
+        if len(mask_data["rles"]) == 0:
+            return mask_data
+
+        # Filter small disconnected regions and holes
+        new_masks = []
+        scores = []
+        for rle in mask_data["rles"]:
+            mask = rle_to_mask(rle)
+
+            mask, changed = remove_small_regions(mask, min_area, mode="holes")
+            unchanged = not changed
+            mask, changed = remove_small_regions(mask, min_area, mode="islands")
+            unchanged = unchanged and not changed
+
+            new_masks.append(torch.as_tensor(mask).unsqueeze(0))
+            # Give score=0 to changed masks and score=1 to unchanged masks
+            # so NMS will prefer ones that didn't need postprocessing
+            scores.append(float(unchanged))
+
+        # Recalculate boxes and remove any new duplicates
+        masks = torch.cat(new_masks, dim=0)
+        boxes = batched_mask_to_box(masks)
+        keep_by_nms = batched_nms(
+            boxes.float(),
+            torch.as_tensor(scores),
+            torch.zeros(len(boxes)),  # categories
+            iou_threshold=nms_thresh,
+        )
+
+        # Only recalculate RLEs for masks that have changed
+        for i_mask in keep_by_nms:
+            if scores[i_mask] == 0.0:
+                mask_torch = masks[i_mask].unsqueeze(0)
+                mask_data["rles"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]
+                mask_data["boxes"][i_mask] = boxes[i_mask]  # update res directly
+        mask_data.filter(keep_by_nms)
+
+        return mask_data

segment_anything/build_sam.py

@@ -0,0 +1,131 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from functools import partial
+
+from .modeling import ImageEncoderViT, MaskDecoder, PromptEncoder, Sam, TwoWayTransformer, HourglassImageEncoderViT
+
+
+def build_sam_vit_h(checkpoint=None, **hourglass_kwargs):
+    return _build_sam(
+        encoder_embed_dim=1280,
+        encoder_depth=32,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[7, 15, 23, 31],
+        checkpoint=checkpoint,
+        hourglass_kwargs=hourglass_kwargs,
+    )
+
+
+build_sam = build_sam_vit_h
+
+
+def build_sam_vit_l(checkpoint=None, **hourglass_kwargs):
+    return _build_sam(
+        encoder_embed_dim=1024,
+        encoder_depth=24,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[5, 11, 17, 23],
+        checkpoint=checkpoint,
+        hourglass_kwargs=hourglass_kwargs,
+    )
+
+
+def build_sam_vit_b(checkpoint=None, **hourglass_kwargs):
+    return _build_sam(
+        encoder_embed_dim=768,
+        encoder_depth=12,
+        encoder_num_heads=12,
+        encoder_global_attn_indexes=[2, 5, 8, 11],
+        checkpoint=checkpoint,
+        hourglass_kwargs=hourglass_kwargs,
+    )
+
+
+sam_model_registry = {
+    "default": build_sam,
+    "vit_h": build_sam,
+    "vit_l": build_sam_vit_l,
+    "vit_b": build_sam_vit_b,
+}
+
+
+def _build_sam(
+    encoder_embed_dim,
+    encoder_depth,
+    encoder_num_heads,
+    encoder_global_attn_indexes,
+    checkpoint=None,
+    hourglass_kwargs={},
+):
+    prompt_embed_dim = 256
+    image_size = 1024
+    vit_patch_size = 16
+    image_embedding_size = image_size // vit_patch_size
+    use_hourglass = hourglass_kwargs.pop("use_hourglass", False)
+    if use_hourglass:
+        image_encoder = HourglassImageEncoderViT(
+            depth=encoder_depth,
+            embed_dim=encoder_embed_dim,
+            img_size=image_size,
+            mlp_ratio=4,
+            norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
+            num_heads=encoder_num_heads,
+            patch_size=vit_patch_size,
+            qkv_bias=True,
+            use_rel_pos=True,
+            global_attn_indexes=encoder_global_attn_indexes,
+            window_size=14,
+            out_chans=prompt_embed_dim,
+            **hourglass_kwargs,
+        )
+    else:
+        image_encoder = ImageEncoderViT(
+            depth=encoder_depth,
+            embed_dim=encoder_embed_dim,
+            img_size=image_size,
+            mlp_ratio=4,
+            norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
+            num_heads=encoder_num_heads,
+            patch_size=vit_patch_size,
+            qkv_bias=True,
+            use_rel_pos=True,
+            global_attn_indexes=encoder_global_attn_indexes,
+            window_size=14,
+            out_chans=prompt_embed_dim,
+        )
+
+    sam = Sam(
+        image_encoder=image_encoder,
+        prompt_encoder=PromptEncoder(
+            embed_dim=prompt_embed_dim,
+            image_embedding_size=(image_embedding_size, image_embedding_size),
+            input_image_size=(image_size, image_size),
+            mask_in_chans=16,
+        ),
+        mask_decoder=MaskDecoder(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+        ),
+        pixel_mean=[123.675, 116.28, 103.53],
+        pixel_std=[58.395, 57.12, 57.375],
+    )
+    sam.eval()
+    if checkpoint is not None:
+        with open(checkpoint, "rb") as f:
+            state_dict = torch.load(f)
+        sam.load_state_dict(state_dict)
+    return sam

segment_anything/modeling/__init__.py

@@ -0,0 +1,12 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .sam import Sam
+from .image_encoder import ImageEncoderViT
+from .hourglass_image_encoder import HourglassImageEncoderViT
+from .mask_decoder import MaskDecoder
+from .prompt_encoder import PromptEncoder
+from .transformer import TwoWayTransformer

segment_anything/modeling/common.py

@@ -0,0 +1,43 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+
+from typing import Type
+
+
+class MLPBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        mlp_dim: int,
+        act: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        super().__init__()
+        self.lin1 = nn.Linear(embedding_dim, mlp_dim)
+        self.lin2 = nn.Linear(mlp_dim, embedding_dim)
+        self.act = act()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lin2(self.act(self.lin1(x)))
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x

segment_anything/modeling/hourglass_image_encoder.py

@@ -0,0 +1,418 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+from typing import Optional, Tuple, Type
+
+from .common import LayerNorm2d, MLPBlock
+
+from .image_encoder import (
+    window_partition,
+    window_unpartition,
+    add_decomposed_rel_pos,
+    ImageEncoderViT,
+    Block,
+    Attention,
+)
+
+
+class TokenClusteringBlock(nn.Module):
+    def __init__(self, num_spixels=None, n_iters=5, temperture=0.05, window_size=7):
+        super().__init__()
+        if isinstance(num_spixels, tuple):
+            assert len(num_spixels) == 2
+        elif num_spixels is not None:
+            x = int(math.sqrt(num_spixels))
+            assert x * x == num_spixels
+            num_spixels = (x, x)
+        self.num_spixels = num_spixels
+        self.n_iters = n_iters
+        self.temperture = temperture
+        assert window_size % 2 == 1
+        self.r = window_size // 2
+
+    def calc_init_centroid(self, images, num_spixels_width, num_spixels_height):
+        """
+        calculate initial superpixels
+
+        Args:
+            images: torch.Tensor
+                A Tensor of shape (B, C, H, W)
+            spixels_width: int
+                initial superpixel width
+            spixels_height: int
+                initial superpixel height
+
+        Return:
+            centroids: torch.Tensor
+                A Tensor of shape (B, C, H * W)
+            init_label_map: torch.Tensor
+                A Tensor of shape (B, H * W)
+            num_spixels_width: int
+                A number of superpixels in each column
+            num_spixels_height: int
+                A number of superpixels int each raw
+        """
+        batchsize, channels, height, width = images.shape
+        device = images.device
+
+        centroids = torch.nn.functional.adaptive_avg_pool2d(
+            images, (num_spixels_height, num_spixels_width)
+        )
+
+        with torch.no_grad():
+            num_spixels = num_spixels_width * num_spixels_height
+            labels = (
+                torch.arange(num_spixels, device=device)
+                .reshape(1, 1, *centroids.shape[-2:])
+                .type_as(centroids)
+            )
+            init_label_map = torch.nn.functional.interpolate(
+                labels, size=(height, width), mode="nearest"
+            ).type_as(centroids)
+            init_label_map = init_label_map.repeat(batchsize, 1, 1, 1)
+
+        init_label_map = init_label_map.reshape(batchsize, -1)
+        centroids = centroids.reshape(batchsize, channels, -1)
+
+        return centroids, init_label_map
+
+    def forward(self, pixel_features, num_spixels=None):
+        if num_spixels is None:
+            num_spixels = self.num_spixels
+            assert num_spixels is not None
+        else:
+            if isinstance(num_spixels, tuple):
+                assert len(num_spixels) == 2
+            else:
+                x = int(math.sqrt(num_spixels))
+                assert x * x == num_spixels
+                num_spixels = (x, x)
+        pixel_features = pixel_features.permute(0, 3, 1, 2)
+        num_spixels_height, num_spixels_width = num_spixels
+        num_spixels = num_spixels_width * num_spixels_height
+        spixel_features, init_label_map = self.calc_init_centroid(
+            pixel_features, num_spixels_width, num_spixels_height
+        )
+
+        device = init_label_map.device
+        spixels_number = torch.arange(num_spixels, device=device)[None, :, None]
+        relative_labels_widths = init_label_map[:, None] % num_spixels_width - spixels_number % num_spixels_width
+        relative_labels_heights = torch.div(init_label_map[:, None], num_spixels_width, rounding_mode='trunc') - torch.div(spixels_number, num_spixels_width, rounding_mode='trunc')
+        mask = torch.logical_and(torch.abs(relative_labels_widths) <= self.r, torch.abs(relative_labels_heights) <= self.r)
+        mask_dist = (~mask) * 1e16
+
+        pixel_features = pixel_features.reshape(*pixel_features.shape[:2], -1)  # (B, C, L)
+        permuted_pixel_features = pixel_features.permute(0, 2, 1)       # (B, L, C)
+
+        for _ in range(self.n_iters):
+            dist_matrix = self.pairwise_dist(pixel_features, spixel_features)    # (B, L', L)
+            dist_matrix += mask_dist
+            affinity_matrix = (-dist_matrix * self.temperture).softmax(1)
+            spixel_features = torch.bmm(affinity_matrix.detach(), permuted_pixel_features)
+            spixel_features = spixel_features / affinity_matrix.detach().sum(2, keepdim=True).clamp_(min=1e-16)
+            spixel_features = spixel_features.permute(0, 2, 1)
+        
+        dist_matrix = self.pairwise_dist(pixel_features, spixel_features)
+        hard_labels = torch.argmin(dist_matrix, dim=1)
+
+        B, C, _ = spixel_features.shape
+        spixel_features = spixel_features.permute(0, 2, 1).reshape(B, num_spixels_height, num_spixels_width, C)
+        return spixel_features, hard_labels
+
+    def pairwise_dist(self, f1, f2):
+        return ((f1 * f1).sum(dim=1).unsqueeze(1)
+                + (f2 * f2).sum(dim=1).unsqueeze(2)
+                - 2 * torch.einsum("bcm, bcn -> bmn", f2, f1))
+
+    def extra_repr(self):
+        return f"num_spixels={self.num_spixels}, n_iters={self.n_iters}"
+
+
+def naive_unpool(f_regions, region_indices):
+    _, _, C = f_regions.shape
+    N, L = region_indices.shape
+    index = region_indices.view(N, L, 1).expand(N, L, C)
+    result = f_regions.gather(1, index)
+    return result
+
+
+class State:
+    def __init__(self, unpooling):
+        self.unpooling = unpooling
+        self.__updated = False
+
+    @property
+    def updated(self):
+        return self.__updated
+
+    def get(self, name, default=None):
+        return getattr(self, name, default)
+
+    def update_state(self, **states: dict):
+        self.__updated = True
+        for k, v in states.items():
+            setattr(self, k, v)
+
+    def call(self, input: torch.Tensor):
+        return self.unpooling(input, self)
+
+
+class UnpoolingBase(nn.Module):
+    def forward(self, x, state: State):
+        if not state.updated:
+            return x, False
+        return self._forward(x, state)
+
+    def derive_unpooler(self):
+        return State(self)
+
+
+class NaiveUnpooling(UnpoolingBase):
+    def _forward(self, x, state: State):
+        return naive_unpool(x, state.hard_labels), False
+
+
+class TokenReconstructionBlock(UnpoolingBase):
+    def __init__(self, k=3, temperture=0.05):
+        super().__init__()
+
+        self.k = k
+        self.temperture = temperture
+
+    def _forward(self, x, state: State):
+        feat = state.feat_before_pooling
+        sfeat = state.feat_after_pooling
+        ds = (
+            (feat * feat).sum(dim=2).unsqueeze(2)
+            + (sfeat * sfeat).sum(dim=2).unsqueeze(1)
+            - 2 * torch.einsum("bnc, bmc -> bnm", feat, sfeat)
+        )  # distance between features and super-features
+        ds[ds < 0] = 0
+        weight = torch.exp(-self.temperture * ds)
+        if self.k >= 0:
+            topk, indices = torch.topk(weight, k=self.k, dim=2)
+            mink = torch.min(topk, dim=-1).values
+            mink = mink.unsqueeze(-1).repeat(1, 1, weight.shape[-1])
+            mask = torch.ge(weight, mink)
+            zero = Variable(torch.zeros_like(weight)).cuda()
+            attention = torch.where(mask, weight, zero)
+        attention = F.normalize(attention, dim=2)
+        ret = torch.einsum("bnm, bmc -> bnc", attention, x)
+
+        return ret, False
+
+
+
+class HourglassImageEncoderViT(ImageEncoderViT):
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        out_chans: int = 256,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        global_attn_indexes: Tuple[int, ...] = (),
+        hourglass_clustering_location: int = -1,
+        hourglass_num_cluster: int = None,
+        hourglass_cluster_iters: int = 3,
+        hourglass_temperture: float = 0.1,
+        hourglass_cluster_window_size: int = 12,
+        hourglass_reconstruction_k: int = 36,
+    ) -> None:
+        """
+        Args:
+            img_size (int): Input image size.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_attn_indexes (list): Indexes for blocks using global attention.
+        """
+        super().__init__(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            depth=depth,
+            num_heads=num_heads,
+            mlp_ratio=mlp_ratio,
+            out_chans=out_chans,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            act_layer=act_layer,
+            use_abs_pos=use_abs_pos,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            window_size=window_size,
+            global_attn_indexes=global_attn_indexes,
+        )
+
+        self.window_size = window_size
+        self.ws_new = int(math.sqrt(hourglass_num_cluster))
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            block = HourglassBlock(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=use_rel_pos,
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=(window_size if i < hourglass_clustering_location else self.ws_new) if i not in global_attn_indexes else 0,
+                window_size_ckpt=window_size,
+                input_size=(img_size // patch_size, img_size // patch_size),
+            )
+            self.blocks.append(block)
+
+        self.clustering_location = hourglass_clustering_location
+        self.token_clustering_block = TokenClusteringBlock(
+            num_spixels=hourglass_num_cluster, 
+            n_iters=hourglass_cluster_iters, 
+            temperture=hourglass_temperture, 
+            window_size=hourglass_cluster_window_size,
+        )
+        self.token_reconstruction_block = TokenReconstructionBlock(
+            k=hourglass_reconstruction_k,
+            temperture=hourglass_temperture,
+        )
+
+    def cluster(self, x, reconstructer):
+        # x: B, H, W, C
+        H, W = x.shape[1:3]
+        x, pad_hw = window_partition(x, self.window_size)  # B*Nw, WH, WW, C
+        Bnw, _, _, C = x.shape
+
+        reconstructer.update_state(
+            feat_before_pooling=x.view(-1, self.window_size * self.window_size, C)
+        )
+        x, hard_labels = self.token_clustering_block(x)  # B*H*W, Wh, Ww, C
+        reconstructer.update_state(hard_labels=hard_labels)
+        reconstructer.update_state(feat_after_pooling=x.view(Bnw, -1, C))
+
+        # merge window
+        # Reverse window partition
+        h = pad_hw[0] // self.window_size * x.shape[1]
+        w = pad_hw[1] // self.window_size * x.shape[2]
+        x = window_unpartition(x, self.ws_new, (h, w), (h, w))
+        # out: B, h, w, C
+        return x, pad_hw
+
+    def reconstruct(self, x, H, W, recontructer, pad_hw):
+        # x: B, h, w, C
+        x, _ = window_partition(x, self.ws_new)   # B*Nw, Wh, Ww, C
+        Bnw, _, _, C = x.shape
+        x = x.view(Bnw, -1, C)
+
+        x, _ = recontructer.call(x) # B*Nw, WH*WW, C
+
+        # merge windows
+        x = x.view(-1, self.window_size, self.window_size, C)
+        x = window_unpartition(x, self.window_size, pad_hw, (H, W)) # B, H, W, C
+        return x
+
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+
+        H, W = x.shape[1], x.shape[2]
+        reconstructer = self.token_reconstruction_block.derive_unpooler()
+        reconstructer.update_state(hw_shape=(H, W))
+
+        for i, blk in enumerate(self.blocks):
+            if i == self.clustering_location:
+                x, pad_hw = self.cluster(x, reconstructer)
+            x = blk(x)
+
+        x = self.reconstruct(x, H, W, reconstructer, pad_hw)
+
+        x = self.neck(x.permute(0, 3, 1, 2))
+
+        return x
+
+
+class HourglassBlock(Block):
+    """Transformer blocks with support of window attention and residual propagation blocks"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+        window_size_ckpt: int = 0,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then
+                use global attention.
+            input_size (int or None): Input resolution for calculating the relative positional
+                parameter size.
+        """
+        super(HourglassBlock, self).__init__(
+            dim=dim,
+            num_heads=num_heads,
+            mlp_ratio=mlp_ratio,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            act_layer=act_layer,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            window_size=window_size,
+            input_size=input_size,
+        )
+
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size_ckpt, window_size_ckpt),
+        )

segment_anything/modeling/image_encoder.py

@@ -0,0 +1,395 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Type
+
+from .common import LayerNorm2d, MLPBlock
+
+
+# This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
+class ImageEncoderViT(nn.Module):
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        out_chans: int = 256,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        global_attn_indexes: Tuple[int, ...] = (),
+    ) -> None:
+        """
+        Args:
+            img_size (int): Input image size.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_attn_indexes (list): Indexes for blocks using global attention.
+        """
+        super().__init__()
+        self.img_size = img_size
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+
+        self.pos_embed: Optional[nn.Parameter] = None
+        if use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            self.pos_embed = nn.Parameter(
+                torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim)
+            )
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=use_rel_pos,
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i not in global_attn_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+            )
+            self.blocks.append(block)
+
+        self.neck = nn.Sequential(
+            nn.Conv2d(
+                embed_dim,
+                out_chans,
+                kernel_size=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+            nn.Conv2d(
+                out_chans,
+                out_chans,
+                kernel_size=3,
+                padding=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+
+        for blk in self.blocks:
+            x = blk(x)
+
+        x = self.neck(x.permute(0, 3, 1, 2))
+
+        return x
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention and residual propagation blocks"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then
+                use global attention.
+            input_size (int or None): Input resolution for calculating the relative positional
+                parameter size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+        )
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
+
+        self.window_size = window_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.attn(x)
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool:  If True, add a learnable bias to query, key, value.
+            rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (int or None): Input resolution for calculating the relative positional
+                parameter size.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        self.use_rel_pos = use_rel_pos
+        if self.use_rel_pos:
+            assert (
+                input_size is not None
+            ), "Input size must be provided if using relative positional encoding."
+            # initialize relative positional embeddings
+            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
+            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (3, B, nHead, H * W, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        # q, k, v with shape (B * nHead, H * W, C)
+        q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
+
+        attn = (q * self.scale) @ k.transpose(-2, -1)
+
+        if self.use_rel_pos:
+            attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
+
+        attn = attn.softmax(dim=-1)
+        x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
+        x = self.proj(x)
+
+        return x
+
+
+def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
+) -> torch.Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of
+        query and key sizes.
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def add_decomposed_rel_pos(
+    attn: torch.Tensor,
+    q: torch.Tensor,
+    rel_pos_h: torch.Tensor,
+    rel_pos_w: torch.Tensor,
+    q_size: Tuple[int, int],
+    k_size: Tuple[int, int],
+) -> torch.Tensor:
+    """
+    Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
+    https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py   # noqa B950
+    Args:
+        attn (Tensor): attention map.
+        q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
+        rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
+        rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
+        q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
+        k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
+
+    Returns:
+        attn (Tensor): attention map with added relative positional embeddings.
+    """
+    q_h, q_w = q_size
+    k_h, k_w = k_size
+    Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+    Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
+
+    attn = (
+        attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
+    ).view(B, q_h * q_w, k_h * k_w)
+
+    return attn
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int):  embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x

segment_anything/modeling/mask_decoder.py

@@ -0,0 +1,176 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from typing import List, Tuple, Type
+
+from .common import LayerNorm2d
+
+
+class MaskDecoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a
+        tranformer architecture.
+
+        Arguments:
+          transformer_dim (int): the channel dimension of the transformer
+          transformer (nn.Module): the transformer used to predict masks
+          num_multimask_outputs (int): the number of masks to predict
+            when disambiguating masks
+          activation (nn.Module): the type of activation to use when
+            upscaling masks
+          iou_head_depth (int): the depth of the MLP used to predict
+            mask quality
+          iou_head_hidden_dim (int): the hidden dimension of the MLP
+            used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
+            activation(),
+        )
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [
+                MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
+                for i in range(self.num_mask_tokens)
+            ]
+        )
+
+        self.iou_prediction_head = MLP(
+            transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth
+        )
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Arguments:
+          image_embeddings (torch.Tensor): the embeddings from the image encoder
+          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+          multimask_output (bool): Whether to return multiple masks or a single
+            mask.
+
+        Returns:
+          torch.Tensor: batched predicted masks
+          torch.Tensor: batched predictions of mask quality
+        """
+        masks, iou_pred = self.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=image_pe,
+            sparse_prompt_embeddings=sparse_prompt_embeddings,
+            dense_prompt_embeddings=dense_prompt_embeddings,
+        )
+
+        # Select the correct mask or masks for outptu
+        if multimask_output:
+            mask_slice = slice(1, None)
+        else:
+            mask_slice = slice(0, 1)
+        masks = masks[:, mask_slice, :, :]
+        iou_pred = iou_pred[:, mask_slice]
+
+        # Prepare output
+        return masks, iou_pred
+
+    def predict_masks(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Predicts masks. See 'forward' for more details."""
+        # Concatenate output tokens
+        output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
+        output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
+        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
+
+        # Expand per-image data in batch direction to be per-mask
+        src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
+        src = src + dense_prompt_embeddings
+        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
+        b, c, h, w = src.shape
+
+        # Run the transformer
+        hs, src = self.transformer(src, pos_src, tokens)
+        iou_token_out = hs[:, 0, :]
+        mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]
+
+        # Upscale mask embeddings and predict masks using the mask tokens
+        src = src.transpose(1, 2).view(b, c, h, w)
+        upscaled_embedding = self.output_upscaling(src)
+        hyper_in_list: List[torch.Tensor] = []
+        for i in range(self.num_mask_tokens):
+            hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding.shape
+        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
+
+        # Generate mask quality predictions
+        iou_pred = self.iou_prediction_head(iou_token_out)
+
+        return masks, iou_pred
+
+
+# Lightly adapted from
+# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        sigmoid_output: bool = False,
+    ) -> None:
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
+        )
+        self.sigmoid_output = sigmoid_output
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return x

segment_anything/modeling/prompt_encoder.py

@@ -0,0 +1,214 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+from torch import nn
+
+from typing import Any, Optional, Tuple, Type
+
+from .common import LayerNorm2d
+
+
+class PromptEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        image_embedding_size: Tuple[int, int],
+        input_image_size: Tuple[int, int],
+        mask_in_chans: int,
+        activation: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        """
+        Encodes prompts for input to SAM's mask decoder.
+
+        Arguments:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts,
+        applied to a dense set of points the shape of the image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape
+            1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(
+        self,
+        points: torch.Tensor,
+        labels: torch.Tensor,
+        pad: bool,
+    ) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
+        point_embedding[labels == -1] = 0.0
+        point_embedding[labels == -1] += self.not_a_point_embed.weight
+        point_embedding[labels == 0] += self.point_embeddings[0].weight
+        point_embedding[labels == 1] += self.point_embeddings[1].weight
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        mask_embedding = self.mask_downscaling(masks)
+        return mask_embedding
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """
+        Gets the batch size of the output given the batch size of the input prompts.
+        """
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense
+        embeddings.
+
+        Arguments:
+          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
+            and labels to embed.
+          boxes (torch.Tensor or none): boxes to embed
+          masks (torch.Tensor or none): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape
+            BxNx(embed_dim), where N is determined by the number of input points
+            and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape
+            Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """
+    Positional encoding using random spatial frequencies.
+    """
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer(
+            "positional_encoding_gaussian_matrix",
+            scale * torch.randn((2, num_pos_feats)),
+        )
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(
+        self, coords_input: torch.Tensor, image_size: Tuple[int, int]
+    ) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C

segment_anything/modeling/sam.py

@@ -0,0 +1,174 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from typing import Any, Dict, List, Tuple
+
+from .image_encoder import ImageEncoderViT
+from .mask_decoder import MaskDecoder
+from .prompt_encoder import PromptEncoder
+
+
+class Sam(nn.Module):
+    mask_threshold: float = 0.0
+    image_format: str = "RGB"
+
+    def __init__(
+        self,
+        image_encoder,
+        prompt_encoder: PromptEncoder,
+        mask_decoder: MaskDecoder,
+        pixel_mean: List[float] = [123.675, 116.28, 103.53],
+        pixel_std: List[float] = [58.395, 57.12, 57.375],
+    ) -> None:
+        """
+        SAM predicts object masks from an image and input prompts.
+
+        Arguments:
+          image_encoder (ImageEncoderViT): The backbone used to encode the
+            image into image embeddings that allow for efficient mask prediction.
+          prompt_encoder (PromptEncoder): Encodes various types of input prompts.
+          mask_decoder (MaskDecoder): Predicts masks from the image embeddings
+            and encoded prompts.
+          pixel_mean (list(float)): Mean values for normalizing pixels in the input image.
+          pixel_std (list(float)): Std values for normalizing pixels in the input image.
+        """
+        super().__init__()
+        self.image_encoder = image_encoder
+        self.prompt_encoder = prompt_encoder
+        self.mask_decoder = mask_decoder
+        self.register_buffer("pixel_mean", torch.Tensor(pixel_mean).view(-1, 1, 1), False)
+        self.register_buffer("pixel_std", torch.Tensor(pixel_std).view(-1, 1, 1), False)
+
+    @property
+    def device(self) -> Any:
+        return self.pixel_mean.device
+
+    @torch.no_grad()
+    def forward(
+        self,
+        batched_input: List[Dict[str, Any]],
+        multimask_output: bool,
+    ) -> List[Dict[str, torch.Tensor]]:
+        """
+        Predicts masks end-to-end from provided images and prompts.
+        If prompts are not known in advance, using SamPredictor is
+        recommended over calling the model directly.
+
+        Arguments:
+          batched_input (list(dict)): A list over input images, each a
+            dictionary with the following keys. A prompt key can be
+            excluded if it is not present.
+              'image': The image as a torch tensor in 3xHxW format,
+                already transformed for input to the model.
+              'original_size': (tuple(int, int)) The original size of
+                the image before transformation, as (H, W).
+              'point_coords': (torch.Tensor) Batched point prompts for
+                this image, with shape BxNx2. Already transformed to the
+                input frame of the model.
+              'point_labels': (torch.Tensor) Batched labels for point prompts,
+                with shape BxN.
+              'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.
+                Already transformed to the input frame of the model.
+              'mask_inputs': (torch.Tensor) Batched mask inputs to the model,
+                in the form Bx1xHxW.
+          multimask_output (bool): Whether the model should predict multiple
+            disambiguating masks, or return a single mask.
+
+        Returns:
+          (list(dict)): A list over input images, where each element is
+            as dictionary with the following keys.
+              'masks': (torch.Tensor) Batched binary mask predictions,
+                with shape BxCxHxW, where B is the number of input promts,
+                C is determiend by multimask_output, and (H, W) is the
+                original size of the image.
+              'iou_predictions': (torch.Tensor) The model's predictions
+                of mask quality, in shape BxC.
+              'low_res_logits': (torch.Tensor) Low resolution logits with
+                shape BxCxHxW, where H=W=256. Can be passed as mask input
+                to subsequent iterations of prediction.
+        """
+        input_images = torch.stack([self.preprocess(x["image"]) for x in batched_input], dim=0)
+        image_embeddings = self.image_encoder(input_images)
+
+        outputs = []
+        for image_record, curr_embedding in zip(batched_input, image_embeddings):
+            if "point_coords" in image_record:
+                points = (image_record["point_coords"], image_record["point_labels"])
+            else:
+                points = None
+            sparse_embeddings, dense_embeddings = self.prompt_encoder(
+                points=points,
+                boxes=image_record.get("boxes", None),
+                masks=image_record.get("mask_inputs", None),
+            )
+            low_res_masks, iou_predictions = self.mask_decoder(
+                image_embeddings=curr_embedding.unsqueeze(0),
+                image_pe=self.prompt_encoder.get_dense_pe(),
+                sparse_prompt_embeddings=sparse_embeddings,
+                dense_prompt_embeddings=dense_embeddings,
+                multimask_output=multimask_output,
+            )
+            masks = self.postprocess_masks(
+                low_res_masks,
+                input_size=image_record["image"].shape[-2:],
+                original_size=image_record["original_size"],
+            )
+            masks = masks > self.mask_threshold
+            outputs.append(
+                {
+                    "masks": masks,
+                    "iou_predictions": iou_predictions,
+                    "low_res_logits": low_res_masks,
+                }
+            )
+        return outputs
+
+    def postprocess_masks(
+        self,
+        masks: torch.Tensor,
+        input_size: Tuple[int, ...],
+        original_size: Tuple[int, ...],
+    ) -> torch.Tensor:
+        """
+        Remove padding and upscale masks to the original image size.
+
+        Arguments:
+          masks (torch.Tensor): Batched masks from the mask_decoder,
+            in BxCxHxW format.
+          input_size (tuple(int, int)): The size of the image input to the
+            model, in (H, W) format. Used to remove padding.
+          original_size (tuple(int, int)): The original size of the image
+            before resizing for input to the model, in (H, W) format.
+
+        Returns:
+          (torch.Tensor): Batched masks in BxCxHxW format, where (H, W)
+            is given by original_size.
+        """
+        masks = F.interpolate(
+            masks,
+            (self.image_encoder.img_size, self.image_encoder.img_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+        masks = masks[..., : input_size[0], : input_size[1]]
+        masks = F.interpolate(masks, original_size, mode="bilinear", align_corners=False)
+        return masks
+
+    def preprocess(self, x: torch.Tensor) -> torch.Tensor:
+        """Normalize pixel values and pad to a square input."""
+        # Normalize colors
+        x = (x - self.pixel_mean) / self.pixel_std
+
+        # Pad
+        h, w = x.shape[-2:]
+        padh = self.image_encoder.img_size - h
+        padw = self.image_encoder.img_size - w
+        x = F.pad(x, (0, padw, 0, padh))
+        return x

segment_anything/modeling/transformer.py

@@ -0,0 +1,240 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from torch import Tensor, nn
+
+import math
+from typing import Tuple, Type
+
+from .common import MLPBlock
+
+
+class TwoWayTransformer(nn.Module):
+    def __init__(
+        self,
+        depth: int,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+    ) -> None:
+        """
+        A transformer decoder that attends to an input image using
+        queries whose positional embedding is supplied.
+
+        Args:
+          depth (int): number of layers in the transformer
+          embedding_dim (int): the channel dimension for the input embeddings
+          num_heads (int): the number of heads for multihead attention. Must
+            divide embedding_dim
+          mlp_dim (int): the channel dimension internal to the MLP block
+          activation (nn.Module): the activation to use in the MLP block
+        """
+        super().__init__()
+        self.depth = depth
+        self.embedding_dim = embedding_dim
+        self.num_heads = num_heads
+        self.mlp_dim = mlp_dim
+        self.layers = nn.ModuleList()
+
+        for i in range(depth):
+            self.layers.append(
+                TwoWayAttentionBlock(
+                    embedding_dim=embedding_dim,
+                    num_heads=num_heads,
+                    mlp_dim=mlp_dim,
+                    activation=activation,
+                    attention_downsample_rate=attention_downsample_rate,
+                    skip_first_layer_pe=(i == 0),
+                )
+            )
+
+        self.final_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm_final_attn = nn.LayerNorm(embedding_dim)
+
+    def forward(
+        self,
+        image_embedding: Tensor,
+        image_pe: Tensor,
+        point_embedding: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          image_embedding (torch.Tensor): image to attend to. Should be shape
+            B x embedding_dim x h x w for any h and w.
+          image_pe (torch.Tensor): the positional encoding to add to the image. Must
+            have the same shape as image_embedding.
+          point_embedding (torch.Tensor): the embedding to add to the query points.
+            Must have shape B x N_points x embedding_dim for any N_points.
+
+        Returns:
+          torch.Tensor: the processed point_embedding
+          torch.Tensor: the processed image_embedding
+        """
+        # BxCxHxW -> BxHWxC == B x N_image_tokens x C
+        bs, c, h, w = image_embedding.shape
+        image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
+        image_pe = image_pe.flatten(2).permute(0, 2, 1)
+
+        # Prepare queries
+        queries = point_embedding
+        keys = image_embedding
+
+        # Apply transformer blocks and final layernorm
+        for layer in self.layers:
+            queries, keys = layer(
+                queries=queries,
+                keys=keys,
+                query_pe=point_embedding,
+                key_pe=image_pe,
+            )
+
+        # Apply the final attenion layer from the points to the image
+        q = queries + point_embedding
+        k = keys + image_pe
+        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm_final_attn(queries)
+
+        return queries, keys
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int = 2048,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+        skip_first_layer_pe: bool = False,
+    ) -> None:
+        """
+        A transformer block with four layers: (1) self-attention of sparse
+        inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
+        block on sparse inputs, and (4) cross attention of dense inputs to sparse
+        inputs.
+
+        Arguments:
+          embedding_dim (int): the channel dimension of the embeddings
+          num_heads (int): the number of heads in the attention layers
+          mlp_dim (int): the hidden dimension of the mlp block
+          activation (nn.Module): the activation of the mlp block
+          skip_first_layer_pe (bool): skip the PE on the first layer
+        """
+        super().__init__()
+        self.self_attn = Attention(embedding_dim, num_heads)
+        self.norm1 = nn.LayerNorm(embedding_dim)
+
+        self.cross_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm2 = nn.LayerNorm(embedding_dim)
+
+        self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)
+        self.norm3 = nn.LayerNorm(embedding_dim)
+
+        self.norm4 = nn.LayerNorm(embedding_dim)
+        self.cross_attn_image_to_token = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+
+        self.skip_first_layer_pe = skip_first_layer_pe
+
+    def forward(
+        self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        # Self attention block
+        if self.skip_first_layer_pe:
+            queries = self.self_attn(q=queries, k=queries, v=queries)
+        else:
+            q = queries + query_pe
+            attn_out = self.self_attn(q=q, k=q, v=queries)
+            queries = queries + attn_out
+        queries = self.norm1(queries)
+
+        # Cross attention block, tokens attending to image embedding
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm2(queries)
+
+        # MLP block
+        mlp_out = self.mlp(queries)
+        queries = queries + mlp_out
+        queries = self.norm3(queries)
+
+        # Cross attention block, image embedding attending to tokens
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
+        keys = keys + attn_out
+        keys = self.norm4(keys)
+
+        return queries, keys
+
+
+class Attention(nn.Module):
+    """
+    An attention layer that allows for downscaling the size of the embedding
+    after projection to queries, keys, and values.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        downsample_rate: int = 1,
+    ) -> None:
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.internal_dim = embedding_dim // downsample_rate
+        self.num_heads = num_heads
+        assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
+
+        self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.k_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.v_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+    def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
+        b, n, c = x.shape
+        x = x.reshape(b, n, num_heads, c // num_heads)
+        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head
+
+    def _recombine_heads(self, x: Tensor) -> Tensor:
+        b, n_heads, n_tokens, c_per_head = x.shape
+        x = x.transpose(1, 2)
+        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        # Attention
+        _, _, _, c_per_head = q.shape
+        attn = q @ k.permute(0, 1, 3, 2)  # B x N_heads x N_tokens x N_tokens
+        attn = attn / math.sqrt(c_per_head)
+        attn = torch.softmax(attn, dim=-1)
+
+        # Get output
+        out = attn @ v
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out

segment_anything/predictor.py

@@ -0,0 +1,269 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+
+from segment_anything.modeling import Sam
+
+from typing import Optional, Tuple
+
+from .utils.transforms import ResizeLongestSide
+
+
+class SamPredictor:
+    def __init__(
+        self,
+        sam_model: Sam,
+    ) -> None:
+        """
+        Uses SAM to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model (Sam): The model to use for mask prediction.
+        """
+        super().__init__()
+        self.model = sam_model
+        self.transform = ResizeLongestSide(sam_model.image_encoder.img_size)
+        self.reset_image()
+
+    def set_image(
+        self,
+        image: np.ndarray,
+        image_format: str = "RGB",
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray): The image for calculating masks. Expects an
+            image in HWC uint8 format, with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        assert image_format in [
+            "RGB",
+            "BGR",
+        ], f"image_format must be in ['RGB', 'BGR'], is {image_format}."
+        if image_format != self.model.image_format:
+            image = image[..., ::-1]
+
+        # Transform the image to the form expected by the model
+        input_image = self.transform.apply_image(image)
+        input_image_torch = torch.as_tensor(input_image, device=self.device)
+        input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
+
+        self.set_torch_image(input_image_torch, image.shape[:2])
+
+    @torch.no_grad()
+    def set_torch_image(
+        self,
+        transformed_image: torch.Tensor,
+        original_image_size: Tuple[int, ...],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method. Expects the input
+        image to be already transformed to the format expected by the model.
+
+        Arguments:
+          transformed_image (torch.Tensor): The input image, with shape
+            1x3xHxW, which has been transformed with ResizeLongestSide.
+          original_image_size (tuple(int, int)): The size of the image
+            before transformation, in (H, W) format.
+        """
+        assert (
+            len(transformed_image.shape) == 4
+            and transformed_image.shape[1] == 3
+            and max(*transformed_image.shape[2:]) == self.model.image_encoder.img_size
+        ), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
+        self.reset_image()
+
+        self.original_size = original_image_size
+        self.input_size = tuple(transformed_image.shape[-2:])
+        input_image = self.model.preprocess(transformed_image)
+        self.features = self.model.image_encoder(input_image)
+        self.is_image_set = True
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self.is_image_set:
+            raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
+
+        # Transform input prompts
+        coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
+        if point_coords is not None:
+            assert (
+                point_labels is not None
+            ), "point_labels must be supplied if point_coords is supplied."
+            point_coords = self.transform.apply_coords(point_coords, self.original_size)
+            coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)
+            labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
+            coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
+        if box is not None:
+            box = self.transform.apply_boxes(box, self.original_size)
+            box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            box_torch = box_torch[None, :]
+        if mask_input is not None:
+            mask_input_torch = torch.as_tensor(mask_input, dtype=torch.float, device=self.device)
+            mask_input_torch = mask_input_torch[None, :, :, :]
+
+        masks, iou_predictions, low_res_masks = self.predict_torch(
+            coords_torch,
+            labels_torch,
+            box_torch,
+            mask_input_torch,
+            multimask_output,
+            return_logits=return_logits,
+        )
+
+        masks = masks[0].detach().cpu().numpy()
+        iou_predictions = iou_predictions[0].detach().cpu().numpy()
+        low_res_masks = low_res_masks[0].detach().cpu().numpy()
+        return masks, iou_predictions, low_res_masks
+
+    @torch.no_grad()
+    def predict_torch(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using ResizeLongestSide.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self.is_image_set:
+            raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
+
+        if point_coords is not None:
+            points = (point_coords, point_labels)
+        else:
+            points = None
+
+        # Embed prompts
+        sparse_embeddings, dense_embeddings = self.model.prompt_encoder(
+            points=points,
+            boxes=boxes,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        low_res_masks, iou_predictions = self.model.mask_decoder(
+            image_embeddings=self.features,
+            image_pe=self.model.prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self.model.postprocess_masks(low_res_masks, self.input_size, self.original_size)
+
+        if not return_logits:
+            masks = masks > self.model.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self.is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) to generate an embedding."
+            )
+        assert self.features is not None, "Features must exist if an image has been set."
+        return self.features
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_image(self) -> None:
+        """Resets the currently set image."""
+        self.is_image_set = False
+        self.features = None
+        self.orig_h = None
+        self.orig_w = None
+        self.input_h = None
+        self.input_w = None

segment_anything/utils/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

segment_anything/utils/amg.py

@@ -0,0 +1,346 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+
+import math
+from copy import deepcopy
+from itertools import product
+from typing import Any, Dict, Generator, ItemsView, List, Tuple
+
+
+class MaskData:
+    """
+    A structure for storing masks and their related data in batched format.
+    Implements basic filtering and concatenation.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        for v in kwargs.values():
+            assert isinstance(
+                v, (list, np.ndarray, torch.Tensor)
+            ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats = dict(**kwargs)
+
+    def __setitem__(self, key: str, item: Any) -> None:
+        assert isinstance(
+            item, (list, np.ndarray, torch.Tensor)
+        ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats[key] = item
+
+    def __delitem__(self, key: str) -> None:
+        del self._stats[key]
+
+    def __getitem__(self, key: str) -> Any:
+        return self._stats[key]
+
+    def items(self) -> ItemsView[str, Any]:
+        return self._stats.items()
+
+    def filter(self, keep: torch.Tensor) -> None:
+        for k, v in self._stats.items():
+            if v is None:
+                self._stats[k] = None
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = v[torch.as_tensor(keep, device=v.device)]
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = v[keep.detach().cpu().numpy()]
+            elif isinstance(v, list) and keep.dtype == torch.bool:
+                self._stats[k] = [a for i, a in enumerate(v) if keep[i]]
+            elif isinstance(v, list):
+                self._stats[k] = [v[i] for i in keep]
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def cat(self, new_stats: "MaskData") -> None:
+        for k, v in new_stats.items():
+            if k not in self._stats or self._stats[k] is None:
+                self._stats[k] = deepcopy(v)
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = torch.cat([self._stats[k], v], dim=0)
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = np.concatenate([self._stats[k], v], axis=0)
+            elif isinstance(v, list):
+                self._stats[k] = self._stats[k] + deepcopy(v)
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def to_numpy(self) -> None:
+        for k, v in self._stats.items():
+            if isinstance(v, torch.Tensor):
+                self._stats[k] = v.detach().cpu().numpy()
+
+
+def is_box_near_crop_edge(
+    boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0
+) -> torch.Tensor:
+    """Filter masks at the edge of a crop, but not at the edge of the original image."""
+    crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)
+    orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)
+    boxes = uncrop_boxes_xyxy(boxes, crop_box).float()
+    near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)
+    near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)
+    near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)
+    return torch.any(near_crop_edge, dim=1)
+
+
+def box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:
+    box_xywh = deepcopy(box_xyxy)
+    box_xywh[2] = box_xywh[2] - box_xywh[0]
+    box_xywh[3] = box_xywh[3] - box_xywh[1]
+    return box_xywh
+
+
+def batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:
+    assert len(args) > 0 and all(
+        len(a) == len(args[0]) for a in args
+    ), "Batched iteration must have inputs of all the same size."
+    n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)
+    for b in range(n_batches):
+        yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]
+
+
+def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
+    """
+    Encodes masks to an uncompressed RLE, in the format expected by
+    pycoco tools.
+    """
+    # Put in fortran order and flatten h,w
+    b, h, w = tensor.shape
+    tensor = tensor.permute(0, 2, 1).flatten(1)
+
+    # Compute change indices
+    diff = tensor[:, 1:] ^ tensor[:, :-1]
+    change_indices = diff.nonzero()
+
+    # Encode run length
+    out = []
+    for i in range(b):
+        cur_idxs = change_indices[change_indices[:, 0] == i, 1]
+        cur_idxs = torch.cat(
+            [
+                torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
+                cur_idxs + 1,
+                torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
+            ]
+        )
+        btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
+        counts = [] if tensor[i, 0] == 0 else [0]
+        counts.extend(btw_idxs.detach().cpu().tolist())
+        out.append({"size": [h, w], "counts": counts})
+    return out
+
+
+def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
+    """Compute a binary mask from an uncompressed RLE."""
+    h, w = rle["size"]
+    mask = np.empty(h * w, dtype=bool)
+    idx = 0
+    parity = False
+    for count in rle["counts"]:
+        mask[idx : idx + count] = parity
+        idx += count
+        parity ^= True
+    mask = mask.reshape(w, h)
+    return mask.transpose()  # Put in C order
+
+
+def area_from_rle(rle: Dict[str, Any]) -> int:
+    return sum(rle["counts"][1::2])
+
+
+def calculate_stability_score(
+    masks: torch.Tensor, mask_threshold: float, threshold_offset: float
+) -> torch.Tensor:
+    """
+    Computes the stability score for a batch of masks. The stability
+    score is the IoU between the binary masks obtained by thresholding
+    the predicted mask logits at high and low values.
+    """
+    # One mask is always contained inside the other.
+    # Save memory by preventing unnecesary cast to torch.int64
+    intersections = (
+        (masks > (mask_threshold + threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    unions = (
+        (masks > (mask_threshold - threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    return intersections / unions
+
+
+def build_point_grid(n_per_side: int) -> np.ndarray:
+    """Generates a 2D grid of points evenly spaced in [0,1]x[0,1]."""
+    offset = 1 / (2 * n_per_side)
+    points_one_side = np.linspace(offset, 1 - offset, n_per_side)
+    points_x = np.tile(points_one_side[None, :], (n_per_side, 1))
+    points_y = np.tile(points_one_side[:, None], (1, n_per_side))
+    points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)
+    return points
+
+
+def build_all_layer_point_grids(
+    n_per_side: int, n_layers: int, scale_per_layer: int
+) -> List[np.ndarray]:
+    """Generates point grids for all crop layers."""
+    points_by_layer = []
+    for i in range(n_layers + 1):
+        n_points = int(n_per_side / (scale_per_layer**i))
+        points_by_layer.append(build_point_grid(n_points))
+    return points_by_layer
+
+
+def generate_crop_boxes(
+    im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float
+) -> Tuple[List[List[int]], List[int]]:
+    """
+    Generates a list of crop boxes of different sizes. Each layer
+    has (2**i)**2 boxes for the ith layer.
+    """
+    crop_boxes, layer_idxs = [], []
+    im_h, im_w = im_size
+    short_side = min(im_h, im_w)
+
+    # Original image
+    crop_boxes.append([0, 0, im_w, im_h])
+    layer_idxs.append(0)
+
+    def crop_len(orig_len, n_crops, overlap):
+        return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))
+
+    for i_layer in range(n_layers):
+        n_crops_per_side = 2 ** (i_layer + 1)
+        overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))
+
+        crop_w = crop_len(im_w, n_crops_per_side, overlap)
+        crop_h = crop_len(im_h, n_crops_per_side, overlap)
+
+        crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]
+        crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]
+
+        # Crops in XYWH format
+        for x0, y0 in product(crop_box_x0, crop_box_y0):
+            box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]
+            crop_boxes.append(box)
+            layer_idxs.append(i_layer + 1)
+
+    return crop_boxes, layer_idxs
+
+
+def uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)
+    # Check if boxes has a channel dimension
+    if len(boxes.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return boxes + offset
+
+
+def uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0]], device=points.device)
+    # Check if points has a channel dimension
+    if len(points.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return points + offset
+
+
+def uncrop_masks(
+    masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int
+) -> torch.Tensor:
+    x0, y0, x1, y1 = crop_box
+    if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:
+        return masks
+    # Coordinate transform masks
+    pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)
+    pad = (x0, pad_x - x0, y0, pad_y - y0)
+    return torch.nn.functional.pad(masks, pad, value=0)
+
+
+def remove_small_regions(
+    mask: np.ndarray, area_thresh: float, mode: str
+) -> Tuple[np.ndarray, bool]:
+    """
+    Removes small disconnected regions and holes in a mask. Returns the
+    mask and an indicator of if the mask has been modified.
+    """
+    import cv2  # type: ignore
+
+    assert mode in ["holes", "islands"]
+    correct_holes = mode == "holes"
+    working_mask = (correct_holes ^ mask).astype(np.uint8)
+    n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)
+    sizes = stats[:, -1][1:]  # Row 0 is background label
+    small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]
+    if len(small_regions) == 0:
+        return mask, False
+    fill_labels = [0] + small_regions
+    if not correct_holes:
+        fill_labels = [i for i in range(n_labels) if i not in fill_labels]
+        # If every region is below threshold, keep largest
+        if len(fill_labels) == 0:
+            fill_labels = [int(np.argmax(sizes)) + 1]
+    mask = np.isin(regions, fill_labels)
+    return mask, True
+
+
+def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
+    from pycocotools import mask as mask_utils  # type: ignore
+
+    h, w = uncompressed_rle["size"]
+    rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
+    rle["counts"] = rle["counts"].decode("utf-8")  # Necessary to serialize with json
+    return rle
+
+
+def batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:
+    """
+    Calculates boxes in XYXY format around masks. Return [0,0,0,0] for
+    an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.
+    """
+    # torch.max below raises an error on empty inputs, just skip in this case
+    if torch.numel(masks) == 0:
+        return torch.zeros(*masks.shape[:-2], 4, device=masks.device)
+
+    # Normalize shape to CxHxW
+    shape = masks.shape
+    h, w = shape[-2:]
+    if len(shape) > 2:
+        masks = masks.flatten(0, -3)
+    else:
+        masks = masks.unsqueeze(0)
+
+    # Get top and bottom edges
+    in_height, _ = torch.max(masks, dim=-1)
+    in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]
+    bottom_edges, _ = torch.max(in_height_coords, dim=-1)
+    in_height_coords = in_height_coords + h * (~in_height)
+    top_edges, _ = torch.min(in_height_coords, dim=-1)
+
+    # Get left and right edges
+    in_width, _ = torch.max(masks, dim=-2)
+    in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]
+    right_edges, _ = torch.max(in_width_coords, dim=-1)
+    in_width_coords = in_width_coords + w * (~in_width)
+    left_edges, _ = torch.min(in_width_coords, dim=-1)
+
+    # If the mask is empty the right edge will be to the left of the left edge.
+    # Replace these boxes with [0, 0, 0, 0]
+    empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)
+    out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)
+    out = out * (~empty_filter).unsqueeze(-1)
+
+    # Return to original shape
+    if len(shape) > 2:
+        out = out.reshape(*shape[:-2], 4)
+    else:
+        out = out[0]
+
+    return out

segment_anything/utils/onnx.py

@@ -0,0 +1,144 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from typing import Tuple
+
+from ..modeling import Sam
+from .amg import calculate_stability_score
+
+
+class SamOnnxModel(nn.Module):
+    """
+    This model should not be called directly, but is used in ONNX export.
+    It combines the prompt encoder, mask decoder, and mask postprocessing of Sam,
+    with some functions modified to enable model tracing. Also supports extra
+    options controlling what information. See the ONNX export script for details.
+    """
+
+    def __init__(
+        self,
+        model: Sam,
+        return_single_mask: bool,
+        use_stability_score: bool = False,
+        return_extra_metrics: bool = False,
+    ) -> None:
+        super().__init__()
+        self.mask_decoder = model.mask_decoder
+        self.model = model
+        self.img_size = model.image_encoder.img_size
+        self.return_single_mask = return_single_mask
+        self.use_stability_score = use_stability_score
+        self.stability_score_offset = 1.0
+        self.return_extra_metrics = return_extra_metrics
+
+    @staticmethod
+    def resize_longest_image_size(
+        input_image_size: torch.Tensor, longest_side: int
+    ) -> torch.Tensor:
+        input_image_size = input_image_size.to(torch.float32)
+        scale = longest_side / torch.max(input_image_size)
+        transformed_size = scale * input_image_size
+        transformed_size = torch.floor(transformed_size + 0.5).to(torch.int64)
+        return transformed_size
+
+    def _embed_points(self, point_coords: torch.Tensor, point_labels: torch.Tensor) -> torch.Tensor:
+        point_coords = point_coords + 0.5
+        point_coords = point_coords / self.img_size
+        point_embedding = self.model.prompt_encoder.pe_layer._pe_encoding(point_coords)
+        point_labels = point_labels.unsqueeze(-1).expand_as(point_embedding)
+
+        point_embedding = point_embedding * (point_labels != -1)
+        point_embedding = point_embedding + self.model.prompt_encoder.not_a_point_embed.weight * (
+            point_labels == -1
+        )
+
+        for i in range(self.model.prompt_encoder.num_point_embeddings):
+            point_embedding = point_embedding + self.model.prompt_encoder.point_embeddings[
+                i
+            ].weight * (point_labels == i)
+
+        return point_embedding
+
+    def _embed_masks(self, input_mask: torch.Tensor, has_mask_input: torch.Tensor) -> torch.Tensor:
+        mask_embedding = has_mask_input * self.model.prompt_encoder.mask_downscaling(input_mask)
+        mask_embedding = mask_embedding + (
+            1 - has_mask_input
+        ) * self.model.prompt_encoder.no_mask_embed.weight.reshape(1, -1, 1, 1)
+        return mask_embedding
+
+    def mask_postprocessing(self, masks: torch.Tensor, orig_im_size: torch.Tensor) -> torch.Tensor:
+        masks = F.interpolate(
+            masks,
+            size=(self.img_size, self.img_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        prepadded_size = self.resize_longest_image_size(orig_im_size, self.img_size)
+        masks = masks[..., : int(prepadded_size[0]), : int(prepadded_size[1])]
+
+        orig_im_size = orig_im_size.to(torch.int64)
+        h, w = orig_im_size[0], orig_im_size[1]
+        masks = F.interpolate(masks, size=(h, w), mode="bilinear", align_corners=False)
+        return masks
+
+    def select_masks(
+        self, masks: torch.Tensor, iou_preds: torch.Tensor, num_points: int
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Determine if we should return the multiclick mask or not from the number of points.
+        # The reweighting is used to avoid control flow.
+        score_reweight = torch.tensor(
+            [[1000] + [0] * (self.model.mask_decoder.num_mask_tokens - 1)]
+        ).to(iou_preds.device)
+        score = iou_preds + (num_points - 2.5) * score_reweight
+        best_idx = torch.argmax(score, dim=1)
+        masks = masks[torch.arange(masks.shape[0]), best_idx, :, :].unsqueeze(1)
+        iou_preds = iou_preds[torch.arange(masks.shape[0]), best_idx].unsqueeze(1)
+
+        return masks, iou_preds
+
+    @torch.no_grad()
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        point_coords: torch.Tensor,
+        point_labels: torch.Tensor,
+        mask_input: torch.Tensor,
+        has_mask_input: torch.Tensor,
+        orig_im_size: torch.Tensor,
+    ):
+        sparse_embedding = self._embed_points(point_coords, point_labels)
+        dense_embedding = self._embed_masks(mask_input, has_mask_input)
+
+        masks, scores = self.model.mask_decoder.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=self.model.prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embedding,
+            dense_prompt_embeddings=dense_embedding,
+        )
+
+        if self.use_stability_score:
+            scores = calculate_stability_score(
+                masks, self.model.mask_threshold, self.stability_score_offset
+            )
+
+        if self.return_single_mask:
+            masks, scores = self.select_masks(masks, scores, point_coords.shape[1])
+
+        upscaled_masks = self.mask_postprocessing(masks, orig_im_size)
+
+        if self.return_extra_metrics:
+            stability_scores = calculate_stability_score(
+                upscaled_masks, self.model.mask_threshold, self.stability_score_offset
+            )
+            areas = (upscaled_masks > self.model.mask_threshold).sum(-1).sum(-1)
+            return upscaled_masks, scores, stability_scores, areas, masks
+
+        return upscaled_masks, scores, masks

segment_anything/utils/transforms.py

@@ -0,0 +1,102 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+from torch.nn import functional as F
+from torchvision.transforms.functional import resize, to_pil_image  # type: ignore
+
+from copy import deepcopy
+from typing import Tuple
+
+
+class ResizeLongestSide:
+    """
+    Resizes images to longest side 'target_length', as well as provides
+    methods for resizing coordinates and boxes. Provides methods for
+    transforming both numpy array and batched torch tensors.
+    """
+
+    def __init__(self, target_length: int) -> None:
+        self.target_length = target_length
+
+    def apply_image(self, image: np.ndarray) -> np.ndarray:
+        """
+        Expects a numpy array with shape HxWxC in uint8 format.
+        """
+        target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
+        return np.array(resize(to_pil_image(image), target_size))
+
+    def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
+        """
+        Expects a numpy array of length 2 in the final dimension. Requires the
+        original image size in (H, W) format.
+        """
+        old_h, old_w = original_size
+        new_h, new_w = self.get_preprocess_shape(
+            original_size[0], original_size[1], self.target_length
+        )
+        coords = deepcopy(coords).astype(float)
+        coords[..., 0] = coords[..., 0] * (new_w / old_w)
+        coords[..., 1] = coords[..., 1] * (new_h / old_h)
+        return coords
+
+    def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
+        """
+        Expects a numpy array shape Bx4. Requires the original image size
+        in (H, W) format.
+        """
+        boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)
+        return boxes.reshape(-1, 4)
+
+    def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:
+        """
+        Expects batched images with shape BxCxHxW and float format. This
+        transformation may not exactly match apply_image. apply_image is
+        the transformation expected by the model.
+        """
+        # Expects an image in BCHW format. May not exactly match apply_image.
+        target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
+        return F.interpolate(
+            image, target_size, mode="bilinear", align_corners=False, antialias=True
+        )
+
+    def apply_coords_torch(
+        self, coords: torch.Tensor, original_size: Tuple[int, ...]
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. Requires the
+        original image size in (H, W) format.
+        """
+        old_h, old_w = original_size
+        new_h, new_w = self.get_preprocess_shape(
+            original_size[0], original_size[1], self.target_length
+        )
+        coords = deepcopy(coords).to(torch.float)
+        coords[..., 0] = coords[..., 0] * (new_w / old_w)
+        coords[..., 1] = coords[..., 1] * (new_h / old_h)
+        return coords
+
+    def apply_boxes_torch(
+        self, boxes: torch.Tensor, original_size: Tuple[int, ...]
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with shape Bx4. Requires the original image
+        size in (H, W) format.
+        """
+        boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)
+        return boxes.reshape(-1, 4)
+
+    @staticmethod
+    def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:
+        """
+        Compute the output size given input size and target long side length.
+        """
+        scale = long_side_length * 1.0 / max(oldh, oldw)
+        newh, neww = oldh * scale, oldw * scale
+        neww = int(neww + 0.5)
+        newh = int(newh + 0.5)
+        return (newh, neww)

setup.cfg

@@ -0,0 +1,11 @@
+[isort]
+line_length=100
+multi_line_output=3
+include_trailing_comma=True
+known_standard_library=numpy,setuptools
+skip_glob=*/__init__.py
+known_myself=segment_anything
+known_third_party=matplotlib,cv2,torch,torchvision,pycocotools,onnx,black,isort
+no_lines_before=STDLIB,THIRDPARTY
+sections=FUTURE,STDLIB,THIRDPARTY,MYSELF,FIRSTPARTY,LOCALFOLDER
+default_section=FIRSTPARTY

setup.py

@@ -0,0 +1,18 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from setuptools import find_packages, setup
+
+setup(
+    name="segment_anything",
+    version="1.0",
+    install_requires=[],
+    packages=find_packages(exclude="notebooks"),
+    extras_require={
+        "all": ["matplotlib", "pycocotools", "opencv-python", "onnx", "onnxruntime"],
+        "dev": ["flake8", "isort", "black", "mypy"],
+    },
+)