Fix installer packaging and demo workflow

.github/workflows/smoke.yml

@@ -23,16 +23,24 @@ jobs:
         run: |
           python -m compileall __init__.py nodes.py install.py
           python - <<'PY'
+          import json
           import tomllib
           from pathlib import Path
 
           pyproject = tomllib.loads(Path("pyproject.toml").read_text(encoding="utf-8"))
           assert pyproject["project"]["name"] == "matanyone2-video-matting"
+          assert pyproject["tool"]["comfy"]["requires-comfyui"] == ">=0.17.2"
           assert Path("README.md").exists()
           assert Path("LICENSE").exists()
           assert Path("assets/registry-banner-20260315b.svg").exists()
           assert Path("assets/demo-showcase-20260315.png").exists()
+          assert Path("assets/matanyone2-demo-input.mp4").exists()
+          assert Path("vendor/MatAnyone2/matanyone2/model/matanyone2.py").exists()
+          assert Path("vendor/segment-anything/segment_anything/predictor.py").exists()
           assert Path("web/matanyone_editor.js").exists()
           assert Path("workflows/matanyone2_demo.json").exists()
+          workflow = json.loads(Path("workflows/matanyone2_demo.json").read_text(encoding="utf-8"))
+          load_video = next(node for node in workflow["nodes"] if node["type"] == "VHS_LoadVideo")
+          assert load_video["widgets_values"]["video"] == "matanyone2-demo-input.mp4"
           print("Smoke validation passed.")
           PY

.gitignore

@@ -5,8 +5,9 @@ __pycache__/
 dist/
 build/
 
-# Vendored repos (cloned by install.py)
-vendor/
+# Bundled vendor sources
+vendor/**/__pycache__/
+vendor/**/*.pyc
 models/
 
 # ComfyUI outputs

README.md

@@ -15,7 +15,8 @@ Interactive video matting inside ComfyUI with a built-in SAM point editor and Ma
 - **Two-pass interactive workflow**: fast first queue to open the editor, full second queue to render the matte.
 - **Built-in SAM overlay**: positive and negative point editing happens directly in ComfyUI.
 - **Multi-target masking**: create and compare multiple mask targets before committing.
-- **Pinned vendor snapshots**: installs pull known-good MatAnyone2 and Segment Anything source snapshots into `vendor/`.
+- **Pinned vendor snapshots**: the package ships known-good MatAnyone2 and Segment Anything source snapshots in `vendor/`.
+- **Bundled starter clip**: the sample workflow ships with a tiny demo video that `install.py` copies into `ComfyUI/input/`.
 - **Manager-ready packaging**: `requirements.txt`, `install.py`, workflow example, and registry metadata are all included.
 
 ## Installation
@@ -29,7 +30,7 @@ Registry page: [registry.comfy.org/nodes/matanyone2-video-matting](https://regis
 Manager will:
 
 1. Install Python dependencies from `requirements.txt`
-2. Run `install.py` to fetch the pinned MatAnyone2 and SAM vendor snapshots
+2. Run `install.py` to validate the bundled vendor snapshots and copy `matanyone2-demo-input.mp4` into `ComfyUI/input/`
 3. Ask for a ComfyUI restart
 
 ### Manual Install
@@ -63,6 +64,8 @@ Models download automatically on first use if they are missing.
 
 `workflows/matanyone2_demo.json` is a ready-to-run example that wires video loading, frame selection, the interactive editor, and final matte export.
 
+It defaults to the bundled `matanyone2-demo-input.mp4` starter clip that `install.py` copies into `ComfyUI/input/`. Swap the `Load Video` node to your own footage whenever you are ready.
+
 ```text
 LoadVideo -> SliceFrames -> SelectFrame -> Interactive SAM -> Matte -> SaveVideo x3
 ```
@@ -70,7 +73,7 @@ LoadVideo -> SliceFrames -> SelectFrame -> Interactive SAM -> Matte -> SaveVideo
 How to use it:
 
 1. Drag `workflows/matanyone2_demo.json` onto the ComfyUI canvas.
-2. Load a source video.
+2. Queue once on the bundled starter clip, or replace `Load Video` with your own clip first.
 3. Queue once to open the editor on the first frame.
 4. Left-click for foreground points and right-click for background points.
 5. Click `Apply` in the editor.
@@ -112,6 +115,8 @@ Nodes are grouped under `MatAnyone2` and `MatAnyone2/SAM` in the Add Node menu.
 
 - If the editor does not open, queue the workflow once so the first-frame preview exists.
 - If the demo workflow shows missing video nodes, install VideoHelperSuite first.
+- If `Load Video` is blank or points at `/input/`, rerun `install.py` so the bundled demo clip is copied into `ComfyUI/input/`, or choose your own video in the VHS node before queueing.
+- If ComfyUI Manager shows a security warning for network access, you are likely on an older registry build from before the vendor sources were bundled directly in the package.
 - If VRAM is tight, use a smaller SAM variant and lower `max_internal_size`.
 - If you already keep checkpoints elsewhere, pass an explicit `checkpoint_path` to the loader nodes.
 

install.py

@@ -1,30 +1,34 @@
 from __future__ import annotations
 
+import hashlib
 import shutil
-import tarfile
-import tempfile
-import urllib.request
 from pathlib import Path
 
 
 ROOT = Path(__file__).resolve().parent
 VENDOR_ROOT = ROOT / "vendor"
 COMMIT_MARKER = ".vendor_commit"
+DEMO_VIDEO_NAME = "matanyone2-demo-input.mp4"
+DEMO_VIDEO_SOURCE = ROOT / "assets" / DEMO_VIDEO_NAME
 
 VENDORS = (
     {
         "name": "MatAnyone2",
         "path": ROOT / "vendor" / "MatAnyone2",
-        "repo": "https://github.com/pq-yang/MatAnyone2",
         "commit": "4aed00afd91e108d8986b062b968678fd1c4f8f7",
-        "archive_url": "https://github.com/pq-yang/MatAnyone2/archive/4aed00afd91e108d8986b062b968678fd1c4f8f7.tar.gz",
+        "required_paths": (
+            "matanyone2/config/eval_matanyone_config.yaml",
+            "matanyone2/model/matanyone2.py",
+        ),
     },
     {
         "name": "segment-anything",
         "path": ROOT / "vendor" / "segment-anything",
-        "repo": "https://github.com/facebookresearch/segment-anything",
         "commit": "dca509fe793f601edb92606367a655c15ac00fdf",
-        "archive_url": "https://github.com/facebookresearch/segment-anything/archive/dca509fe793f601edb92606367a655c15ac00fdf.tar.gz",
+        "required_paths": (
+            "segment_anything/build_sam.py",
+            "segment_anything/predictor.py",
+        ),
     },
 )
 
@@ -36,55 +40,75 @@ def _read_marker(repo_path: Path) -> str | None:
     return marker.read_text(encoding="utf-8").strip() or None
 
 
-def _safe_extract_tar(archive_path: Path, destination: Path) -> Path:
-    with tarfile.open(archive_path, "r:gz") as archive:
-        members = archive.getmembers()
-        top_levels = {member.name.split("/", 1)[0] for member in members if member.name}
-        if len(top_levels) != 1:
-            raise RuntimeError(f"Unexpected archive layout in {archive_path.name}")
+def _detect_input_directory() -> Path | None:
+    try:
+        import folder_paths  # type: ignore
 
-        for member in members:
-            target_path = (destination / member.name).resolve()
-            if destination.resolve() not in target_path.parents and target_path != destination.resolve():
-                raise RuntimeError(f"Unsafe archive entry detected: {member.name}")
+        return Path(folder_paths.get_input_directory()).resolve()
+    except Exception:
+        pass
 
-        try:
-            archive.extractall(destination, filter="data")
-        except TypeError:
-            archive.extractall(destination)
-
-    return destination / next(iter(top_levels))
+    for parent in ROOT.parents:
+        if parent.name == "custom_nodes":
+            return (parent.parent / "input").resolve()
+    return None
 
 
-def _download_file(url: str, destination: Path) -> None:
-    print(f"Downloading {url}")
-    with urllib.request.urlopen(url) as response, destination.open("wb") as output_file:
-        shutil.copyfileobj(response, output_file)
+def _file_sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as file:
+        for chunk in iter(lambda: file.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
 
 
-def ensure_vendor_checkout(vendor: dict[str, str | Path]) -> None:
-    repo_path = Path(vendor["path"])
-    commit = str(vendor["commit"])
-    archive_url = str(vendor["archive_url"])
-
-    if repo_path.exists() and _read_marker(repo_path) == commit:
-        print(f"Using pinned {vendor['name']} snapshot at {repo_path}")
+def ensure_demo_input_video() -> None:
+    if not DEMO_VIDEO_SOURCE.exists():
+        print(f"Bundled demo clip missing at {DEMO_VIDEO_SOURCE}, skipping input copy.")
         return
 
-    VENDOR_ROOT.mkdir(parents=True, exist_ok=True)
-    with tempfile.TemporaryDirectory(prefix="matanyone_vendor_") as temp_dir_raw:
-        temp_dir = Path(temp_dir_raw)
-        archive_path = temp_dir / f"{vendor['name']}.tar.gz"
-        _download_file(archive_url, archive_path)
-        extracted_root = _safe_extract_tar(archive_path, temp_dir)
+    input_dir = _detect_input_directory()
+    if input_dir is None:
+        print("Could not locate ComfyUI input directory, skipping demo clip copy.")
+        return
 
-        if repo_path.exists():
-            shutil.rmtree(repo_path)
-        shutil.copytree(extracted_root, repo_path)
-        (repo_path / COMMIT_MARKER).write_text(f"{commit}\n", encoding="utf-8")
-        print(f"Installed pinned {vendor['name']} snapshot at {repo_path}")
+    input_dir.mkdir(parents=True, exist_ok=True)
+    destination = input_dir / DEMO_VIDEO_NAME
+    if destination.exists():
+        if _file_sha256(destination) == _file_sha256(DEMO_VIDEO_SOURCE):
+            print(f"Using bundled demo clip at {destination}")
+            return
+        print(f"Refreshing bundled demo clip at {destination}")
+    else:
+        print(f"Copying bundled demo clip to {destination}")
+
+    shutil.copy2(DEMO_VIDEO_SOURCE, destination)
+
+
+def ensure_vendor_bundle(vendor: dict[str, str | Path | tuple[str, ...]]) -> None:
+    repo_path = Path(vendor["path"])
+    commit = str(vendor["commit"])
+    required_paths = tuple(str(path) for path in vendor["required_paths"])
+
+    missing_paths = [str(repo_path / relative_path) for relative_path in required_paths if not (repo_path / relative_path).exists()]
+    if missing_paths:
+        missing = ", ".join(missing_paths)
+        raise RuntimeError(
+            f"Bundled {vendor['name']} sources are incomplete. Missing: {missing}. "
+            "Reinstall the node package from the registry or GitHub."
+        )
+
+    marker = _read_marker(repo_path)
+    if marker != commit:
+        raise RuntimeError(
+            f"Bundled {vendor['name']} sources do not match the pinned commit. "
+            f"Expected {commit}, found {marker or 'missing marker'}."
+        )
+
+    print(f"Using bundled {vendor['name']} sources at {repo_path}")
 
 
 if __name__ == "__main__":
     for vendor in VENDORS:
-        ensure_vendor_checkout(vendor)
+        ensure_vendor_bundle(vendor)
+    ensure_demo_input_video()

nodes.py

@@ -738,6 +738,8 @@ def _close_interactive_sam_session(session_id: str) -> None:
 
 
 class MatAnyoneModelLoader:
+    DESCRIPTION = "Loads a MatAnyone or MatAnyone2 checkpoint and auto-downloads it when no path is provided."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -767,6 +769,8 @@ class MatAnyoneModelLoader:
 
 
 class MatAnyoneSAMLoader:
+    DESCRIPTION = "Loads a Segment Anything checkpoint for scripted or reusable SAM refinement."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -796,6 +800,8 @@ class MatAnyoneSAMLoader:
 
 
 class MatAnyoneSliceFrames:
+    DESCRIPTION = "Slices an IMAGE batch down to the frame range you want to edit or matte."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -825,6 +831,8 @@ class MatAnyoneSliceFrames:
 
 
 class MatAnyoneSelectFrame:
+    DESCRIPTION = "Extracts one frame from an IMAGE batch for interactive masking or prompt setup."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -848,6 +856,8 @@ class MatAnyoneSelectFrame:
 
 
 class MatAnyonePromptStart:
+    DESCRIPTION = "Creates an empty SAM prompt container for point-based edits."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {"required": {}}
@@ -862,6 +872,8 @@ class MatAnyonePromptStart:
 
 
 class MatAnyonePromptFromText:
+    DESCRIPTION = "Parses x,y,+/- lines into a reusable SAM prompt structure."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -887,6 +899,8 @@ class MatAnyonePromptFromText:
 
 
 class MatAnyoneAddPoint:
+    DESCRIPTION = "Adds one positive or negative point to an existing SAM prompt."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -911,6 +925,8 @@ class MatAnyoneAddPoint:
 
 
 class MatAnyoneSAMRefine:
+    DESCRIPTION = "Runs SAM against a frame and prompt, returning the chosen mask, preview, logits, and score."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -984,6 +1000,8 @@ class MatAnyoneSAMRefine:
 
 
 class MatAnyoneMergeMasks:
+    DESCRIPTION = "Combines up to four masks into one thresholded output mask."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -1015,6 +1033,8 @@ class MatAnyoneMergeMasks:
 
 
 class MatAnyonePreviewMasks:
+    DESCRIPTION = "Draws one or more mask overlays on the source image for quick inspection."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -1055,6 +1075,8 @@ class MatAnyonePreviewMasks:
 
 
 class MatAnyoneInteractiveSAM:
+    DESCRIPTION = "Opens the built-in multi-target SAM editor and returns the merged first-frame mask."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {
@@ -1153,6 +1175,8 @@ class MatAnyoneInteractiveSAM:
 
 
 class MatAnyoneMatte:
+    DESCRIPTION = "Propagates the first-frame mask through the clip and returns foreground, alpha, and preview outputs."
+
     @classmethod
     def INPUT_TYPES(cls):
         return {

pyproject.toml

@@ -30,4 +30,5 @@ Documentation = "https://github.com/dreamrec/MatAnyone2_ComfyUI#readme"
 [tool.comfy]
 PublisherId = "dreamrec"
 DisplayName = "MatAnyone2 Video Matting"
+requires-comfyui = ">=0.17.2"
 Banner = "https://raw.githubusercontent.com/dreamrec/MatAnyone2_ComfyUI/main/assets/registry-banner-20260315b.svg"

vendor/MatAnyone2/.vendor_commit

@@ -0,0 +1 @@
+4aed00afd91e108d8986b062b968678fd1c4f8f7

vendor/MatAnyone2/LICENSE.txt

@@ -0,0 +1,42 @@
+# S-Lab License 1.0  
+
+Copyright 2023 S-Lab
+
+Redistribution and use for non-commercial purpose in source and 
+binary forms, with or without modification, are permitted provided 
+that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright 
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright 
+   notice, this list of conditions and the following disclaimer in 
+   the documentation and/or other materials provided with the 
+   distribution.
+
+3. Neither the name of the copyright holder nor the names of its 
+   contributors may be used to endorse or promote products derived 
+   from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+In the event that redistribution and/or use for commercial purpose in 
+source or binary forms, with or without modification is required, 
+please contact the contributor(s) of the work.
+
+
+---
+For inquiries permission for commercial use, please consult our team: 
+Peiqing Yang (peiqingyang99@outlook.com),
+Dr. Shangchen Zhou (shangchenzhou@gmail.com),
+Prof. Chen Change Loy (ccloy@ntu.edu.sg)

vendor/MatAnyone2/README.md

@@ -0,0 +1,178 @@
+<div align="center">
+<div style="text-align: center;">
+    <img src="./assets/matanyone2_logo.png" alt="MatAnyone Logo" style="height: 52px;">
+    <h2>Scaling Video Matting via a Learned Quality Evaluator</h2>
+</div>
+
+<div>
+    <a href='https://pq-yang.github.io/' target='_blank'>Peiqing Yang</a><sup>1</sup>&emsp;
+    <a href='https://shangchenzhou.com/' target='_blank'>Shangchen Zhou</a><sup>1†</sup>&emsp;
+    <a href="https://www.linkedin.com/in/kai-hao-794321382/" target='_blank'>Kai Hao</a><sup>1</sup>&emsp;
+    <a href="https://scholar.google.com.sg/citations?user=fMXnSGMAAAAJ&hl=en/" target='_blank'>Qingyi Tao</a><sup>2</sup>&emsp;
+</div>
+<div>
+    <sup>1</sup>S-Lab, Nanyang Technological University&emsp; 
+    <sup>2</sup>SenseTime Research, Singapore&emsp; 
+    <br>
+    <sup>†</sup>Project lead
+</div>
+
+
+<div>
+    <h4 align="center">
+        <a href="https://pq-yang.github.io/projects/MatAnyone2/" target='_blank'>
+        <img src="https://img.shields.io/badge/😈-Project%20Page-blue">
+        </a>
+        <a href="https://arxiv.org/abs/2512.11782" target='_blank'>
+        <img src="https://img.shields.io/badge/arXiv-2501.14677-b31b1b.svg">
+        </a>
+        <a href="https://www.youtube.com/watch?v=tyi8CNyjOhc&lc=Ugw1OS7z5QbW29RZCFZ4AaABAg" target='_blank'>
+        <img src="https://img.shields.io/badge/Demo%20Video-%23FF0000.svg?logo=YouTube&logoColor=white">
+        </a>
+        <a href="https://huggingface.co/spaces/PeiqingYang/MatAnyone" target='_blank'>
+        <img src="https://img.shields.io/badge/Demo-%F0%9F%A4%97%20Hugging%20Face-blue">
+        </a>
+        <img src="https://api.infinitescript.com/badgen/count?name=sczhou/MatAnyone2&ltext=Visitors&color=3977dd">
+    </h4>
+</div>
+
+<strong>MatAnyone 2 is a practical human video matting framework that preserves fine details by avoiding segmentation-like boundaries, while also shows enhanced robustness under challenging real-world conditions.</strong>
+
+<div style="width: 100%; text-align: center; margin:auto;">
+    <img style="width:100%" src="assets/teaser.jpg">
+</div>
+
+:movie_camera: For more visual results, go checkout our <a href="https://pq-yang.github.io/projects/MatAnyone2/" target="_blank">project page</a>
+
+---
+</div>
+
+
+## 📮 Update
+- [2026.03] Release inference codes, evaluation codes, and gradio demo.
+- [2025.12] This repo is created.
+
+
+## 🏄🏻‍♀️ TODO
+- [x] Release inference codes and gradio demo. 
+- [x] Release evaluation codes.
+- [ ] Release training codes for video matting model.
+- [ ] Release checkpoint and training codes for quality evaluator model.
+- [ ] Release real-world video matting dataset **VMReal**.
+
+
+## 🔎 Overview
+![overall_structure](assets/matanyone1vs2.jpg)
+
+## 🔧 Installation
+1. Clone Repo
+    ```bash
+    git clone https://github.com/pq-yang/MatAnyone2
+    cd MatAnyone2
+    ```
+
+2. Create Conda Environment and Install Dependencies
+    ```bash
+    # create new conda env
+    conda create -n matanyone2 python=3.10 -y
+    conda activate matanyone2
+
+    # install python dependencies
+    pip install -e .
+    # [optional] install python dependencies for gradio demo
+    pip3 install -r hugging_face/requirements.txt
+    ```
+
+## 🔥 Inference
+
+### Download Model
+Download our pretrained model from [MatAnyone 2](https://github.com/pq-yang/MatAnyone2/releases/download/v1.0.0/matanyone2.pth) to the `pretrained_models` folder (pretrained model can also be automatically downloaded during the first inference).
+
+The directory structure will be arranged as:
+```
+pretrained_models
+   |- matanyone2.pth
+```
+
+### Quick Test
+We provide some examples in the [`inputs`](./inputs) folder. **For each run, we take a video and its first-frame segmenatation mask as input.** <u>The segmenation mask could be obtained from interactive segmentation models such as [SAM2 demo](https://huggingface.co/spaces/fffiloni/SAM2-Image-Predictor)</u>. For example, the directory structure can be arranged as:
+```
+inputs
+   |- video
+      |- test-sample1          # folder containing all frames
+      |- test-sample2.mp4      # .mp4, .mov, .avi
+   |- mask
+      |- test-sample1.png      # mask for targer person(s)
+      |- test-sample2.png    
+```
+Run the following command to try it out:
+
+```shell
+# intput format: video folder
+python inference_matanyone2.py -i inputs/video/test-sample1 -m inputs/mask/test-sample1.png
+
+# intput format: mp4
+python inference_matanyone2.py -i inputs/video/test-sample2.mp4 -m inputs/mask/test-sample2.png
+
+```
+The results will be saved in the `results` folder, including the foreground output video and the alpha output video. 
+- If you want to save the results as per-frame images, you can set `--save_image`.
+- If you want to set a limit for the maximum input resolution, you can set `--max_size`, and the video will be downsampled if min(w, h) exceeds. By default, we don't set the limit.
+
+## 🎪 Interactive Demo
+To get rid of the preparation for first-frame segmentation mask, we prepare a gradio demo on [hugging face](https://huggingface.co/spaces/PeiqingYang/MatAnyone2) and could also **launch locally**. Just drop your video/image, assign the target masks with a few clicks, and get the the matting results!
+
+*We integrate MatAnyone Series in the demo. [MatAnyone 2](https://github.com/pq-yang/MatAnyone2) is the default model. You can also choose [MatAnyone](https://github.com/pq-yang/MatAnyone) as your processing model in "Model Selection".*
+
+```shell
+cd hugging_face
+
+# install python dependencies
+pip3 install -r requirements.txt # FFmpeg required
+
+# launch the demo
+python app.py
+```
+
+By launching, an interactive interface will appear as follow.
+
+![overall_teaser](assets/teaser_demo.gif)
+
+## 📊 Evaluation
+Please refer to the [evaluation documentation](docs/EVAL.md) for details.
+
+## 🛠️ Data Pipeline
+![data_pipeline](assets/data_pipeline.jpg)
+
+
+## 📑 Citation
+
+   If you find our repo useful for your research, please consider citing our paper:
+
+   ```bibtex
+  @InProceedings{yang2026matanyone2,
+      title     = {{MatAnyone 2}: Scaling Video Matting via a Learned Quality Evaluator},
+      author    = {Yang, Peiqing and Zhou, Shangchen and Hao, Kai and Tao, Qingyi},
+      booktitle = {CVPR},
+      year      = {2026}
+      }
+
+  @inProceedings{yang2025matanyone,
+      title     = {{MatAnyone}: Stable Video Matting with Consistent Memory Propagation},
+      author    = {Yang, Peiqing and Zhou, Shangchen and Zhao, Jixin and Tao, Qingyi and Loy, Chen Change},
+      booktitle = {CVPR},
+      year      = {2025}
+      }
+   ```
+
+## 📝 License
+
+This project is licensed under <a rel="license" href="./LICENSE">NTU S-Lab License 1.0</a>. Redistribution and use should follow this license.
+
+## 👏 Acknowledgement
+
+This project is built upon [MatAnyone](https://github.com/pq-yang/MatAnyone) and [Cutie](https://github.com/hkchengrex/Cutie), with matting dataset files adapted from [RVM](https://github.com/PeterL1n/RobustVideoMatting). The interactive demo is adapted from [ProPainter](https://github.com/sczhou/ProPainter), leveraging segmentation capabilities from [Segment Anything Model](https://github.com/facebookresearch/segment-anything) and [Segment Anything Model 2](https://github.com/facebookresearch/sam2). Thanks for their awesome works!
+
+## 📧 Contact
+
+If you have any questions, please feel free to reach us at `peiqingyang99@outlook.com`. 

vendor/MatAnyone2/matanyone2/__init__.py

@@ -0,0 +1,2 @@
+from matanyone2.inference.inference_core import InferenceCore
+from matanyone2.model.matanyone2 import MatAnyone2

vendor/MatAnyone2/matanyone2/config/eval_matanyone_config.yaml

@@ -0,0 +1,47 @@
+defaults:
+  - _self_
+  - model: base
+  - override hydra/job_logging: custom-no-rank.yaml
+
+hydra:
+  run:
+    dir: ../output/${exp_id}/${dataset}
+  output_subdir: ${now:%Y-%m-%d_%H-%M-%S}-hydra
+
+amp: False
+weights: pretrained_models/matanyone2.pth # default (can be modified from outside)
+output_dir: null # defaults to run_dir; specify this to override
+flip_aug: False
+
+
+# maximum shortest side of the input; -1 means no resizing
+# With eval_vos.py, we usually just use the dataset's size (resizing done in dataloader)
+# this parameter is added for the sole purpose for the GUI in the current codebase
+# InferenceCore will downsize the input and restore the output to the original size if needed
+# if you are using this code for some other project, you can also utilize this parameter
+max_internal_size: -1
+
+# these parameters, when set, override the dataset's default; useful for debugging
+save_all: True
+use_all_masks: False
+use_long_term: False
+mem_every: 5
+
+# only relevant when long_term is not enabled
+max_mem_frames: 5
+
+# only relevant when long_term is enabled
+long_term:
+  count_usage: True
+  max_mem_frames: 10
+  min_mem_frames: 5
+  num_prototypes: 128
+  max_num_tokens: 10000
+  buffer_tokens: 2000
+
+top_k: 30
+stagger_updates: 5
+chunk_size: -1 # number of objects to process in parallel; -1 means unlimited
+save_scores: False
+save_aux: False
+visualize: False

vendor/MatAnyone2/matanyone2/config/hydra/job_logging/custom-no-rank.yaml

@@ -0,0 +1,22 @@
+# python logging configuration for tasks
+version: 1
+formatters:
+  simple:
+    format: '[%(asctime)s][%(levelname)s] - %(message)s'
+    datefmt: '%Y-%m-%d %H:%M:%S'
+handlers:
+  console:
+    class: logging.StreamHandler
+    formatter: simple
+    stream: ext://sys.stdout
+  file:
+    class: logging.FileHandler
+    formatter: simple
+    # absolute file path
+    filename: ${hydra.runtime.output_dir}/${now:%Y-%m-%d_%H-%M-%S}-eval.log
+    mode: w
+root:
+  level: INFO
+  handlers: [console, file]
+
+disable_existing_loggers: false

vendor/MatAnyone2/matanyone2/config/hydra/job_logging/custom.yaml

@@ -0,0 +1,22 @@
+# python logging configuration for tasks
+version: 1
+formatters:
+  simple:
+    format: '[%(asctime)s][%(levelname)s][r${oc.env:LOCAL_RANK}] - %(message)s'
+    datefmt: '%Y-%m-%d %H:%M:%S'
+handlers:
+  console:
+    class: logging.StreamHandler
+    formatter: simple
+    stream: ext://sys.stdout
+  file:
+    class: logging.FileHandler
+    formatter: simple
+    # absolute file path
+    filename: ${hydra.runtime.output_dir}/${now:%Y-%m-%d_%H-%M-%S}-rank${oc.env:LOCAL_RANK}.log
+    mode: w
+root:
+  level: INFO
+  handlers: [console, file]
+
+disable_existing_loggers: false

vendor/MatAnyone2/matanyone2/config/model/base.yaml

@@ -0,0 +1,58 @@
+pixel_mean: [0.485, 0.456, 0.406]
+pixel_std: [0.229, 0.224, 0.225]
+
+pixel_dim: 256
+key_dim: 64
+value_dim: 256
+sensory_dim: 256
+embed_dim: 256
+
+pixel_encoder:
+  type: resnet50
+  ms_dims: [1024, 512, 256, 64, 3] # f16, f8, f4, f2, f1
+
+mask_encoder:
+  type: resnet18
+  final_dim: 256
+
+pixel_pe_scale: 32
+pixel_pe_temperature: 128
+
+object_transformer:
+  embed_dim: ${model.embed_dim}
+  ff_dim: 2048
+  num_heads: 8
+  num_blocks: 3
+  num_queries: 16
+  read_from_pixel:
+    input_norm: False
+    input_add_pe: False
+    add_pe_to_qkv: [True, True, False]
+  read_from_past:
+    add_pe_to_qkv: [True, True, False]
+  read_from_memory:
+    add_pe_to_qkv: [True, True, False]
+  read_from_query:
+    add_pe_to_qkv: [True, True, False]
+    output_norm: False
+  query_self_attention:
+    add_pe_to_qkv: [True, True, False]
+  pixel_self_attention:
+    add_pe_to_qkv: [True, True, False]
+
+object_summarizer:
+  embed_dim: ${model.object_transformer.embed_dim}
+  num_summaries: ${model.object_transformer.num_queries}
+  add_pe: True
+
+aux_loss:
+  sensory:
+    enabled: True
+    weight: 0.01
+  query:
+    enabled: True
+    weight: 0.01
+
+mask_decoder:
+  # first value must equal embed_dim
+  up_dims: [256, 128, 128, 64, 16]

vendor/MatAnyone2/matanyone2/inference/image_feature_store.py

@@ -0,0 +1,56 @@
+import warnings
+from typing import Iterable
+import torch
+from matanyone2.model.matanyone2 import MatAnyone2
+
+
+class ImageFeatureStore:
+    """
+    A cache for image features.
+    These features might be reused at different parts of the inference pipeline.
+    This class provide an interface for reusing these features.
+    It is the user's responsibility to delete redundant features.
+
+    Feature of a frame should be associated with a unique index -- typically the frame id.
+    """
+    def __init__(self, network: MatAnyone2, no_warning: bool = False):
+        self.network = network
+        self._store = {}
+        self.no_warning = no_warning
+
+    def _encode_feature(self, index: int, image: torch.Tensor, last_feats=None) -> None:
+        ms_features, pix_feat = self.network.encode_image(image, last_feats=last_feats)
+        key, shrinkage, selection = self.network.transform_key(ms_features[0])
+        self._store[index] = (ms_features, pix_feat, key, shrinkage, selection)
+   
+    def get_all_features(self, images: torch.Tensor) -> (Iterable[torch.Tensor], torch.Tensor):
+        seq_length = images.shape[0]
+        ms_features, pix_feat = self.network.encode_image(images, seq_length)
+        key, shrinkage, selection = self.network.transform_key(ms_features[0])
+        for index in range(seq_length):
+            self._store[index] = ([f[index].unsqueeze(0) for f in ms_features], pix_feat[index].unsqueeze(0), key[index].unsqueeze(0), shrinkage[index].unsqueeze(0), selection[index].unsqueeze(0))
+
+    def get_features(self, index: int,
+                     image: torch.Tensor, last_feats=None) -> (Iterable[torch.Tensor], torch.Tensor):
+        if index not in self._store:
+            self._encode_feature(index, image, last_feats)
+
+        return self._store[index][:2]
+
+    def get_key(self, index: int,
+                image: torch.Tensor, last_feats=None) -> (torch.Tensor, torch.Tensor, torch.Tensor):
+        if index not in self._store:
+            self._encode_feature(index, image, last_feats)
+
+        return self._store[index][2:]
+
+    def delete(self, index: int) -> None:
+        if index in self._store:
+            del self._store[index]
+
+    def __len__(self):
+        return len(self._store)
+
+    def __del__(self):
+        if len(self._store) > 0 and not self.no_warning:
+            warnings.warn(f'Leaking {self._store.keys()} in the image feature store')

vendor/MatAnyone2/matanyone2/inference/inference_core.py

@@ -0,0 +1,550 @@
+import logging
+from omegaconf import DictConfig
+from typing import List, Optional, Iterable, Union,Tuple
+
+import os
+import cv2
+import torch
+import imageio
+import tempfile
+import numpy as np
+from tqdm import tqdm
+from PIL import Image
+import torch.nn.functional as F
+
+from matanyone2.inference.memory_manager import MemoryManager
+from matanyone2.inference.object_manager import ObjectManager
+from matanyone2.inference.image_feature_store import ImageFeatureStore
+from matanyone2.model.matanyone2 import MatAnyone2
+from matanyone2.utils.tensor_utils import pad_divide_by, unpad, aggregate
+from matanyone2.utils.inference_utils import gen_dilate, gen_erosion, read_frame_from_videos
+from matanyone2.utils.device import get_default_device, safe_autocast
+
+
+log = logging.getLogger()
+
+
+class InferenceCore:
+
+    def __init__(self,
+                 network: Union[MatAnyone2,str],
+                 cfg: DictConfig = None,
+                 *,
+                 image_feature_store: ImageFeatureStore = None,
+                 device: Optional[Union[str, torch.device]] = None
+                 ):
+        if device is None:
+            device = get_default_device()
+        self.device = device
+        if isinstance(network, str):
+            network = MatAnyone2.from_pretrained(network)
+        network.to(device)
+        network.eval()
+        self.network = network  
+        cfg = cfg if cfg is not None else network.cfg
+        self.cfg = cfg
+        self.mem_every = cfg.mem_every
+        stagger_updates = cfg.stagger_updates
+        self.chunk_size = cfg.chunk_size
+        self.save_aux = cfg.save_aux
+        self.max_internal_size = cfg.max_internal_size
+        self.flip_aug = cfg.flip_aug
+
+        self.curr_ti = -1
+        self.last_mem_ti = 0
+        # at which time indices should we update the sensory memory
+        if stagger_updates >= self.mem_every:
+            self.stagger_ti = set(range(1, self.mem_every + 1))
+        else:
+            self.stagger_ti = set(
+                np.round(np.linspace(1, self.mem_every, stagger_updates)).astype(int))
+        self.object_manager = ObjectManager()
+        self.memory = MemoryManager(cfg=cfg, object_manager=self.object_manager)
+
+        if image_feature_store is None:
+            self.image_feature_store = ImageFeatureStore(self.network)
+        else:
+            self.image_feature_store = image_feature_store
+
+        self.last_mask = None
+        self.last_pix_feat = None
+        self.last_msk_value = None
+
+    def clear_memory(self):
+        self.curr_ti = -1
+        self.last_mem_ti = 0
+        self.memory = MemoryManager(cfg=self.cfg, object_manager=self.object_manager)
+
+    def clear_non_permanent_memory(self):
+        self.curr_ti = -1
+        self.last_mem_ti = 0
+        self.memory.clear_non_permanent_memory()
+
+    def clear_sensory_memory(self):
+        self.curr_ti = -1
+        self.last_mem_ti = 0
+        self.memory.clear_sensory_memory()
+
+    def update_config(self, cfg):
+        self.mem_every = cfg['mem_every']
+        self.memory.update_config(cfg)
+    
+    def clear_temp_mem(self):
+        self.memory.clear_work_mem()
+        # self.object_manager = ObjectManager()
+        self.memory.clear_obj_mem()
+        # self.memory.clear_sensory_memory()
+
+    def _add_memory(self,
+                    image: torch.Tensor,
+                    pix_feat: torch.Tensor,
+                    prob: torch.Tensor,
+                    key: torch.Tensor,
+                    shrinkage: torch.Tensor,
+                    selection: torch.Tensor,
+                    *,
+                    is_deep_update: bool = True,
+                    force_permanent: bool = False) -> None:
+        """
+        Memorize the given segmentation in all memory stores.
+
+        The batch dimension is 1 if flip augmentation is not used.
+        image: RGB image, (1/2)*3*H*W
+        pix_feat: from the key encoder, (1/2)*_*H*W
+        prob: (1/2)*num_objects*H*W, in [0, 1]
+        key/shrinkage/selection: for anisotropic l2, (1/2)*_*H*W
+        selection can be None if not using long-term memory
+        is_deep_update: whether to use deep update (e.g. with the mask encoder)
+        force_permanent: whether to force the memory to be permanent
+        """
+        if prob.shape[1] == 0:
+            # nothing to add
+            log.warn('Trying to add an empty object mask to memory!')
+            return
+
+        if force_permanent:
+            as_permanent = 'all'
+        else:
+            as_permanent = 'first'
+
+        self.memory.initialize_sensory_if_needed(key, self.object_manager.all_obj_ids)
+        msk_value, sensory, obj_value, _ = self.network.encode_mask(
+            image,
+            pix_feat,
+            self.memory.get_sensory(self.object_manager.all_obj_ids),
+            prob,
+            deep_update=is_deep_update,
+            chunk_size=self.chunk_size,
+            need_weights=self.save_aux)
+        self.memory.add_memory(key,
+                               shrinkage,
+                               msk_value,
+                               obj_value,
+                               self.object_manager.all_obj_ids,
+                               selection=selection,
+                               as_permanent=as_permanent)
+        self.last_mem_ti = self.curr_ti
+        if is_deep_update:
+            self.memory.update_sensory(sensory, self.object_manager.all_obj_ids)
+        self.last_msk_value = msk_value
+
+    def _segment(self,
+                 key: torch.Tensor,
+                 selection: torch.Tensor,
+                 pix_feat: torch.Tensor,
+                 ms_features: Iterable[torch.Tensor],
+                 update_sensory: bool = True) -> torch.Tensor:
+        """
+        Produce a segmentation using the given features and the memory
+
+        The batch dimension is 1 if flip augmentation is not used.
+        key/selection: for anisotropic l2: (1/2) * _ * H * W
+        pix_feat: from the key encoder, (1/2) * _ * H * W
+        ms_features: an iterable of multiscale features from the encoder, each is (1/2)*_*H*W
+                      with strides 16, 8, and 4 respectively
+        update_sensory: whether to update the sensory memory
+
+        Returns: (num_objects+1)*H*W normalized probability; the first channel is the background
+        """
+        bs = key.shape[0]
+        if self.flip_aug:
+            assert bs == 2
+        else:
+            assert bs == 1
+
+        if not self.memory.engaged:
+            log.warn('Trying to segment without any memory!')
+            return torch.zeros((1, key.shape[-2] * 16, key.shape[-1] * 16),
+                               device=key.device,
+                               dtype=key.dtype)
+        
+        uncert_output = None
+
+        if self.curr_ti == 0: # ONLY for the first frame for prediction
+            memory_readout = self.memory.read_first_frame(self.last_msk_value, pix_feat, self.last_mask, self.network, uncert_output=uncert_output)
+        else:
+            memory_readout = self.memory.read(pix_feat, key, selection, self.last_mask, self.network, uncert_output=uncert_output, last_msk_value=self.last_msk_value, ti=self.curr_ti, 
+                                              last_pix_feat=self.last_pix_feat, last_pred_mask=self.last_mask)
+        memory_readout = self.object_manager.realize_dict(memory_readout)
+
+        sensory, _, pred_prob_with_bg = self.network.segment(ms_features,
+                                                        memory_readout,
+                                                        self.memory.get_sensory(
+                                                            self.object_manager.all_obj_ids),
+                                                        chunk_size=self.chunk_size,
+                                                        update_sensory=update_sensory)
+        # remove batch dim
+        if self.flip_aug:
+            # average predictions of the non-flipped and flipped version
+            pred_prob_with_bg = (pred_prob_with_bg[0] +
+                                 torch.flip(pred_prob_with_bg[1], dims=[-1])) / 2
+        else:
+            pred_prob_with_bg = pred_prob_with_bg[0]
+        if update_sensory:
+            self.memory.update_sensory(sensory, self.object_manager.all_obj_ids)
+        return pred_prob_with_bg
+    
+    def pred_all_flow(self, images):
+        self.total_len = images.shape[0]
+        images, self.pad = pad_divide_by(images, 16)
+        images = images.unsqueeze(0)  # add the batch dimension: (1,t,c,h,w)
+        
+        self.flows_forward, self.flows_backward = self.network.pred_forward_backward_flow(images)
+
+    def encode_all_images(self, images):
+        images, self.pad = pad_divide_by(images, 16)
+        self.image_feature_store.get_all_features(images) # t c h w
+        return images
+
+    def step(self,
+             image: torch.Tensor,
+             mask: Optional[torch.Tensor] = None,
+             objects: Optional[List[int]] = None,
+             *,
+             idx_mask: bool = False,
+             end: bool = False,
+             delete_buffer: bool = True,
+             force_permanent: bool = False,
+             matting: bool = True,
+             first_frame_pred: bool = False) -> torch.Tensor:
+        """
+        Take a step with a new incoming image.
+        If there is an incoming mask with new objects, we will memorize them.
+        If there is no incoming mask, we will segment the image using the memory.
+        In both cases, we will update the memory and return a segmentation.
+
+        image: 3*H*W
+        mask: H*W (if idx mask) or len(objects)*H*W or None
+        objects: list of object ids that are valid in the mask Tensor.
+                The ids themselves do not need to be consecutive/in order, but they need to be 
+                in the same position in the list as the corresponding mask
+                in the tensor in non-idx-mask mode.
+                objects is ignored if the mask is None. 
+                If idx_mask is False and objects is None, we sequentially infer the object ids.
+        idx_mask: if True, mask is expected to contain an object id at every pixel.
+                  If False, mask should have multiple channels with each channel representing one object.
+        end: if we are at the end of the sequence, we do not need to update memory
+            if unsure just set it to False 
+        delete_buffer: whether to delete the image feature buffer after this step
+        force_permanent: the memory recorded this frame will be added to the permanent memory
+        """
+        if objects is None and mask is not None:
+            assert not idx_mask
+            objects = list(range(1, mask.shape[0] + 1))
+
+        # resize input if needed -- currently only used for the GUI
+        resize_needed = False
+        if self.max_internal_size > 0:
+            h, w = image.shape[-2:]
+            min_side = min(h, w)
+            if min_side > self.max_internal_size:
+                resize_needed = True
+                new_h = int(h / min_side * self.max_internal_size)
+                new_w = int(w / min_side * self.max_internal_size)
+                image = F.interpolate(image.unsqueeze(0),
+                                      size=(new_h, new_w),
+                                      mode='bilinear',
+                                      align_corners=False)[0]
+                if mask is not None:
+                    if idx_mask:
+                        mask = F.interpolate(mask.unsqueeze(0).unsqueeze(0).float(),
+                                             size=(new_h, new_w),
+                                             mode='nearest-exact',
+                                             align_corners=False)[0, 0].round().long()
+                    else:
+                        mask = F.interpolate(mask.unsqueeze(0),
+                                             size=(new_h, new_w),
+                                             mode='bilinear',
+                                             align_corners=False)[0]
+
+        self.curr_ti += 1
+
+        image, self.pad = pad_divide_by(image, 16) # DONE alreay for 3DCNN!!
+        image = image.unsqueeze(0)  # add the batch dimension
+        if self.flip_aug:
+            image = torch.cat([image, torch.flip(image, dims=[-1])], dim=0)
+
+        # whether to update the working memory
+        is_mem_frame = ((self.curr_ti - self.last_mem_ti >= self.mem_every) or
+                        (mask is not None)) and (not end)
+        # segment when there is no input mask or when the input mask is incomplete
+        need_segment = (mask is None) or (self.object_manager.num_obj > 0
+                                          and not self.object_manager.has_all(objects))
+        update_sensory = ((self.curr_ti - self.last_mem_ti) in self.stagger_ti) and (not end)
+
+        # reinit if it is the first frame for prediction
+        if first_frame_pred:
+            self.curr_ti = 0
+            self.last_mem_ti = 0
+            is_mem_frame = True
+            need_segment = True
+            update_sensory = True
+
+        # encoding the image
+        ms_feat, pix_feat = self.image_feature_store.get_features(self.curr_ti, image)
+        key, shrinkage, selection = self.image_feature_store.get_key(self.curr_ti, image)
+
+        # segmentation from memory if needed
+        if need_segment:
+            pred_prob_with_bg = self._segment(key,
+                                              selection,
+                                              pix_feat,
+                                              ms_feat,
+                                              update_sensory=update_sensory)
+
+        # use the input mask if provided
+        if mask is not None:
+            # inform the manager of the new objects, and get a list of temporary id
+            # temporary ids -- indicates the position of objects in the tensor
+            # (starts with 1 due to the background channel)
+            corresponding_tmp_ids, _ = self.object_manager.add_new_objects(objects)
+
+            mask, _ = pad_divide_by(mask, 16)
+            if need_segment:
+                # merge predicted mask with the incomplete input mask
+                pred_prob_no_bg = pred_prob_with_bg[1:]
+                # use the mutual exclusivity of segmentation
+                if idx_mask:
+                    pred_prob_no_bg[:, mask > 0] = 0
+                else:
+                    pred_prob_no_bg[:, mask.max(0) > 0.5] = 0
+
+                new_masks = []
+                for mask_id, tmp_id in enumerate(corresponding_tmp_ids):
+                    if idx_mask:
+                        this_mask = (mask == objects[mask_id]).type_as(pred_prob_no_bg)
+                    else:
+                        this_mask = mask[tmp_id]
+                    if tmp_id > pred_prob_no_bg.shape[0]:
+                        new_masks.append(this_mask.unsqueeze(0))
+                    else:
+                        # +1 for padding the background channel
+                        pred_prob_no_bg[tmp_id - 1] = this_mask
+                # new_masks are always in the order of tmp_id
+                mask = torch.cat([pred_prob_no_bg, *new_masks], dim=0)
+            elif idx_mask:
+                # simply convert cls to one-hot representation
+                if len(objects) == 0:
+                    if delete_buffer:
+                        self.image_feature_store.delete(self.curr_ti)
+                    log.warn('Trying to insert an empty mask as memory!')
+                    return torch.zeros((1, key.shape[-2] * 16, key.shape[-1] * 16),
+                                       device=key.device,
+                                       dtype=key.dtype)
+                mask = torch.stack(
+                    [mask == objects[mask_id] for mask_id, _ in enumerate(corresponding_tmp_ids)],
+                    dim=0)
+            if matting:
+                mask = mask.unsqueeze(0).float() / 255.
+                pred_prob_with_bg = torch.cat([1-mask, mask], 0)
+            else:
+                pred_prob_with_bg = aggregate(mask, dim=0)
+                pred_prob_with_bg = torch.softmax(pred_prob_with_bg, dim=0)
+
+        self.last_mask = pred_prob_with_bg[1:].unsqueeze(0)
+        if self.flip_aug:
+            self.last_mask = torch.cat(
+                [self.last_mask, torch.flip(self.last_mask, dims=[-1])], dim=0)
+        self.last_pix_feat = pix_feat
+
+        # save as memory if needed
+        if is_mem_frame or force_permanent:
+            # clear the memory for given mask and add the first predicted mask
+            if first_frame_pred:
+                self.clear_temp_mem()
+            self._add_memory(image,
+                             pix_feat,
+                             self.last_mask,
+                             key,
+                             shrinkage,
+                             selection,
+                             force_permanent=force_permanent,
+                             is_deep_update=True)
+        else: # compute self.last_msk_value for non-memory frame 
+            msk_value, _, _, _ = self.network.encode_mask(
+                            image,
+                            pix_feat,
+                            self.memory.get_sensory(self.object_manager.all_obj_ids),
+                            self.last_mask,
+                            deep_update=False,
+                            chunk_size=self.chunk_size,
+                            need_weights=self.save_aux)
+            self.last_msk_value = msk_value
+
+        if delete_buffer:
+            self.image_feature_store.delete(self.curr_ti)
+
+        output_prob = unpad(pred_prob_with_bg, self.pad)
+        if resize_needed:
+            # restore output to the original size
+            output_prob = F.interpolate(output_prob.unsqueeze(0),
+                                        size=(h, w),
+                                        mode='bilinear',
+                                        align_corners=False)[0]
+
+        return output_prob
+
+    def delete_objects(self, objects: List[int]) -> None:
+        """
+        Delete the given objects from the memory.
+        """
+        self.object_manager.delete_objects(objects)
+        self.memory.purge_except(self.object_manager.all_obj_ids)
+
+    def output_prob_to_mask(self, output_prob: torch.Tensor, matting: bool = True) -> torch.Tensor:
+        if matting:
+            new_mask = output_prob[1:].squeeze(0)
+        else:
+            mask = torch.argmax(output_prob, dim=0)
+
+            # index in tensor != object id -- remap the ids here
+            new_mask = torch.zeros_like(mask)
+            for tmp_id, obj in self.object_manager.tmp_id_to_obj.items():
+                new_mask[mask == tmp_id] = obj.id
+
+        return new_mask
+
+    @torch.inference_mode()
+    @safe_autocast()
+    def process_video(
+        self,
+        input_path: str,
+        mask_path: str,
+        output_path: str = None,
+        n_warmup: int = 10,
+        r_erode: int = 10,
+        r_dilate: int = 10,
+        suffix: str = "",
+        save_image: bool = False,
+        max_size: int = -1,
+    ) -> Tuple:
+        """
+        Process a video for object segmentation and matting.
+        This method processes a video file by performing object segmentation and matting on each frame.
+        It supports warmup frames, mask erosion/dilation, and various output options.
+        Args:
+            input_path (str): Path to the input video file
+            mask_path (str): Path to the mask image file used for initial segmentation
+            output_path (str, optional): Directory path where output files will be saved. Defaults to a temporary directory
+            n_warmup (int, optional): Number of warmup frames to use. Defaults to 10
+            r_erode (int, optional): Erosion radius for mask processing. Defaults to 10
+            r_dilate (int, optional): Dilation radius for mask processing. Defaults to 10
+            suffix (str, optional): Suffix to append to output filename. Defaults to ""
+            save_image (bool, optional): Whether to save individual frames. Defaults to False
+            max_size (int, optional): Maximum size for frame dimension. Use -1 for no limit. Defaults to -1
+        Returns:
+            Tuple[str, str]: A tuple containing:
+                - Path to the output foreground video file (str)
+                - Path to the output alpha matte video file (str)
+        Output:
+            - Saves processed video files with foreground (_fgr) and alpha matte (_pha)
+            - If save_image=True, saves individual frames in separate directories
+        """
+        output_path = output_path if output_path is not None else tempfile.TemporaryDirectory().name
+        r_erode = int(r_erode)
+        r_dilate = int(r_dilate)
+        n_warmup = int(n_warmup)
+        max_size = int(max_size)
+
+        vframes, fps, length, video_name = read_frame_from_videos(input_path)
+        repeated_frames = vframes[0].unsqueeze(0).repeat(n_warmup, 1, 1, 1)
+        vframes = torch.cat([repeated_frames, vframes], dim=0).float()
+        length += n_warmup
+
+        new_h, new_w = vframes.shape[-2:]
+        if max_size > 0:
+            h, w = new_h, new_w
+            min_side = min(h, w)
+            if min_side > max_size:
+                new_h = int(h / min_side * max_size)
+                new_w = int(w / min_side * max_size)
+                vframes = F.interpolate(vframes, size=(new_h, new_w), mode="area")
+
+        os.makedirs(output_path, exist_ok=True)
+        if suffix:
+            video_name = f"{video_name}_{suffix}"
+        if save_image:
+            os.makedirs(f"{output_path}/{video_name}", exist_ok=True)
+            os.makedirs(f"{output_path}/{video_name}/pha", exist_ok=True)
+            os.makedirs(f"{output_path}/{video_name}/fgr", exist_ok=True)
+
+        mask = np.array(Image.open(mask_path).convert("L"))
+        if r_dilate > 0:
+            mask = gen_dilate(mask, r_dilate, r_dilate)
+        if r_erode > 0:
+            mask = gen_erosion(mask, r_erode, r_erode)
+        
+        mask = torch.from_numpy(mask).float().to(self.device)
+        if max_size > 0:
+            mask = F.interpolate(
+                mask.unsqueeze(0).unsqueeze(0), size=(new_h, new_w), mode="nearest"
+            )[0, 0]
+
+        bgr = (np.array([120, 255, 155], dtype=np.float32) / 255).reshape((1, 1, 3))
+        objects = [1]
+
+        phas = []
+        fgrs = []
+        for ti in tqdm(range(length)):
+            image = vframes[ti]
+            image_np = np.array(image.permute(1, 2, 0))
+            image = (image / 255.0).float().to(self.device)
+
+            if ti == 0:
+                output_prob = self.step(image, mask, objects=objects)
+                output_prob = self.step(image, first_frame_pred=True)
+            else:
+                if ti <= n_warmup:
+                    output_prob = self.step(image, first_frame_pred=True)
+                else:
+                    output_prob = self.step(image)
+
+            mask = self.output_prob_to_mask(output_prob)
+            pha = mask.unsqueeze(2).cpu().numpy()
+            com_np = image_np / 255.0 * pha + bgr * (1 - pha)
+
+            if ti > (n_warmup - 1):
+                com_np = (com_np * 255).astype(np.uint8)
+                pha = (pha * 255).astype(np.uint8)
+                fgrs.append(com_np)
+                phas.append(pha)
+                if save_image:
+                    cv2.imwrite(
+                        f"{output_path}/{video_name}/pha/{str(ti - n_warmup).zfill(5)}.png",
+                        pha,
+                    )
+                    cv2.imwrite(
+                        f"{output_path}/{video_name}/fgr/{str(ti - n_warmup).zfill(5)}.png",
+                        com_np[..., [2, 1, 0]],
+                    )
+
+        fgrs = np.array(fgrs)
+        phas = np.array(phas)
+        
+        fgr_filename = f"{output_path}/{video_name}_fgr.mp4"
+        alpha_filename = f"{output_path}/{video_name}_pha.mp4"
+        
+        imageio.mimwrite(fgr_filename, fgrs, fps=fps, quality=7)
+        imageio.mimwrite(alpha_filename, phas, fps=fps, quality=7)
+        
+        return (fgr_filename,alpha_filename)

vendor/MatAnyone2/matanyone2/inference/kv_memory_store.py

@@ -0,0 +1,348 @@
+from typing import Dict, List, Optional, Literal
+from collections import defaultdict
+import torch
+
+
+def _add_last_dim(dictionary, key, new_value, prepend=False):
+    # append/prepend a new value to the last dimension of a tensor in a dictionary
+    # if the key does not exist, put the new value in
+    # append by default
+    if key in dictionary:
+        dictionary[key] = torch.cat([dictionary[key], new_value], -1)
+    else:
+        dictionary[key] = new_value
+
+
+class KeyValueMemoryStore:
+    """
+    Works for key/value pairs type storage
+    e.g., working and long-term memory
+    """
+    def __init__(self, save_selection: bool = False, save_usage: bool = False):
+        """
+        We store keys and values of objects that first appear in the same frame in a bucket.
+        Each bucket contains a set of object ids.
+        Each bucket is associated with a single key tensor
+            and a dictionary of value tensors indexed by object id.
+
+        The keys and values are stored as the concatenation of a permanent part and a temporary part.
+        """
+        self.save_selection = save_selection
+        self.save_usage = save_usage
+
+        self.global_bucket_id = 0  # does not reduce even if buckets are removed
+        self.buckets: Dict[int, List[int]] = {}  # indexed by bucket id
+        self.k: Dict[int, torch.Tensor] = {}  # indexed by bucket id
+        self.v: Dict[int, torch.Tensor] = {}  # indexed by object id
+
+        # indexed by bucket id; the end point of permanent memory
+        self.perm_end_pt: Dict[int, int] = defaultdict(int)
+
+        # shrinkage and selection are just like the keys
+        self.s = {}
+        if self.save_selection:
+            self.e = {}  # does not contain the permanent memory part
+
+        # usage
+        if self.save_usage:
+            self.use_cnt = {}  # indexed by bucket id, does not contain the permanent memory part
+            self.life_cnt = {}  # indexed by bucket id, does not contain the permanent memory part
+
+    def add(self,
+            key: torch.Tensor,
+            values: Dict[int, torch.Tensor],
+            shrinkage: torch.Tensor,
+            selection: torch.Tensor,
+            supposed_bucket_id: int = -1,
+            as_permanent: Literal['no', 'first', 'all'] = 'no') -> None:
+        """
+        key: (1/2)*C*N
+        values: dict of values ((1/2)*C*N), object ids are used as keys
+        shrinkage: (1/2)*1*N
+        selection: (1/2)*C*N
+
+        supposed_bucket_id: used to sync the bucket id between working and long-term memory
+        if provided, the input should all be in a single bucket indexed by this id
+        as_permanent: whether to store the input as permanent memory
+            'no': don't
+            'first': only store it as permanent memory if the bucket is empty
+            'all': always store it as permanent memory
+        """
+        bs = key.shape[0]
+        ne = key.shape[-1]
+        assert len(key.shape) == 3
+        assert len(shrinkage.shape) == 3
+        assert not self.save_selection or len(selection.shape) == 3
+        assert as_permanent in ['no', 'first', 'all']
+
+        # add the value and create new buckets if necessary
+        if supposed_bucket_id >= 0:
+            enabled_buckets = [supposed_bucket_id]
+            bucket_exist = supposed_bucket_id in self.buckets
+            for obj, value in values.items():
+                if bucket_exist:
+                    assert obj in self.v
+                    assert obj in self.buckets[supposed_bucket_id]
+                    _add_last_dim(self.v, obj, value, prepend=(as_permanent == 'all'))
+                else:
+                    assert obj not in self.v
+                    self.v[obj] = value
+            self.buckets[supposed_bucket_id] = list(values.keys())
+        else:
+            new_bucket_id = None
+            enabled_buckets = set()
+            for obj, value in values.items():
+                assert len(value.shape) == 3
+                if obj in self.v:
+                    _add_last_dim(self.v, obj, value, prepend=(as_permanent == 'all'))
+                    bucket_used = [
+                        bucket_id for bucket_id, object_ids in self.buckets.items()
+                        if obj in object_ids
+                    ]
+                    assert len(bucket_used) == 1  # each object should only be in one bucket
+                    enabled_buckets.add(bucket_used[0])
+                else:
+                    self.v[obj] = value
+                    if new_bucket_id is None:
+                        # create new bucket
+                        new_bucket_id = self.global_bucket_id
+                        self.global_bucket_id += 1
+                        self.buckets[new_bucket_id] = []
+                    # put the new object into the corresponding bucket
+                    self.buckets[new_bucket_id].append(obj)
+                    enabled_buckets.add(new_bucket_id)
+
+        # increment the permanent size if necessary
+        add_as_permanent = {}  # indexed by bucket id
+        for bucket_id in enabled_buckets:
+            add_as_permanent[bucket_id] = False
+            if as_permanent == 'all':
+                self.perm_end_pt[bucket_id] += ne
+                add_as_permanent[bucket_id] = True
+            elif as_permanent == 'first':
+                if self.perm_end_pt[bucket_id] == 0:
+                    self.perm_end_pt[bucket_id] = ne
+                    add_as_permanent[bucket_id] = True
+
+        # create new counters for usage if necessary
+        if self.save_usage and as_permanent != 'all':
+            new_count = torch.zeros((bs, ne), device=key.device, dtype=torch.float32)
+            new_life = torch.zeros((bs, ne), device=key.device, dtype=torch.float32) + 1e-7
+
+        # add the key to every bucket
+        for bucket_id in self.buckets:
+            if bucket_id not in enabled_buckets:
+                # if we are not adding new values to a bucket, we should skip it
+                continue
+
+            _add_last_dim(self.k, bucket_id, key, prepend=add_as_permanent[bucket_id])
+            _add_last_dim(self.s, bucket_id, shrinkage, prepend=add_as_permanent[bucket_id])
+            if not add_as_permanent[bucket_id]:
+                if self.save_selection:
+                    _add_last_dim(self.e, bucket_id, selection)
+                if self.save_usage:
+                    _add_last_dim(self.use_cnt, bucket_id, new_count)
+                    _add_last_dim(self.life_cnt, bucket_id, new_life)
+
+    def update_bucket_usage(self, bucket_id: int, usage: torch.Tensor) -> None:
+        # increase all life count by 1
+        # increase use of indexed elements
+        if not self.save_usage:
+            return
+
+        usage = usage[:, self.perm_end_pt[bucket_id]:]
+        if usage.shape[-1] == 0:
+            # if there is no temporary memory, we don't need to update
+            return
+        self.use_cnt[bucket_id] += usage.view_as(self.use_cnt[bucket_id])
+        self.life_cnt[bucket_id] += 1
+
+    def sieve_by_range(self, bucket_id: int, start: int, end: int, min_size: int) -> None:
+        # keep only the temporary elements *outside* of this range (with some boundary conditions)
+        # the permanent elements are ignored in this computation
+        # i.e., concat (a[:start], a[end:])
+        # bucket with size <= min_size are not modified
+
+        assert start >= 0
+        assert end <= 0
+
+        object_ids = self.buckets[bucket_id]
+        bucket_num_elements = self.k[bucket_id].shape[-1] - self.perm_end_pt[bucket_id]
+        if bucket_num_elements <= min_size:
+            return
+
+        if end == 0:
+            # negative 0 would not work as the end index!
+            # effectively make the second part an empty slice
+            end = self.k[bucket_id].shape[-1] + 1
+
+        p_size = self.perm_end_pt[bucket_id]
+        start = start + p_size
+
+        k = self.k[bucket_id]
+        s = self.s[bucket_id]
+        if self.save_selection:
+            e = self.e[bucket_id]
+        if self.save_usage:
+            use_cnt = self.use_cnt[bucket_id]
+            life_cnt = self.life_cnt[bucket_id]
+
+        self.k[bucket_id] = torch.cat([k[:, :, :start], k[:, :, end:]], -1)
+        self.s[bucket_id] = torch.cat([s[:, :, :start], s[:, :, end:]], -1)
+        if self.save_selection:
+            self.e[bucket_id] = torch.cat([e[:, :, :start - p_size], e[:, :, end:]], -1)
+        if self.save_usage:
+            self.use_cnt[bucket_id] = torch.cat([use_cnt[:, :start - p_size], use_cnt[:, end:]], -1)
+            self.life_cnt[bucket_id] = torch.cat([life_cnt[:, :start - p_size], life_cnt[:, end:]],
+                                                 -1)
+        for obj_id in object_ids:
+            v = self.v[obj_id]
+            self.v[obj_id] = torch.cat([v[:, :, :start], v[:, :, end:]], -1)
+
+    def remove_old_memory(self, bucket_id: int, max_len: int) -> None:
+        self.sieve_by_range(bucket_id, 0, -max_len, max_len)
+
+    def remove_obsolete_features(self, bucket_id: int, max_size: int) -> None:
+        # for long-term memory only
+        object_ids = self.buckets[bucket_id]
+
+        assert self.perm_end_pt[bucket_id] == 0  # permanent memory should be empty in LT memory
+
+        # normalize with life duration
+        usage = self.get_usage(bucket_id)
+        bs = usage.shape[0]
+
+        survivals = []
+
+        for bi in range(bs):
+            _, survived = torch.topk(usage[bi], k=max_size)
+            survivals.append(survived.flatten())
+            assert survived.shape[-1] == survivals[0].shape[-1]
+
+        self.k[bucket_id] = torch.stack(
+            [self.k[bucket_id][bi, :, survived] for bi, survived in enumerate(survivals)], 0)
+        self.s[bucket_id] = torch.stack(
+            [self.s[bucket_id][bi, :, survived] for bi, survived in enumerate(survivals)], 0)
+
+        if self.save_selection:
+            # Long-term memory does not store selection so this should not be needed
+            self.e[bucket_id] = torch.stack(
+                [self.e[bucket_id][bi, :, survived] for bi, survived in enumerate(survivals)], 0)
+        for obj_id in object_ids:
+            self.v[obj_id] = torch.stack(
+                [self.v[obj_id][bi, :, survived] for bi, survived in enumerate(survivals)], 0)
+
+        self.use_cnt[bucket_id] = torch.stack(
+            [self.use_cnt[bucket_id][bi, survived] for bi, survived in enumerate(survivals)], 0)
+        self.life_cnt[bucket_id] = torch.stack(
+            [self.life_cnt[bucket_id][bi, survived] for bi, survived in enumerate(survivals)], 0)
+
+    def get_usage(self, bucket_id: int) -> torch.Tensor:
+        # return normalized usage
+        if not self.save_usage:
+            raise RuntimeError('I did not count usage!')
+        else:
+            usage = self.use_cnt[bucket_id] / self.life_cnt[bucket_id]
+            return usage
+
+    def get_all_sliced(
+        self, bucket_id: int, start: int, end: int
+    ) -> (torch.Tensor, torch.Tensor, torch.Tensor, Dict[int, torch.Tensor], torch.Tensor):
+        # return k, sk, ek, value, normalized usage in order, sliced by start and end
+        # this only queries the temporary memory
+
+        assert start >= 0
+        assert end <= 0
+
+        p_size = self.perm_end_pt[bucket_id]
+        start = start + p_size
+
+        if end == 0:
+            # negative 0 would not work as the end index!
+            k = self.k[bucket_id][:, :, start:]
+            sk = self.s[bucket_id][:, :, start:]
+            ek = self.e[bucket_id][:, :, start - p_size:] if self.save_selection else None
+            value = {obj_id: self.v[obj_id][:, :, start:] for obj_id in self.buckets[bucket_id]}
+            usage = self.get_usage(bucket_id)[:, start - p_size:] if self.save_usage else None
+        else:
+            k = self.k[bucket_id][:, :, start:end]
+            sk = self.s[bucket_id][:, :, start:end]
+            ek = self.e[bucket_id][:, :, start - p_size:end] if self.save_selection else None
+            value = {obj_id: self.v[obj_id][:, :, start:end] for obj_id in self.buckets[bucket_id]}
+            usage = self.get_usage(bucket_id)[:, start - p_size:end] if self.save_usage else None
+
+        return k, sk, ek, value, usage
+
+    def purge_except(self, obj_keep_idx: List[int]):
+        # purge certain objects from the memory except the one listed
+        obj_keep_idx = set(obj_keep_idx)
+
+        # remove objects that are not in the keep list from the buckets
+        buckets_to_remove = []
+        for bucket_id, object_ids in self.buckets.items():
+            self.buckets[bucket_id] = [obj_id for obj_id in object_ids if obj_id in obj_keep_idx]
+            if len(self.buckets[bucket_id]) == 0:
+                buckets_to_remove.append(bucket_id)
+
+        # remove object values that are not in the keep list
+        self.v = {k: v for k, v in self.v.items() if k in obj_keep_idx}
+
+        # remove buckets that are empty
+        for bucket_id in buckets_to_remove:
+            del self.buckets[bucket_id]
+            del self.k[bucket_id]
+            del self.s[bucket_id]
+            if self.save_selection:
+                del self.e[bucket_id]
+            if self.save_usage:
+                del self.use_cnt[bucket_id]
+                del self.life_cnt[bucket_id]
+
+    def clear_non_permanent_memory(self):
+        # clear all non-permanent memory
+        for bucket_id in self.buckets:
+            self.sieve_by_range(bucket_id, 0, 0, 0)
+
+    def get_v_size(self, obj_id: int) -> int:
+        return self.v[obj_id].shape[-1]
+
+    def size(self, bucket_id: int) -> int:
+        if bucket_id not in self.k:
+            return 0
+        else:
+            return self.k[bucket_id].shape[-1]
+
+    def perm_size(self, bucket_id: int) -> int:
+        return self.perm_end_pt[bucket_id]
+
+    def non_perm_size(self, bucket_id: int) -> int:
+        return self.size(bucket_id) - self.perm_size(bucket_id)
+
+    def engaged(self, bucket_id: Optional[int] = None) -> bool:
+        if bucket_id is None:
+            return len(self.buckets) > 0
+        else:
+            return bucket_id in self.buckets
+
+    @property
+    def num_objects(self) -> int:
+        return len(self.v)
+
+    @property
+    def key(self) -> Dict[int, torch.Tensor]:
+        return self.k
+
+    @property
+    def value(self) -> Dict[int, torch.Tensor]:
+        return self.v
+
+    @property
+    def shrinkage(self) -> Dict[int, torch.Tensor]:
+        return self.s
+
+    @property
+    def selection(self) -> Dict[int, torch.Tensor]:
+        return self.e
+
+    def __contains__(self, key):
+        return key in self.v

vendor/MatAnyone2/matanyone2/inference/memory_manager.py

@@ -0,0 +1,453 @@
+import logging
+from omegaconf import DictConfig
+from typing import List, Dict
+import torch
+
+from matanyone2.inference.object_manager import ObjectManager
+from matanyone2.inference.kv_memory_store import KeyValueMemoryStore
+from matanyone2.model.matanyone2 import MatAnyone2
+from matanyone2.model.utils.memory_utils import get_similarity, do_softmax
+
+log = logging.getLogger()
+
+
+class MemoryManager:
+    """
+    Manages all three memory stores and the transition between working/long-term memory
+    """
+    def __init__(self, cfg: DictConfig, object_manager: ObjectManager):
+        self.object_manager = object_manager
+        self.sensory_dim = cfg.model.sensory_dim
+        self.top_k = cfg.top_k
+        self.chunk_size = cfg.chunk_size
+
+        self.save_aux = cfg.save_aux
+
+        self.use_long_term = cfg.use_long_term
+        self.count_long_term_usage = cfg.long_term.count_usage
+        # subtract 1 because the first-frame is now counted as "permanent memory"
+        # and is not counted towards max_mem_frames
+        # but we want to keep the hyperparameters consistent as before for the same behavior
+        if self.use_long_term:
+            self.max_mem_frames = cfg.long_term.max_mem_frames - 1
+            self.min_mem_frames = cfg.long_term.min_mem_frames - 1
+            self.num_prototypes = cfg.long_term.num_prototypes
+            self.max_long_tokens = cfg.long_term.max_num_tokens
+            self.buffer_tokens = cfg.long_term.buffer_tokens
+        else:
+            self.max_mem_frames = cfg.max_mem_frames - 1
+
+        # dimensions will be inferred from input later
+        self.CK = self.CV = None
+        self.H = self.W = None
+
+        # The sensory memory is stored as a dictionary indexed by object ids
+        # each of shape bs * C^h * H * W
+        self.sensory = {}
+
+        # a dictionary indexed by object ids, each of shape bs * T * Q * C
+        self.obj_v = {}
+
+        self.work_mem = KeyValueMemoryStore(save_selection=self.use_long_term,
+                                            save_usage=self.use_long_term)
+        if self.use_long_term:
+            self.long_mem = KeyValueMemoryStore(save_usage=self.count_long_term_usage)
+
+        self.config_stale = True
+        self.engaged = False
+
+    def update_config(self, cfg: DictConfig) -> None:
+        self.config_stale = True
+        self.top_k = cfg['top_k']
+
+        assert self.use_long_term == cfg.use_long_term, 'cannot update this'
+        assert self.count_long_term_usage == cfg.long_term.count_usage, 'cannot update this'
+
+        self.use_long_term = cfg.use_long_term
+        self.count_long_term_usage = cfg.long_term.count_usage
+        if self.use_long_term:
+            self.max_mem_frames = cfg.long_term.max_mem_frames - 1
+            self.min_mem_frames = cfg.long_term.min_mem_frames - 1
+            self.num_prototypes = cfg.long_term.num_prototypes
+            self.max_long_tokens = cfg.long_term.max_num_tokens
+            self.buffer_tokens = cfg.long_term.buffer_tokens
+        else:
+            self.max_mem_frames = cfg.max_mem_frames - 1
+
+    def _readout(self, affinity, v, uncert_mask=None) -> torch.Tensor:
+        # affinity: bs*N*HW
+        # v: bs*C*N or bs*num_objects*C*N
+        # returns bs*C*HW or bs*num_objects*C*HW
+        if len(v.shape) == 3:
+            # single object
+            if uncert_mask is not None:
+                return v @ affinity * uncert_mask
+            else:
+                return v @ affinity
+        else:
+            bs, num_objects, C, N = v.shape
+            v = v.view(bs, num_objects * C, N)
+            out = v @ affinity
+            if uncert_mask is not None:
+                uncert_mask = uncert_mask.flatten(start_dim=2).expand(-1, C, -1)
+                out = out * uncert_mask
+            return out.view(bs, num_objects, C, -1)
+
+    def _get_mask_by_ids(self, mask: torch.Tensor, obj_ids: List[int]) -> torch.Tensor:
+        # -1 because the mask does not contain the background channel
+        return mask[:, [self.object_manager.find_tmp_by_id(obj) - 1 for obj in obj_ids]]
+
+    def _get_sensory_by_ids(self, obj_ids: List[int]) -> torch.Tensor:
+        return torch.stack([self.sensory[obj] for obj in obj_ids], dim=1)
+
+    def _get_object_mem_by_ids(self, obj_ids: List[int]) -> torch.Tensor:
+        return torch.stack([self.obj_v[obj] for obj in obj_ids], dim=1)
+
+    def _get_visual_values_by_ids(self, obj_ids: List[int]) -> torch.Tensor:
+        # All the values that the object ids refer to should have the same shape
+        value = torch.stack([self.work_mem.value[obj] for obj in obj_ids], dim=1)
+        if self.use_long_term and obj_ids[0] in self.long_mem.value:
+            lt_value = torch.stack([self.long_mem.value[obj] for obj in obj_ids], dim=1)
+            value = torch.cat([lt_value, value], dim=-1)
+
+        return value
+    
+    def read_first_frame(self, last_msk_value, pix_feat: torch.Tensor, 
+             last_mask: torch.Tensor, network: MatAnyone2, uncert_output=None) -> Dict[int, torch.Tensor]:
+        """
+        Read from all memory stores and returns a single memory readout tensor for each object
+
+        pix_feat: (1/2) x C x H x W
+        query_key: (1/2) x C^k x H x W
+        selection:  (1/2) x C^k x H x W
+        last_mask: (1/2) x num_objects x H x W (at stride 16)
+        return a dict of memory readouts, indexed by object indices. Each readout is C*H*W
+        """
+        h, w = pix_feat.shape[-2:]
+        bs = pix_feat.shape[0]
+        assert last_mask.shape[0] == bs
+
+        """
+        Compute affinity and perform readout
+        """
+        all_readout_mem = {}
+        buckets = self.work_mem.buckets
+        for bucket_id, bucket in buckets.items():
+
+            if self.chunk_size < 1:
+                object_chunks = [bucket]
+            else:
+                object_chunks = [
+                    bucket[i:i + self.chunk_size] for i in range(0, len(bucket), self.chunk_size)
+                ]
+
+            for objects in object_chunks:
+                this_sensory = self._get_sensory_by_ids(objects)
+                this_last_mask = self._get_mask_by_ids(last_mask, objects)
+                this_msk_value = self._get_visual_values_by_ids(objects)  # (1/2)*num_objects*C*N
+                pixel_readout = network.pixel_fusion(pix_feat, last_msk_value, this_sensory,
+                                                     this_last_mask)
+                this_obj_mem = self._get_object_mem_by_ids(objects).unsqueeze(2)
+                readout_memory, aux_features = network.readout_query(pixel_readout, this_obj_mem)
+                for i, obj in enumerate(objects):
+                    all_readout_mem[obj] = readout_memory[:, i]
+
+                if self.save_aux:
+                    aux_output = {
+                        # 'sensory': this_sensory,
+                        # 'pixel_readout': pixel_readout,
+                        'q_logits': aux_features['logits'] if aux_features else None,
+                        # 'q_weights': aux_features['q_weights'] if aux_features else None,
+                        # 'p_weights': aux_features['p_weights'] if aux_features else None,
+                        # 'attn_mask': aux_features['attn_mask'].float() if aux_features else None,
+                    }
+                    self.aux = aux_output
+
+        return all_readout_mem
+
+    def read(self, pix_feat: torch.Tensor, query_key: torch.Tensor, selection: torch.Tensor,
+             last_mask: torch.Tensor, network: MatAnyone2, uncert_output=None, last_msk_value=None, ti=None,
+             last_pix_feat=None, last_pred_mask=None) -> Dict[int, torch.Tensor]:
+        """
+        Read from all memory stores and returns a single memory readout tensor for each object
+
+        pix_feat: (1/2) x C x H x W
+        query_key: (1/2) x C^k x H x W
+        selection:  (1/2) x C^k x H x W
+        last_mask: (1/2) x num_objects x H x W (at stride 16)
+        return a dict of memory readouts, indexed by object indices. Each readout is C*H*W
+        """
+        h, w = pix_feat.shape[-2:]
+        bs = pix_feat.shape[0]
+        assert query_key.shape[0] == bs
+        assert selection.shape[0] == bs
+        assert last_mask.shape[0] == bs
+
+        uncert_mask = uncert_output["mask"] if uncert_output is not None else None
+
+        query_key = query_key.flatten(start_dim=2)  # bs*C^k*HW
+        selection = selection.flatten(start_dim=2)  # bs*C^k*HW
+        """
+        Compute affinity and perform readout
+        """
+        all_readout_mem = {}
+        buckets = self.work_mem.buckets
+        for bucket_id, bucket in buckets.items():
+            if self.use_long_term and self.long_mem.engaged(bucket_id):
+                # Use long-term memory
+                long_mem_size = self.long_mem.size(bucket_id)
+                memory_key = torch.cat([self.long_mem.key[bucket_id], self.work_mem.key[bucket_id]],
+                                       -1)
+                shrinkage = torch.cat(
+                    [self.long_mem.shrinkage[bucket_id], self.work_mem.shrinkage[bucket_id]], -1)
+
+                similarity = get_similarity(memory_key, shrinkage, query_key, selection)
+                affinity, usage = do_softmax(similarity,
+                                             top_k=self.top_k,
+                                             inplace=True,
+                                             return_usage=True)
+                """
+                Record memory usage for working and long-term memory
+                """
+                # ignore the index return for long-term memory
+                work_usage = usage[:, long_mem_size:]
+                self.work_mem.update_bucket_usage(bucket_id, work_usage)
+
+                if self.count_long_term_usage:
+                    # ignore the index return for working memory
+                    long_usage = usage[:, :long_mem_size]
+                    self.long_mem.update_bucket_usage(bucket_id, long_usage)
+            else:
+                # no long-term memory
+                memory_key = self.work_mem.key[bucket_id]
+                shrinkage = self.work_mem.shrinkage[bucket_id]
+                similarity = get_similarity(memory_key, shrinkage, query_key, selection, uncert_mask=uncert_mask)
+
+                if self.use_long_term:
+                    affinity, usage = do_softmax(similarity,
+                                                 top_k=self.top_k,
+                                                 inplace=True,
+                                                 return_usage=True)
+                    self.work_mem.update_bucket_usage(bucket_id, usage)
+                else:
+                    affinity = do_softmax(similarity, top_k=self.top_k, inplace=True)
+
+            if self.chunk_size < 1:
+                object_chunks = [bucket]
+            else:
+                object_chunks = [
+                    bucket[i:i + self.chunk_size] for i in range(0, len(bucket), self.chunk_size)
+                ]
+
+            for objects in object_chunks:
+                this_sensory = self._get_sensory_by_ids(objects)
+                this_last_mask = self._get_mask_by_ids(last_mask, objects)
+                this_msk_value = self._get_visual_values_by_ids(objects)  # (1/2)*num_objects*C*N
+                visual_readout = self._readout(affinity,
+                                               this_msk_value, uncert_mask).view(bs, len(objects), self.CV, h, w)
+                
+                uncert_output = network.pred_uncertainty(last_pix_feat, pix_feat, last_pred_mask, visual_readout[:,0]-last_msk_value[:,0])
+
+                if uncert_output is not None:
+                    uncert_prob = uncert_output["prob"].unsqueeze(1) # b n 1 h w
+                    visual_readout = visual_readout*uncert_prob + last_msk_value*(1-uncert_prob)
+
+                pixel_readout = network.pixel_fusion(pix_feat, visual_readout, this_sensory,
+                                                     this_last_mask)
+                this_obj_mem = self._get_object_mem_by_ids(objects).unsqueeze(2)
+                readout_memory, aux_features = network.readout_query(pixel_readout, this_obj_mem)
+                for i, obj in enumerate(objects):
+                    all_readout_mem[obj] = readout_memory[:, i]
+
+                if self.save_aux:
+                    aux_output = {
+                        # 'sensory': this_sensory,
+                        # 'pixel_readout': pixel_readout,
+                        'q_logits': aux_features['logits'] if aux_features else None,
+                        # 'q_weights': aux_features['q_weights'] if aux_features else None,
+                        # 'p_weights': aux_features['p_weights'] if aux_features else None,
+                        # 'attn_mask': aux_features['attn_mask'].float() if aux_features else None,
+                    }
+                    self.aux = aux_output
+
+        return all_readout_mem
+
+    def add_memory(self,
+                   key: torch.Tensor,
+                   shrinkage: torch.Tensor,
+                   msk_value: torch.Tensor,
+                   obj_value: torch.Tensor,
+                   objects: List[int],
+                   selection: torch.Tensor = None,
+                   *,
+                   as_permanent: bool = False) -> None:
+        # key: (1/2)*C*H*W
+        # msk_value: (1/2)*num_objects*C*H*W
+        # obj_value: (1/2)*num_objects*Q*C
+        # objects contains a list of object ids corresponding to the objects in msk_value/obj_value
+        bs = key.shape[0]
+        assert shrinkage.shape[0] == bs
+        assert msk_value.shape[0] == bs
+        assert obj_value.shape[0] == bs
+
+        self.engaged = True
+        if self.H is None or self.config_stale:
+            self.config_stale = False
+            self.H, self.W = msk_value.shape[-2:]
+            self.HW = self.H * self.W
+            # convert from num. frames to num. tokens
+            self.max_work_tokens = self.max_mem_frames * self.HW
+            if self.use_long_term:
+                self.min_work_tokens = self.min_mem_frames * self.HW
+
+        # key:   bs*C*N
+        # value: bs*num_objects*C*N
+        key = key.flatten(start_dim=2)
+        shrinkage = shrinkage.flatten(start_dim=2)
+        self.CK = key.shape[1]
+
+        msk_value = msk_value.flatten(start_dim=3)
+        self.CV = msk_value.shape[2]
+
+        if selection is not None:
+            # not used in non-long-term mode
+            selection = selection.flatten(start_dim=2)
+
+        # insert object values into object memory
+        for obj_id, obj in enumerate(objects):
+            if obj in self.obj_v:
+                """streaming average
+                each self.obj_v[obj] is (1/2)*num_summaries*(embed_dim+1)
+                first embed_dim keeps track of the sum of embeddings
+                the last dim keeps the total count
+                averaging in done inside the object transformer
+
+                incoming obj_value is (1/2)*num_objects*num_summaries*(embed_dim+1)
+                self.obj_v[obj] = torch.cat([self.obj_v[obj], obj_value[:, obj_id]], dim=0)
+                """
+                last_acc = self.obj_v[obj][:, :, -1]
+                new_acc = last_acc + obj_value[:, obj_id, :, -1]
+
+                self.obj_v[obj][:, :, :-1] = (self.obj_v[obj][:, :, :-1] +
+                                              obj_value[:, obj_id, :, :-1])
+                self.obj_v[obj][:, :, -1] = new_acc
+            else:
+                self.obj_v[obj] = obj_value[:, obj_id]
+
+        # convert mask value tensor into a dict for insertion
+        msk_values = {obj: msk_value[:, obj_id] for obj_id, obj in enumerate(objects)}
+        self.work_mem.add(key,
+                          msk_values,
+                          shrinkage,
+                          selection=selection,
+                          as_permanent=as_permanent)
+
+        for bucket_id in self.work_mem.buckets.keys():
+            # long-term memory cleanup
+            if self.use_long_term:
+                # Do memory compressed if needed
+                if self.work_mem.non_perm_size(bucket_id) >= self.max_work_tokens:
+                    # Remove obsolete features if needed
+                    if self.long_mem.non_perm_size(bucket_id) >= (self.max_long_tokens -
+                                                         self.num_prototypes):
+                        self.long_mem.remove_obsolete_features(
+                            bucket_id,
+                            self.max_long_tokens - self.num_prototypes - self.buffer_tokens)
+
+                    self.compress_features(bucket_id)
+            else:
+                # FIFO
+                self.work_mem.remove_old_memory(bucket_id, self.max_work_tokens)
+
+    def purge_except(self, obj_keep_idx: List[int]) -> None:
+        # purge certain objects from the memory except the one listed
+        self.work_mem.purge_except(obj_keep_idx)
+        if self.use_long_term and self.long_mem.engaged():
+            self.long_mem.purge_except(obj_keep_idx)
+        self.sensory = {k: v for k, v in self.sensory.items() if k in obj_keep_idx}
+
+        if not self.work_mem.engaged():
+            # everything is removed!
+            self.engaged = False
+
+    def compress_features(self, bucket_id: int) -> None:
+
+        # perform memory consolidation
+        prototype_key, prototype_value, prototype_shrinkage = self.consolidation(
+            *self.work_mem.get_all_sliced(bucket_id, 0, -self.min_work_tokens))
+
+        # remove consolidated working memory
+        self.work_mem.sieve_by_range(bucket_id,
+                                     0,
+                                     -self.min_work_tokens,
+                                     min_size=self.min_work_tokens)
+
+        # add to long-term memory
+        self.long_mem.add(prototype_key,
+                          prototype_value,
+                          prototype_shrinkage,
+                          selection=None,
+                          supposed_bucket_id=bucket_id)
+
+    def consolidation(self, candidate_key: torch.Tensor, candidate_shrinkage: torch.Tensor,
+                      candidate_selection: torch.Tensor, candidate_value: Dict[int, torch.Tensor],
+                      usage: torch.Tensor) -> (torch.Tensor, Dict[int, torch.Tensor], torch.Tensor):
+        # find the indices with max usage
+        bs = candidate_key.shape[0]
+        assert bs in [1, 2]
+
+        prototype_key = []
+        prototype_selection = []
+        for bi in range(bs):
+            _, max_usage_indices = torch.topk(usage[bi], k=self.num_prototypes, dim=-1, sorted=True)
+            prototype_indices = max_usage_indices.flatten()
+            prototype_key.append(candidate_key[bi, :, prototype_indices])
+            prototype_selection.append(candidate_selection[bi, :, prototype_indices])
+        prototype_key = torch.stack(prototype_key, dim=0)
+        prototype_selection = torch.stack(prototype_selection, dim=0)
+        """
+        Potentiation step
+        """
+        similarity = get_similarity(candidate_key, candidate_shrinkage, prototype_key,
+                                    prototype_selection)
+        affinity = do_softmax(similarity)
+
+        # readout the values
+        prototype_value = {k: self._readout(affinity, v) for k, v in candidate_value.items()}
+
+        # readout the shrinkage term
+        prototype_shrinkage = self._readout(affinity, candidate_shrinkage)
+
+        return prototype_key, prototype_value, prototype_shrinkage
+
+    def initialize_sensory_if_needed(self, sample_key: torch.Tensor, ids: List[int]):
+        for obj in ids:
+            if obj not in self.sensory:
+                # also initializes the sensory memory
+                bs, _, h, w = sample_key.shape
+                self.sensory[obj] = torch.zeros((bs, self.sensory_dim, h, w),
+                                                device=sample_key.device)
+
+    def update_sensory(self, sensory: torch.Tensor, ids: List[int]):
+        # sensory: 1*num_objects*C*H*W
+        for obj_id, obj in enumerate(ids):
+            self.sensory[obj] = sensory[:, obj_id]
+
+    def get_sensory(self, ids: List[int]):
+        # returns (1/2)*num_objects*C*H*W
+        return self._get_sensory_by_ids(ids)
+    
+    def clear_non_permanent_memory(self):
+        self.work_mem.clear_non_permanent_memory()
+        if self.use_long_term:
+            self.long_mem.clear_non_permanent_memory()
+
+    def clear_sensory_memory(self):
+        self.sensory = {}
+    
+    def clear_work_mem(self):
+        self.work_mem = KeyValueMemoryStore(save_selection=self.use_long_term,
+                                            save_usage=self.use_long_term)
+    
+    def clear_obj_mem(self):
+        self.obj_v = {}

vendor/MatAnyone2/matanyone2/inference/object_info.py

@@ -0,0 +1,24 @@
+class ObjectInfo:
+    """
+    Store meta information for an object
+    """
+    def __init__(self, id: int):
+        self.id = id
+        self.poke_count = 0  # count number of detections missed
+
+    def poke(self) -> None:
+        self.poke_count += 1
+
+    def unpoke(self) -> None:
+        self.poke_count = 0
+
+    def __hash__(self):
+        return hash(self.id)
+
+    def __eq__(self, other):
+        if type(other) == int:
+            return self.id == other
+        return self.id == other.id
+
+    def __repr__(self):
+        return f'(ID: {self.id})'

vendor/MatAnyone2/matanyone2/inference/object_manager.py

@@ -0,0 +1,149 @@
+from typing import Union, List, Dict
+
+import torch
+from matanyone2.inference.object_info import ObjectInfo
+
+
+class ObjectManager:
+    """
+    Object IDs are immutable. The same ID always represent the same object.
+    Temporary IDs are the positions of each object in the tensor. It changes as objects get removed.
+    Temporary IDs start from 1.
+    """
+
+    def __init__(self):
+        self.obj_to_tmp_id: Dict[ObjectInfo, int] = {}
+        self.tmp_id_to_obj: Dict[int, ObjectInfo] = {}
+        self.obj_id_to_obj: Dict[int, ObjectInfo] = {}
+
+        self.all_historical_object_ids: List[int] = []
+
+    def _recompute_obj_id_to_obj_mapping(self) -> None:
+        self.obj_id_to_obj = {obj.id: obj for obj in self.obj_to_tmp_id}
+
+    def add_new_objects(
+            self, objects: Union[List[ObjectInfo], ObjectInfo,
+                                 List[int]]) -> (List[int], List[int]):
+        if not isinstance(objects, list):
+            objects = [objects]
+
+        corresponding_tmp_ids = []
+        corresponding_obj_ids = []
+        for obj in objects:
+            if isinstance(obj, int):
+                obj = ObjectInfo(id=obj)
+
+            if obj in self.obj_to_tmp_id:
+                # old object
+                corresponding_tmp_ids.append(self.obj_to_tmp_id[obj])
+                corresponding_obj_ids.append(obj.id)
+            else:
+                # new object
+                new_obj = ObjectInfo(id=obj.id)
+
+                # new object
+                new_tmp_id = len(self.obj_to_tmp_id) + 1
+                self.obj_to_tmp_id[new_obj] = new_tmp_id
+                self.tmp_id_to_obj[new_tmp_id] = new_obj
+                self.all_historical_object_ids.append(new_obj.id)
+                corresponding_tmp_ids.append(new_tmp_id)
+                corresponding_obj_ids.append(new_obj.id)
+
+        self._recompute_obj_id_to_obj_mapping()
+        assert corresponding_tmp_ids == sorted(corresponding_tmp_ids)
+        return corresponding_tmp_ids, corresponding_obj_ids
+
+    def delete_objects(self, obj_ids_to_remove: Union[int, List[int]]) -> None:
+        # delete an object or a list of objects
+        # re-sort the tmp ids
+        if isinstance(obj_ids_to_remove, int):
+            obj_ids_to_remove = [obj_ids_to_remove]
+
+        new_tmp_id = 1
+        total_num_id = len(self.obj_to_tmp_id)
+
+        local_obj_to_tmp_id = {}
+        local_tmp_to_obj_id = {}
+
+        for tmp_iter in range(1, total_num_id + 1):
+            obj = self.tmp_id_to_obj[tmp_iter]
+            if obj.id not in obj_ids_to_remove:
+                local_obj_to_tmp_id[obj] = new_tmp_id
+                local_tmp_to_obj_id[new_tmp_id] = obj
+                new_tmp_id += 1
+
+        self.obj_to_tmp_id = local_obj_to_tmp_id
+        self.tmp_id_to_obj = local_tmp_to_obj_id
+        self._recompute_obj_id_to_obj_mapping()
+
+    def purge_inactive_objects(self,
+                               max_missed_detection_count: int) -> (bool, List[int], List[int]):
+        # remove tmp ids of objects that are removed
+        obj_id_to_be_deleted = []
+        tmp_id_to_be_deleted = []
+        tmp_id_to_keep = []
+        obj_id_to_keep = []
+
+        for obj in self.obj_to_tmp_id:
+            if obj.poke_count > max_missed_detection_count:
+                obj_id_to_be_deleted.append(obj.id)
+                tmp_id_to_be_deleted.append(self.obj_to_tmp_id[obj])
+            else:
+                tmp_id_to_keep.append(self.obj_to_tmp_id[obj])
+                obj_id_to_keep.append(obj.id)
+
+        purge_activated = len(obj_id_to_be_deleted) > 0
+        if purge_activated:
+            self.delete_objects(obj_id_to_be_deleted)
+        return purge_activated, tmp_id_to_keep, obj_id_to_keep
+
+    def tmp_to_obj_cls(self, mask) -> torch.Tensor:
+        # remap tmp id cls representation to the true object id representation
+        new_mask = torch.zeros_like(mask)
+        for tmp_id, obj in self.tmp_id_to_obj.items():
+            new_mask[mask == tmp_id] = obj.id
+        return new_mask
+
+    def get_tmp_to_obj_mapping(self) -> Dict[int, ObjectInfo]:
+        # returns the mapping in a dict format for saving it with pickle
+        return {obj.id: tmp_id for obj, tmp_id in self.tmp_id_to_obj.items()}
+
+    def realize_dict(self, obj_dict, dim=1) -> torch.Tensor:
+        # turns a dict indexed by obj id into a tensor, ordered by tmp IDs
+        output = []
+        for _, obj in self.tmp_id_to_obj.items():
+            if obj.id not in obj_dict:
+                raise NotImplementedError
+            output.append(obj_dict[obj.id])
+        output = torch.stack(output, dim=dim)
+        return output
+
+    def make_one_hot(self, cls_mask) -> torch.Tensor:
+        output = []
+        for _, obj in self.tmp_id_to_obj.items():
+            output.append(cls_mask == obj.id)
+        if len(output) == 0:
+            output = torch.zeros((0, *cls_mask.shape), dtype=torch.bool, device=cls_mask.device)
+        else:
+            output = torch.stack(output, dim=0)
+        return output
+
+    @property
+    def all_obj_ids(self) -> List[int]:
+        return [k.id for k in self.obj_to_tmp_id]
+
+    @property
+    def num_obj(self) -> int:
+        return len(self.obj_to_tmp_id)
+
+    def has_all(self, objects: List[int]) -> bool:
+        for obj in objects:
+            if obj not in self.obj_to_tmp_id:
+                return False
+        return True
+
+    def find_object_by_id(self, obj_id) -> ObjectInfo:
+        return self.obj_id_to_obj[obj_id]
+
+    def find_tmp_by_id(self, obj_id) -> int:
+        return self.obj_to_tmp_id[self.obj_id_to_obj[obj_id]]

vendor/MatAnyone2/matanyone2/inference/utils/args_utils.py

@@ -0,0 +1,30 @@
+import logging
+from omegaconf import DictConfig
+
+log = logging.getLogger()
+
+
+def get_dataset_cfg(cfg: DictConfig):
+    dataset_name = cfg.dataset
+    data_cfg = cfg.datasets[dataset_name]
+
+    potential_overrides = [
+        'image_directory',
+        'mask_directory',
+        'json_directory',
+        'size',
+        'save_all',
+        'use_all_masks',
+        'use_long_term',
+        'mem_every',
+    ]
+
+    for override in potential_overrides:
+        if cfg[override] is not None:
+            log.info(f'Overriding config {override} from {data_cfg[override]} to {cfg[override]}')
+            data_cfg[override] = cfg[override]
+        # escalte all potential overrides to the top-level config
+        if override in data_cfg:
+            cfg[override] = data_cfg[override]
+
+    return data_cfg

vendor/MatAnyone2/matanyone2/model/aux_modules.py

@@ -0,0 +1,93 @@
+"""
+For computing auxiliary outputs for auxiliary losses
+"""
+from typing import Dict
+from omegaconf import DictConfig
+import torch
+import torch.nn as nn
+
+from matanyone2.model.group_modules import GConv2d
+from matanyone2.utils.tensor_utils import aggregate
+
+
+class LinearPredictor(nn.Module):
+    def __init__(self, x_dim: int, pix_dim: int):
+        super().__init__()
+        self.projection = GConv2d(x_dim, pix_dim + 1, kernel_size=1)
+
+    def forward(self, pix_feat: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
+        # pixel_feat: B*pix_dim*H*W
+        # x: B*num_objects*x_dim*H*W
+        num_objects = x.shape[1]
+        x = self.projection(x)
+
+        pix_feat = pix_feat.unsqueeze(1).expand(-1, num_objects, -1, -1, -1)
+        logits = (pix_feat * x[:, :, :-1]).sum(dim=2) + x[:, :, -1]
+        return logits
+
+
+class DirectPredictor(nn.Module):
+    def __init__(self, x_dim: int):
+        super().__init__()
+        self.projection = GConv2d(x_dim, 1, kernel_size=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # x: B*num_objects*x_dim*H*W
+        logits = self.projection(x).squeeze(2)
+        return logits
+
+
+class AuxComputer(nn.Module):
+    def __init__(self, cfg: DictConfig):
+        super().__init__()
+
+        use_sensory_aux = cfg.model.aux_loss.sensory.enabled
+        self.use_query_aux = cfg.model.aux_loss.query.enabled
+        self.use_sensory_aux = use_sensory_aux
+
+        sensory_dim = cfg.model.sensory_dim
+        embed_dim = cfg.model.embed_dim
+
+        if use_sensory_aux:  
+            self.sensory_aux = LinearPredictor(sensory_dim, embed_dim)
+
+    def _aggregate_with_selector(self, logits: torch.Tensor, selector: torch.Tensor) -> torch.Tensor:
+        prob = torch.sigmoid(logits)
+        if selector is not None:
+            prob = prob * selector
+        logits = aggregate(prob, dim=1)
+        return logits
+
+    def forward(self, pix_feat: torch.Tensor, aux_input: Dict[str, torch.Tensor],
+                selector: torch.Tensor, seg_pass=False) -> Dict[str, torch.Tensor]:
+        sensory = aux_input['sensory']
+        q_logits = aux_input['q_logits']
+
+        aux_output = {}
+        aux_output['attn_mask'] = aux_input['attn_mask']
+
+        if self.use_sensory_aux:
+            # B*num_objects*H*W
+            logits = self.sensory_aux(pix_feat, sensory)
+            aux_output['sensory_logits'] = self._aggregate_with_selector(logits, selector)
+        if self.use_query_aux:
+            # B*num_objects*num_levels*H*W
+            aux_output['q_logits'] = self._aggregate_with_selector(
+                torch.stack(q_logits, dim=2),
+                selector.unsqueeze(2) if selector is not None else None)
+
+        return aux_output
+    
+    def compute_mask(self, aux_input: Dict[str, torch.Tensor],
+                selector: torch.Tensor) -> Dict[str, torch.Tensor]:
+        # sensory = aux_input['sensory']
+        q_logits = aux_input['q_logits']
+
+        aux_output = {}
+
+        # B*num_objects*num_levels*H*W
+        aux_output['q_logits'] = self._aggregate_with_selector(
+            torch.stack(q_logits, dim=2),
+            selector.unsqueeze(2) if selector is not None else None)
+
+        return aux_output

vendor/MatAnyone2/matanyone2/model/big_modules.py

@@ -0,0 +1,366 @@
+"""
+big_modules.py - This file stores higher-level network blocks.
+
+x - usually denotes features that are shared between objects.
+g - usually denotes features that are not shared between objects 
+    with an extra "num_objects" dimension (batch_size * num_objects * num_channels * H * W).
+
+The trailing number of a variable usually denotes the stride
+"""
+
+from typing import Iterable
+from omegaconf import DictConfig
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from matanyone2.model.group_modules import MainToGroupDistributor, GroupFeatureFusionBlock, GConv2d
+from matanyone2.model.utils import resnet
+from matanyone2.model.modules import SensoryDeepUpdater, SensoryUpdater_fullscale, DecoderFeatureProcessor, MaskUpsampleBlock
+from matanyone2.utils.device import safe_autocast
+
+class UncertPred(nn.Module):
+    def __init__(self, model_cfg: DictConfig):
+        super().__init__()
+        self.conv1x1_v2 = nn.Conv2d(model_cfg.pixel_dim*2 + 1 + model_cfg.value_dim, 64, kernel_size=1, stride=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv3x3 = nn.Conv2d(64, 32, kernel_size=3, stride=1, padding=1, groups=1, bias=False, dilation=1)
+        self.bn2 = nn.BatchNorm2d(32)
+        self.conv3x3_out = nn.Conv2d(32, 1, kernel_size=3, stride=1, padding=1, groups=1, bias=False, dilation=1)
+    
+    def forward(self, last_frame_feat: torch.Tensor, cur_frame_feat: torch.Tensor, last_mask: torch.Tensor, mem_val_diff:torch.Tensor):
+        last_mask = F.interpolate(last_mask, size=last_frame_feat.shape[-2:], mode='area')
+        x = torch.cat([last_frame_feat, cur_frame_feat, last_mask, mem_val_diff], dim=1)
+        x = self.conv1x1_v2(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.conv3x3(x)
+        x = self.bn2(x)
+        x = self.relu(x)
+        x = self.conv3x3_out(x)
+        return x
+    
+    # override the default train() to freeze BN statistics
+    def train(self, mode=True):
+        self.training = False
+        for module in self.children():
+            module.train(False)
+        return self
+
+class PixelEncoder(nn.Module):
+    def __init__(self, model_cfg: DictConfig):
+        super().__init__()
+
+        self.is_resnet = 'resnet' in model_cfg.pixel_encoder.type
+        # if model_cfg.pretrained_resnet is set in the model_cfg we get the value
+        # else default to True
+        is_pretrained_resnet = getattr(model_cfg,"pretrained_resnet",True)
+        if self.is_resnet:
+            if model_cfg.pixel_encoder.type == 'resnet18':
+                network = resnet.resnet18(pretrained=is_pretrained_resnet)
+            elif model_cfg.pixel_encoder.type == 'resnet50':
+                network = resnet.resnet50(pretrained=is_pretrained_resnet)
+            else:
+                raise NotImplementedError
+            self.conv1 = network.conv1
+            self.bn1 = network.bn1
+            self.relu = network.relu
+            self.maxpool = network.maxpool
+
+            self.res2 = network.layer1
+            self.layer2 = network.layer2
+            self.layer3 = network.layer3
+        else:
+            raise NotImplementedError
+
+    def forward(self, x: torch.Tensor, seq_length=None) -> (torch.Tensor, torch.Tensor, torch.Tensor):
+        f1 = x
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        f2 = x
+        x = self.maxpool(x)
+        f4 = self.res2(x)
+        f8 = self.layer2(f4)
+        f16 = self.layer3(f8)
+
+        return f16, f8, f4, f2, f1
+
+    # override the default train() to freeze BN statistics
+    def train(self, mode=True):
+        self.training = False
+        for module in self.children():
+            module.train(False)
+        return self
+
+
+class KeyProjection(nn.Module):
+    def __init__(self, model_cfg: DictConfig):
+        super().__init__()
+        in_dim = model_cfg.pixel_encoder.ms_dims[0]
+        mid_dim = model_cfg.pixel_dim
+        key_dim = model_cfg.key_dim
+
+        self.pix_feat_proj = nn.Conv2d(in_dim, mid_dim, kernel_size=1)
+        self.key_proj = nn.Conv2d(mid_dim, key_dim, kernel_size=3, padding=1)
+        # shrinkage
+        self.d_proj = nn.Conv2d(mid_dim, 1, kernel_size=3, padding=1)
+        # selection
+        self.e_proj = nn.Conv2d(mid_dim, key_dim, kernel_size=3, padding=1)
+
+        nn.init.orthogonal_(self.key_proj.weight.data)
+        nn.init.zeros_(self.key_proj.bias.data)
+
+    def forward(self, x: torch.Tensor, *, need_s: bool,
+                need_e: bool) -> (torch.Tensor, torch.Tensor, torch.Tensor):
+        x = self.pix_feat_proj(x)
+        shrinkage = self.d_proj(x)**2 + 1 if (need_s) else None
+        selection = torch.sigmoid(self.e_proj(x)) if (need_e) else None
+
+        return self.key_proj(x), shrinkage, selection
+
+
+class MaskEncoder(nn.Module):
+    def __init__(self, model_cfg: DictConfig, single_object=False):
+        super().__init__()
+        pixel_dim = model_cfg.pixel_dim
+        value_dim = model_cfg.value_dim
+        sensory_dim = model_cfg.sensory_dim
+        final_dim = model_cfg.mask_encoder.final_dim
+
+        self.single_object = single_object
+        extra_dim = 1 if single_object else 2
+
+        # if model_cfg.pretrained_resnet is set in the model_cfg we get the value
+        # else default to True
+        is_pretrained_resnet = getattr(model_cfg,"pretrained_resnet",True)
+        if model_cfg.mask_encoder.type == 'resnet18':
+            network = resnet.resnet18(pretrained=is_pretrained_resnet, extra_dim=extra_dim)
+        elif model_cfg.mask_encoder.type == 'resnet50':
+            network = resnet.resnet50(pretrained=is_pretrained_resnet, extra_dim=extra_dim)
+        else:
+            raise NotImplementedError
+        self.conv1 = network.conv1
+        self.bn1 = network.bn1
+        self.relu = network.relu
+        self.maxpool = network.maxpool
+
+        self.layer1 = network.layer1
+        self.layer2 = network.layer2
+        self.layer3 = network.layer3
+
+        self.distributor = MainToGroupDistributor()
+        self.fuser = GroupFeatureFusionBlock(pixel_dim, final_dim, value_dim)
+
+        self.sensory_update = SensoryDeepUpdater(value_dim, sensory_dim)
+
+    def forward(self,
+                image: torch.Tensor,
+                pix_feat: torch.Tensor,
+                sensory: torch.Tensor,
+                masks: torch.Tensor,
+                others: torch.Tensor,
+                *,
+                deep_update: bool = True,
+                chunk_size: int = -1) -> (torch.Tensor, torch.Tensor):
+        # ms_features are from the key encoder
+        # we only use the first one (lowest resolution), following XMem
+        if self.single_object:
+            g = masks.unsqueeze(2)
+        else:
+            g = torch.stack([masks, others], dim=2)
+
+        g = self.distributor(image, g)
+
+        batch_size, num_objects = g.shape[:2]
+        if chunk_size < 1 or chunk_size >= num_objects:
+            chunk_size = num_objects
+            fast_path = True
+            new_sensory = sensory
+        else:
+            if deep_update:
+                new_sensory = torch.empty_like(sensory)
+            else:
+                new_sensory = sensory
+            fast_path = False
+
+        # chunk-by-chunk inference
+        all_g = []
+        for i in range(0, num_objects, chunk_size):
+            if fast_path:
+                g_chunk = g
+            else:
+                g_chunk = g[:, i:i + chunk_size]
+            actual_chunk_size = g_chunk.shape[1]
+            g_chunk = g_chunk.flatten(start_dim=0, end_dim=1)
+
+            g_chunk = self.conv1(g_chunk)
+            g_chunk = self.bn1(g_chunk)      # 1/2, 64
+            g_chunk = self.maxpool(g_chunk)  # 1/4, 64
+            g_chunk = self.relu(g_chunk)
+
+            g_chunk = self.layer1(g_chunk)   # 1/4
+            g_chunk = self.layer2(g_chunk)   # 1/8
+            g_chunk = self.layer3(g_chunk)   # 1/16
+
+            g_chunk = g_chunk.view(batch_size, actual_chunk_size, *g_chunk.shape[1:])
+            g_chunk = self.fuser(pix_feat, g_chunk)
+            all_g.append(g_chunk)
+            if deep_update:
+                if fast_path:
+                    new_sensory = self.sensory_update(g_chunk, sensory)
+                else:
+                    new_sensory[:, i:i + chunk_size] = self.sensory_update(
+                        g_chunk, sensory[:, i:i + chunk_size])
+        g = torch.cat(all_g, dim=1)
+
+        return g, new_sensory
+
+    # override the default train() to freeze BN statistics
+    def train(self, mode=True):
+        self.training = False
+        for module in self.children():
+            module.train(False)
+        return self
+
+
+class PixelFeatureFuser(nn.Module):
+    def __init__(self, model_cfg: DictConfig, single_object=False):
+        super().__init__()
+        value_dim = model_cfg.value_dim
+        sensory_dim = model_cfg.sensory_dim
+        pixel_dim = model_cfg.pixel_dim
+        embed_dim = model_cfg.embed_dim
+        self.single_object = single_object
+
+        self.fuser = GroupFeatureFusionBlock(pixel_dim, value_dim, embed_dim)
+        if self.single_object:
+            self.sensory_compress = GConv2d(sensory_dim + 1, value_dim, kernel_size=1)
+        else:
+            self.sensory_compress = GConv2d(sensory_dim + 2, value_dim, kernel_size=1)
+
+    def forward(self,
+                pix_feat: torch.Tensor,
+                pixel_memory: torch.Tensor,
+                sensory_memory: torch.Tensor,
+                last_mask: torch.Tensor,
+                last_others: torch.Tensor,
+                *,
+                chunk_size: int = -1) -> torch.Tensor:
+        batch_size, num_objects = pixel_memory.shape[:2]
+
+        if self.single_object:
+            last_mask = last_mask.unsqueeze(2)
+        else:
+            last_mask = torch.stack([last_mask, last_others], dim=2)
+
+        if chunk_size < 1:
+            chunk_size = num_objects
+
+        # chunk-by-chunk inference
+        all_p16 = []
+        for i in range(0, num_objects, chunk_size):
+            sensory_readout = self.sensory_compress(
+                torch.cat([sensory_memory[:, i:i + chunk_size], last_mask[:, i:i + chunk_size]], 2))
+            p16 = pixel_memory[:, i:i + chunk_size] + sensory_readout
+            p16 = self.fuser(pix_feat, p16)
+            all_p16.append(p16)
+        p16 = torch.cat(all_p16, dim=1)
+
+        return p16
+
+
+class MaskDecoder(nn.Module):
+    def __init__(self, model_cfg: DictConfig):
+        super().__init__()
+        embed_dim = model_cfg.embed_dim
+        sensory_dim = model_cfg.sensory_dim
+        ms_image_dims = model_cfg.pixel_encoder.ms_dims
+        up_dims = model_cfg.mask_decoder.up_dims
+
+        assert embed_dim == up_dims[0]
+
+        self.sensory_update = SensoryUpdater_fullscale([up_dims[0], up_dims[1], up_dims[2], up_dims[3], up_dims[4] + 1], sensory_dim,
+                                             sensory_dim)
+
+        self.decoder_feat_proc = DecoderFeatureProcessor(ms_image_dims[1:], up_dims[:-1])
+        self.up_16_8 = MaskUpsampleBlock(up_dims[0], up_dims[1])
+        self.up_8_4 = MaskUpsampleBlock(up_dims[1], up_dims[2])
+        # newly add for alpha matte
+        self.up_4_2 = MaskUpsampleBlock(up_dims[2], up_dims[3])
+        self.up_2_1 = MaskUpsampleBlock(up_dims[3], up_dims[4])
+
+        self.pred_seg = nn.Conv2d(up_dims[-1], 1, kernel_size=3, padding=1)
+        self.pred_mat = nn.Conv2d(up_dims[-1], 1, kernel_size=3, padding=1)
+
+    def forward(self,
+                ms_image_feat: Iterable[torch.Tensor],
+                memory_readout: torch.Tensor,
+                sensory: torch.Tensor,
+                *,
+                chunk_size: int = -1,
+                update_sensory: bool = True,
+                seg_pass: bool = False,
+                last_mask=None,
+                sigmoid_residual=False) -> (torch.Tensor, torch.Tensor):
+
+        batch_size, num_objects = memory_readout.shape[:2]
+        f8, f4, f2, f1 = self.decoder_feat_proc(ms_image_feat[1:])
+        if chunk_size < 1 or chunk_size >= num_objects:
+            chunk_size = num_objects
+            fast_path = True
+            new_sensory = sensory
+        else:
+            if update_sensory:
+                new_sensory = torch.empty_like(sensory)
+            else:
+                new_sensory = sensory
+            fast_path = False
+
+        # chunk-by-chunk inference
+        all_logits = []
+        for i in range(0, num_objects, chunk_size):
+            if fast_path:
+                p16 = memory_readout
+            else:
+                p16 = memory_readout[:, i:i + chunk_size]
+            actual_chunk_size = p16.shape[1]
+
+            p8 = self.up_16_8(p16, f8)
+            p4 = self.up_8_4(p8, f4)
+            p2 = self.up_4_2(p4, f2)
+            p1 = self.up_2_1(p2, f1)
+            with safe_autocast(enabled=False):
+                if seg_pass:
+                    if last_mask is not None:
+                        res = self.pred_seg(F.relu(p1.flatten(start_dim=0, end_dim=1).float()))
+                        if sigmoid_residual:
+                            res = (torch.sigmoid(res) - 0.5) * 2  # regularization: (-1, 1) change on last mask
+                        logits = last_mask + res
+                    else:
+                        logits = self.pred_seg(F.relu(p1.flatten(start_dim=0, end_dim=1).float()))
+                else:
+                    if last_mask is not None:
+                        res = self.pred_mat(F.relu(p1.flatten(start_dim=0, end_dim=1).float()))
+                        if sigmoid_residual:
+                            res = (torch.sigmoid(res) - 0.5) * 2  # regularization: (-1, 1) change on last mask
+                        logits = last_mask + res
+                    else:
+                        logits = self.pred_mat(F.relu(p1.flatten(start_dim=0, end_dim=1).float()))
+            ## SensoryUpdater_fullscale
+            if update_sensory:
+                p1 = torch.cat(
+                    [p1, logits.view(batch_size, actual_chunk_size, 1, *logits.shape[-2:])], 2)
+                if fast_path:
+                    new_sensory = self.sensory_update([p16, p8, p4, p2, p1], sensory)
+                else:
+                    new_sensory[:,
+                                i:i + chunk_size] = self.sensory_update([p16, p8, p4, p2, p1],
+                                                                        sensory[:,
+                                                                                i:i + chunk_size])
+            all_logits.append(logits)
+        logits = torch.cat(all_logits, dim=0)
+        logits = logits.view(batch_size, num_objects, *logits.shape[-2:])
+
+        return new_sensory, logits

vendor/MatAnyone2/matanyone2/model/channel_attn.py

@@ -0,0 +1,39 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class CAResBlock(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, residual: bool = True):
+        super().__init__()
+        self.residual = residual
+        self.conv1 = nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1)
+        self.conv2 = nn.Conv2d(out_dim, out_dim, kernel_size=3, padding=1)
+
+        t = int((abs(math.log2(out_dim)) + 1) // 2)
+        k = t if t % 2 else t + 1
+        self.pool = nn.AdaptiveAvgPool2d(1)
+        self.conv = nn.Conv1d(1, 1, kernel_size=k, padding=(k - 1) // 2, bias=False)
+
+        if self.residual:
+            if in_dim == out_dim:
+                self.downsample = nn.Identity()
+            else:
+                self.downsample = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r = x
+        x = self.conv1(F.relu(x))
+        x = self.conv2(F.relu(x))
+
+        b, c = x.shape[:2]
+        w = self.pool(x).view(b, 1, c)
+        w = self.conv(w).transpose(-1, -2).unsqueeze(-1).sigmoid()  # B*C*1*1
+
+        if self.residual:
+            x = x * w + self.downsample(r)
+        else:
+            x = x * w
+
+        return x

vendor/MatAnyone2/matanyone2/model/group_modules.py

@@ -0,0 +1,126 @@
+from typing import Optional
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from matanyone2.model.channel_attn import CAResBlock
+
+def interpolate_groups(g: torch.Tensor, ratio: float, mode: str,
+                       align_corners: bool) -> torch.Tensor:
+    batch_size, num_objects = g.shape[:2]
+    g = F.interpolate(g.flatten(start_dim=0, end_dim=1),
+                      scale_factor=ratio,
+                      mode=mode,
+                      align_corners=align_corners)
+    g = g.view(batch_size, num_objects, *g.shape[1:])
+    return g
+
+
+def upsample_groups(g: torch.Tensor,
+                    ratio: float = 2,
+                    mode: str = 'bilinear',
+                    align_corners: bool = False) -> torch.Tensor:
+    return interpolate_groups(g, ratio, mode, align_corners)
+
+
+def downsample_groups(g: torch.Tensor,
+                      ratio: float = 1 / 2,
+                      mode: str = 'area',
+                      align_corners: bool = None) -> torch.Tensor:
+    return interpolate_groups(g, ratio, mode, align_corners)
+
+
+class GConv2d(nn.Conv2d):
+    def forward(self, g: torch.Tensor) -> torch.Tensor:
+        batch_size, num_objects = g.shape[:2]
+        g = super().forward(g.flatten(start_dim=0, end_dim=1))
+        return g.view(batch_size, num_objects, *g.shape[1:])
+
+
+class GroupResBlock(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int):
+        super().__init__()
+
+        if in_dim == out_dim:
+            self.downsample = nn.Identity()
+        else:
+            self.downsample = GConv2d(in_dim, out_dim, kernel_size=1)
+
+        self.conv1 = GConv2d(in_dim, out_dim, kernel_size=3, padding=1)
+        self.conv2 = GConv2d(out_dim, out_dim, kernel_size=3, padding=1)
+
+    def forward(self, g: torch.Tensor) -> torch.Tensor:
+        out_g = self.conv1(F.relu(g))
+        out_g = self.conv2(F.relu(out_g))
+
+        g = self.downsample(g)
+
+        return out_g + g
+
+
+class MainToGroupDistributor(nn.Module):
+    def __init__(self,
+                 x_transform: Optional[nn.Module] = None,
+                 g_transform: Optional[nn.Module] = None,
+                 method: str = 'cat',
+                 reverse_order: bool = False):
+        super().__init__()
+
+        self.x_transform = x_transform
+        self.g_transform = g_transform
+        self.method = method
+        self.reverse_order = reverse_order
+
+    def forward(self, x: torch.Tensor, g: torch.Tensor, skip_expand: bool = False) -> torch.Tensor:
+        num_objects = g.shape[1]
+
+        if self.x_transform is not None:
+            x = self.x_transform(x)
+
+        if self.g_transform is not None:
+            g = self.g_transform(g)
+
+        if not skip_expand:
+            x = x.unsqueeze(1).expand(-1, num_objects, -1, -1, -1)
+        if self.method == 'cat':
+            if self.reverse_order:
+                g = torch.cat([g, x], 2)
+            else:
+                g = torch.cat([x, g], 2)
+        elif self.method == 'add':
+            g = x + g
+        elif self.method == 'mulcat':
+            g = torch.cat([x * g, g], dim=2)
+        elif self.method == 'muladd':
+            g = x * g + g
+        else:
+            raise NotImplementedError
+
+        return g
+
+
+class GroupFeatureFusionBlock(nn.Module):
+    def __init__(self, x_in_dim: int, g_in_dim: int, out_dim: int):
+        super().__init__()
+
+        x_transform = nn.Conv2d(x_in_dim, out_dim, kernel_size=1)
+        g_transform = GConv2d(g_in_dim, out_dim, kernel_size=1)
+
+        self.distributor = MainToGroupDistributor(x_transform=x_transform,
+                                                  g_transform=g_transform,
+                                                  method='add')
+        self.block1 = CAResBlock(out_dim, out_dim)
+        self.block2 = CAResBlock(out_dim, out_dim)
+
+    def forward(self, x: torch.Tensor, g: torch.Tensor) -> torch.Tensor:
+        batch_size, num_objects = g.shape[:2]
+
+        g = self.distributor(x, g)
+
+        g = g.flatten(start_dim=0, end_dim=1)
+
+        g = self.block1(g)
+        g = self.block2(g)
+
+        g = g.view(batch_size, num_objects, *g.shape[1:])
+
+        return g

vendor/MatAnyone2/matanyone2/model/matanyone2.py

@@ -0,0 +1,338 @@
+from typing import List, Dict, Iterable, Tuple
+import logging
+from omegaconf import DictConfig
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from omegaconf import OmegaConf
+from huggingface_hub import PyTorchModelHubMixin
+
+from matanyone2.model.big_modules import PixelEncoder, UncertPred, KeyProjection, MaskEncoder, PixelFeatureFuser, MaskDecoder
+from matanyone2.model.aux_modules import AuxComputer
+from matanyone2.model.utils.memory_utils import get_affinity, readout
+from matanyone2.model.transformer.object_transformer import QueryTransformer
+from matanyone2.model.transformer.object_summarizer import ObjectSummarizer
+from matanyone2.utils.tensor_utils import aggregate
+from matanyone2.utils.device import get_default_device, safe_autocast
+
+device = get_default_device()
+
+log = logging.getLogger()
+class MatAnyone2(nn.Module,
+                PyTorchModelHubMixin,
+                library_name="matanyone2",
+                repo_url="https://github.com/pq-yang/MatAnyone2",
+                coders={
+                    DictConfig: (
+                        lambda x: OmegaConf.to_container(x),
+                        lambda data: OmegaConf.create(data),
+                    )
+                },
+        ):
+
+    def __init__(self, cfg: DictConfig, *, single_object=False):
+        super().__init__()
+        self.cfg = cfg
+        model_cfg = cfg.model
+        self.ms_dims = model_cfg.pixel_encoder.ms_dims
+        self.key_dim = model_cfg.key_dim
+        self.value_dim = model_cfg.value_dim
+        self.sensory_dim = model_cfg.sensory_dim
+        self.pixel_dim = model_cfg.pixel_dim
+        self.embed_dim = model_cfg.embed_dim
+        self.single_object = single_object
+
+        log.info(f'Single object: {self.single_object}')
+
+        self.pixel_encoder = PixelEncoder(model_cfg)
+        self.pix_feat_proj = nn.Conv2d(self.ms_dims[0], self.pixel_dim, kernel_size=1)
+        self.key_proj = KeyProjection(model_cfg)
+        self.mask_encoder = MaskEncoder(model_cfg, single_object=single_object)
+        self.mask_decoder = MaskDecoder(model_cfg)
+        self.pixel_fuser = PixelFeatureFuser(model_cfg, single_object=single_object)
+        self.object_transformer = QueryTransformer(model_cfg)
+        self.object_summarizer = ObjectSummarizer(model_cfg)
+        self.aux_computer = AuxComputer(cfg)
+        self.temp_sparity = UncertPred(model_cfg)
+
+        self.register_buffer("pixel_mean", torch.Tensor(model_cfg.pixel_mean).view(-1, 1, 1), False)
+        self.register_buffer("pixel_std", torch.Tensor(model_cfg.pixel_std).view(-1, 1, 1), False)
+
+    def _get_others(self, masks: torch.Tensor) -> torch.Tensor:
+        # for each object, return the sum of masks of all other objects
+        if self.single_object:
+            return None
+
+        num_objects = masks.shape[1]
+        if num_objects >= 1:
+            others = (masks.sum(dim=1, keepdim=True) - masks).clamp(0, 1)
+        else:
+            others = torch.zeros_like(masks)
+        return others
+
+    def pred_uncertainty(self, last_pix_feat: torch.Tensor, cur_pix_feat: torch.Tensor, last_mask: torch.Tensor, mem_val_diff:torch.Tensor):
+        logits = self.temp_sparity(last_frame_feat=last_pix_feat,
+                                cur_frame_feat=cur_pix_feat,
+                                last_mask=last_mask,
+                                mem_val_diff=mem_val_diff)
+
+        prob = torch.sigmoid(logits)
+        mask = (prob > 0) + 0
+
+        uncert_output = {"logits": logits,
+                     "prob": prob,
+                     "mask": mask}
+
+        return uncert_output
+
+    def encode_image(self, image: torch.Tensor, seq_length=None, last_feats=None) -> (Iterable[torch.Tensor], torch.Tensor): # type: ignore
+        self.pixel_mean = self.pixel_mean.to(device)
+        self.pixel_std = self.pixel_std.to(device)
+        image = (image - self.pixel_mean) / self.pixel_std
+        ms_image_feat = self.pixel_encoder(image, seq_length) # f16, f8, f4, f2, f1
+        return ms_image_feat, self.pix_feat_proj(ms_image_feat[0])
+
+    def encode_mask(
+            self,
+            image: torch.Tensor,
+            ms_features: List[torch.Tensor],
+            sensory: torch.Tensor,
+            masks: torch.Tensor,
+            *,
+            deep_update: bool = True,
+            chunk_size: int = -1,
+            need_weights: bool = False) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        image = (image - self.pixel_mean) / self.pixel_std
+        others = self._get_others(masks)
+        mask_value, new_sensory = self.mask_encoder(image,
+                                                    ms_features,
+                                                    sensory,
+                                                    masks,
+                                                    others,
+                                                    deep_update=deep_update,
+                                                    chunk_size=chunk_size)
+        object_summaries, object_logits = self.object_summarizer(masks, mask_value, need_weights)
+        return mask_value, new_sensory, object_summaries, object_logits
+
+    def transform_key(self,
+                      final_pix_feat: torch.Tensor,
+                      *,
+                      need_sk: bool = True,
+                      need_ek: bool = True) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        key, shrinkage, selection = self.key_proj(final_pix_feat, need_s=need_sk, need_e=need_ek)
+        return key, shrinkage, selection
+
+    # Used in training only.
+    # This step is replaced by MemoryManager in test time
+    def read_memory(self, query_key: torch.Tensor, query_selection: torch.Tensor,
+                    memory_key: torch.Tensor, memory_shrinkage: torch.Tensor,
+                    msk_value: torch.Tensor, obj_memory: torch.Tensor, pix_feat: torch.Tensor,
+                    sensory: torch.Tensor, last_mask: torch.Tensor,
+                    selector: torch.Tensor, uncert_output=None, seg_pass=False,
+                    last_pix_feat=None, last_pred_mask=None) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
+        """
+        query_key       : B * CK * H * W
+        query_selection : B * CK * H * W
+        memory_key      : B * CK * T * H * W
+        memory_shrinkage: B * 1  * T * H * W
+        msk_value       : B * num_objects * CV * T * H * W
+        obj_memory      : B * num_objects * T * num_summaries * C
+        pixel_feature   : B * C * H * W
+        """
+        batch_size, num_objects = msk_value.shape[:2]
+
+        uncert_mask = uncert_output["mask"] if uncert_output is not None else None
+
+        # read using visual attention
+        with safe_autocast(enabled=False):
+            affinity = get_affinity(memory_key.float(), memory_shrinkage.float(), query_key.float(),
+                                    query_selection.float(), uncert_mask=uncert_mask)
+
+            msk_value = msk_value.flatten(start_dim=1, end_dim=2).float()
+
+            # B * (num_objects*CV) * H * W
+            pixel_readout = readout(affinity, msk_value, uncert_mask)
+            pixel_readout = pixel_readout.view(batch_size, num_objects, self.value_dim,
+                                            *pixel_readout.shape[-2:])
+            
+            uncert_output = self.pred_uncertainty(last_pix_feat, pix_feat, last_pred_mask, pixel_readout[:,0]-msk_value[:,:,-1])
+            uncert_prob = uncert_output["prob"].unsqueeze(1) # b n 1 h w
+            pixel_readout = pixel_readout*uncert_prob + msk_value[:,:,-1].unsqueeze(1)*(1-uncert_prob)
+                
+        pixel_readout = self.pixel_fusion(pix_feat, pixel_readout, sensory, last_mask)
+
+
+        # read from query transformer
+        mem_readout, aux_features = self.readout_query(pixel_readout, obj_memory, selector=selector, seg_pass=seg_pass)
+
+        aux_output = {
+            'sensory': sensory,
+            'q_logits': aux_features['logits'] if aux_features else None,
+            'attn_mask': aux_features['attn_mask'] if aux_features else None,
+        }
+
+        return mem_readout, aux_output, uncert_output
+    
+    def read_first_frame_memory(self, pixel_readout,
+                    obj_memory: torch.Tensor, pix_feat: torch.Tensor,
+                    sensory: torch.Tensor, last_mask: torch.Tensor,
+                    selector: torch.Tensor, seg_pass=False) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
+        """
+        query_key       : B * CK * H * W
+        query_selection : B * CK * H * W
+        memory_key      : B * CK * T * H * W
+        memory_shrinkage: B * 1  * T * H * W
+        msk_value       : B * num_objects * CV * T * H * W
+        obj_memory      : B * num_objects * T * num_summaries * C
+        pixel_feature   : B * C * H * W
+        """
+                
+        pixel_readout = self.pixel_fusion(pix_feat, pixel_readout, sensory, last_mask)
+
+        # read from query transformer
+        mem_readout, aux_features = self.readout_query(pixel_readout, obj_memory, selector=selector, seg_pass=seg_pass)
+
+        aux_output = {
+            'sensory': sensory,
+            'q_logits': aux_features['logits'] if aux_features else None,
+            'attn_mask': aux_features['attn_mask'] if aux_features else None,
+        }
+
+        return mem_readout, aux_output
+
+    def pixel_fusion(self,
+                     pix_feat: torch.Tensor,
+                     pixel: torch.Tensor,
+                     sensory: torch.Tensor,
+                     last_mask: torch.Tensor,
+                     *,
+                     chunk_size: int = -1) -> torch.Tensor:
+        last_mask = F.interpolate(last_mask, size=sensory.shape[-2:], mode='area')
+        last_others = self._get_others(last_mask)
+        fused = self.pixel_fuser(pix_feat,
+                                 pixel,
+                                 sensory,
+                                 last_mask,
+                                 last_others,
+                                 chunk_size=chunk_size)
+        return fused
+
+    def readout_query(self,
+                      pixel_readout,
+                      obj_memory,
+                      *,
+                      selector=None,
+                      need_weights=False,
+                      seg_pass=False) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
+        return self.object_transformer(pixel_readout,
+                                       obj_memory,
+                                       selector=selector,
+                                       need_weights=need_weights,
+                                       seg_pass=seg_pass)
+
+    def segment(self,
+                ms_image_feat: List[torch.Tensor],
+                memory_readout: torch.Tensor,
+                sensory: torch.Tensor,
+                *,
+                selector: bool = None,
+                chunk_size: int = -1,
+                update_sensory: bool = True,
+                seg_pass: bool = False,
+                clamp_mat: bool = True,
+                last_mask=None,
+                sigmoid_residual=False,
+                seg_mat=False) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        multi_scale_features is from the key encoder for skip-connection
+        memory_readout is from working/long-term memory
+        sensory is the sensory memory
+        last_mask is the mask from the last frame, supplementing sensory memory
+        selector is 1 if an object exists, and 0 otherwise. We use it to filter padded objects
+            during training.
+        """
+        #### use mat head for seg data
+        if seg_mat:
+            assert seg_pass
+            seg_pass = False
+        ####
+        sensory, logits = self.mask_decoder(ms_image_feat,
+                                        memory_readout,
+                                        sensory,
+                                        chunk_size=chunk_size,
+                                        update_sensory=update_sensory,
+                                        seg_pass = seg_pass,
+                                        last_mask=last_mask,
+                                        sigmoid_residual=sigmoid_residual)
+        if seg_pass:
+            prob = torch.sigmoid(logits)
+            if selector is not None:
+                prob = prob * selector
+
+            # Softmax over all objects[]
+            logits = aggregate(prob, dim=1)
+            prob = F.softmax(logits, dim=1)
+        else:
+            if clamp_mat:
+                logits = logits.clamp(0.0, 1.0)
+            logits = torch.cat([torch.prod(1 - logits, dim=1, keepdim=True), logits], 1)
+            prob = logits
+        
+        return sensory, logits, prob
+
+    def compute_aux(self, pix_feat: torch.Tensor, aux_inputs: Dict[str, torch.Tensor],
+                    selector: torch.Tensor, seg_pass=False) -> Dict[str, torch.Tensor]:
+        return self.aux_computer(pix_feat, aux_inputs, selector, seg_pass=seg_pass)
+
+    def forward(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def load_weights(self, src_dict, init_as_zero_if_needed=False) -> None:
+        if not self.single_object:
+            # Map single-object weight to multi-object weight (4->5 out channels in conv1)
+            for k in list(src_dict.keys()):
+                if k == 'mask_encoder.conv1.weight':
+                    if src_dict[k].shape[1] == 4:
+                        log.info(f'Converting {k} from single object to multiple objects.')
+                        pads = torch.zeros((64, 1, 7, 7), device=src_dict[k].device)
+                        if not init_as_zero_if_needed:
+                            nn.init.orthogonal_(pads)
+                            log.info(f'Randomly initialized padding for {k}.')
+                        else:
+                            log.info(f'Zero-initialized padding for {k}.')
+                        src_dict[k] = torch.cat([src_dict[k], pads], 1)
+                elif k == 'pixel_fuser.sensory_compress.weight':
+                    if src_dict[k].shape[1] == self.sensory_dim + 1:
+                        log.info(f'Converting {k} from single object to multiple objects.')
+                        pads = torch.zeros((self.value_dim, 1, 1, 1), device=src_dict[k].device)
+                        if not init_as_zero_if_needed:
+                            nn.init.orthogonal_(pads)
+                            log.info(f'Randomly initialized padding for {k}.')
+                        else:
+                            log.info(f'Zero-initialized padding for {k}.')
+                        src_dict[k] = torch.cat([src_dict[k], pads], 1)
+        elif self.single_object:
+            """
+            If the model is multiple-object and we are training in single-object, 
+            we strip the last channel of conv1.
+            This is not supposed to happen in standard training except when users are trying to
+            finetune a trained model with single object datasets.
+            """
+            if src_dict['mask_encoder.conv1.weight'].shape[1] == 5:
+                log.warning('Converting mask_encoder.conv1.weight from multiple objects to single object.'
+                            'This is not supposed to happen in standard training.')
+                src_dict['mask_encoder.conv1.weight'] = src_dict['mask_encoder.conv1.weight'][:, :-1]
+                src_dict['pixel_fuser.sensory_compress.weight'] = src_dict['pixel_fuser.sensory_compress.weight'][:, :-1]
+
+        for k in src_dict:
+            if k not in self.state_dict():
+                log.info(f'Key {k} found in src_dict but not in self.state_dict()!!!')
+        for k in self.state_dict():
+            if k not in src_dict:
+                log.info(f'Key {k} found in self.state_dict() but not in src_dict!!!')
+                
+        self.load_state_dict(src_dict, strict=False)
+
+    @property
+    def device(self) -> torch.device:
+        return self.pixel_mean.device

vendor/MatAnyone2/matanyone2/model/modules.py

@@ -0,0 +1,150 @@
+from typing import List, Iterable
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from matanyone2.model.group_modules import MainToGroupDistributor, GroupResBlock, upsample_groups, GConv2d, downsample_groups
+from matanyone2.utils.device import safe_autocast
+
+
+class UpsampleBlock(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, scale_factor: int = 2):
+        super().__init__()
+        self.out_conv = ResBlock(in_dim, out_dim)
+        self.scale_factor = scale_factor
+
+    def forward(self, in_g: torch.Tensor, skip_f: torch.Tensor) -> torch.Tensor:
+        g = F.interpolate(in_g,
+                      scale_factor=self.scale_factor,
+                      mode='bilinear')
+        g = self.out_conv(g)
+        g = g + skip_f
+        return g
+
+class MaskUpsampleBlock(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, scale_factor: int = 2):
+        super().__init__()
+        self.distributor = MainToGroupDistributor(method='add')
+        self.out_conv = GroupResBlock(in_dim, out_dim)
+        self.scale_factor = scale_factor
+
+    def forward(self, in_g: torch.Tensor, skip_f: torch.Tensor) -> torch.Tensor:
+        g = upsample_groups(in_g, ratio=self.scale_factor)
+        g = self.distributor(skip_f, g)
+        g = self.out_conv(g)
+        return g
+    
+
+class DecoderFeatureProcessor(nn.Module):
+    def __init__(self, decoder_dims: List[int], out_dims: List[int]):
+        super().__init__()
+        self.transforms = nn.ModuleList([
+            nn.Conv2d(d_dim, p_dim, kernel_size=1) for d_dim, p_dim in zip(decoder_dims, out_dims)
+        ])
+
+    def forward(self, multi_scale_features: Iterable[torch.Tensor]) -> List[torch.Tensor]:
+        outputs = [func(x) for x, func in zip(multi_scale_features, self.transforms)]
+        return outputs
+
+
+# @torch.jit.script
+def _recurrent_update(h: torch.Tensor, values: torch.Tensor) -> torch.Tensor:
+    # h: batch_size * num_objects * hidden_dim * h * w
+    # values: batch_size * num_objects * (hidden_dim*3) * h * w
+    dim = values.shape[2] // 3
+    forget_gate = torch.sigmoid(values[:, :, :dim])
+    update_gate = torch.sigmoid(values[:, :, dim:dim * 2])
+    new_value = torch.tanh(values[:, :, dim * 2:])
+    new_h = forget_gate * h * (1 - update_gate) + update_gate * new_value
+    return new_h
+
+
+class SensoryUpdater_fullscale(nn.Module):
+    # Used in the decoder, multi-scale feature + GRU
+    def __init__(self, g_dims: List[int], mid_dim: int, sensory_dim: int):
+        super().__init__()
+        self.g16_conv = GConv2d(g_dims[0], mid_dim, kernel_size=1)
+        self.g8_conv = GConv2d(g_dims[1], mid_dim, kernel_size=1)
+        self.g4_conv = GConv2d(g_dims[2], mid_dim, kernel_size=1)
+        self.g2_conv = GConv2d(g_dims[3], mid_dim, kernel_size=1)
+        self.g1_conv = GConv2d(g_dims[4], mid_dim, kernel_size=1)
+
+        self.transform = GConv2d(mid_dim + sensory_dim, sensory_dim * 3, kernel_size=3, padding=1)
+
+        nn.init.xavier_normal_(self.transform.weight)
+
+    def forward(self, g: torch.Tensor, h: torch.Tensor) -> torch.Tensor:
+        g = self.g16_conv(g[0]) + self.g8_conv(downsample_groups(g[1], ratio=1/2)) + \
+            self.g4_conv(downsample_groups(g[2], ratio=1/4)) + \
+            self.g2_conv(downsample_groups(g[3], ratio=1/8)) + \
+            self.g1_conv(downsample_groups(g[4], ratio=1/16))
+
+        with safe_autocast(enabled=False):
+            g = g.float()
+            h = h.float()
+            values = self.transform(torch.cat([g, h], dim=2))
+            new_h = _recurrent_update(h, values)
+
+        return new_h
+
+class SensoryUpdater(nn.Module):
+    # Used in the decoder, multi-scale feature + GRU
+    def __init__(self, g_dims: List[int], mid_dim: int, sensory_dim: int):
+        super().__init__()
+        self.g16_conv = GConv2d(g_dims[0], mid_dim, kernel_size=1)
+        self.g8_conv = GConv2d(g_dims[1], mid_dim, kernel_size=1)
+        self.g4_conv = GConv2d(g_dims[2], mid_dim, kernel_size=1)
+
+        self.transform = GConv2d(mid_dim + sensory_dim, sensory_dim * 3, kernel_size=3, padding=1)
+
+        nn.init.xavier_normal_(self.transform.weight)
+
+    def forward(self, g: torch.Tensor, h: torch.Tensor) -> torch.Tensor:
+        g = self.g16_conv(g[0]) + self.g8_conv(downsample_groups(g[1], ratio=1/2)) + \
+            self.g4_conv(downsample_groups(g[2], ratio=1/4))
+
+        with safe_autocast(enabled=False):
+            g = g.float()
+            h = h.float()
+            values = self.transform(torch.cat([g, h], dim=2))
+            new_h = _recurrent_update(h, values)
+
+        return new_h
+
+
+class SensoryDeepUpdater(nn.Module):
+    def __init__(self, f_dim: int, sensory_dim: int):
+        super().__init__()
+        self.transform = GConv2d(f_dim + sensory_dim, sensory_dim * 3, kernel_size=3, padding=1)
+
+        nn.init.xavier_normal_(self.transform.weight)
+
+    def forward(self, g: torch.Tensor, h: torch.Tensor) -> torch.Tensor:
+        with safe_autocast(enabled=False):
+            g = g.float()
+            h = h.float()
+            values = self.transform(torch.cat([g, h], dim=2))
+            new_h = _recurrent_update(h, values)
+
+        return new_h
+
+  
+class ResBlock(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int):
+        super().__init__()
+
+        if in_dim == out_dim:
+            self.downsample = nn.Identity()
+        else:
+            self.downsample = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+
+        self.conv1 = nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1)
+        self.conv2 = nn.Conv2d(out_dim, out_dim, kernel_size=3, padding=1)
+
+    def forward(self, g: torch.Tensor) -> torch.Tensor:
+        out_g = self.conv1(F.relu(g))
+        out_g = self.conv2(F.relu(out_g))
+
+        g = self.downsample(g)
+
+        return out_g + g

vendor/MatAnyone2/matanyone2/model/transformer/object_summarizer.py

@@ -0,0 +1,90 @@
+from typing import Optional
+from omegaconf import DictConfig
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from matanyone2.model.transformer.positional_encoding import PositionalEncoding
+from matanyone2.utils.device import safe_autocast
+
+
+# @torch.jit.script
+def _weighted_pooling(masks: torch.Tensor, value: torch.Tensor,
+                      logits: torch.Tensor) -> (torch.Tensor, torch.Tensor):
+    # value: B*num_objects*H*W*value_dim
+    # logits: B*num_objects*H*W*num_summaries
+    # masks: B*num_objects*H*W*num_summaries: 1 if allowed
+    weights = logits.sigmoid() * masks
+    # B*num_objects*num_summaries*value_dim
+    sums = torch.einsum('bkhwq,bkhwc->bkqc', weights, value)
+    # B*num_objects*H*W*num_summaries -> B*num_objects*num_summaries*1
+    area = weights.flatten(start_dim=2, end_dim=3).sum(2).unsqueeze(-1)
+
+    # B*num_objects*num_summaries*value_dim
+    return sums, area
+
+
+class ObjectSummarizer(nn.Module):
+    def __init__(self, model_cfg: DictConfig):
+        super().__init__()
+
+        this_cfg = model_cfg.object_summarizer
+        self.value_dim = model_cfg.value_dim
+        self.embed_dim = this_cfg.embed_dim
+        self.num_summaries = this_cfg.num_summaries
+        self.add_pe = this_cfg.add_pe
+        self.pixel_pe_scale = model_cfg.pixel_pe_scale
+        self.pixel_pe_temperature = model_cfg.pixel_pe_temperature
+
+        if self.add_pe:
+            self.pos_enc = PositionalEncoding(self.embed_dim,
+                                              scale=self.pixel_pe_scale,
+                                              temperature=self.pixel_pe_temperature)
+
+        self.input_proj = nn.Linear(self.value_dim, self.embed_dim)
+        self.feature_pred = nn.Sequential(
+            nn.Linear(self.embed_dim, self.embed_dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(self.embed_dim, self.embed_dim),
+        )
+        self.weights_pred = nn.Sequential(
+            nn.Linear(self.embed_dim, self.embed_dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(self.embed_dim, self.num_summaries),
+        )
+
+    def forward(self,
+                masks: torch.Tensor,
+                value: torch.Tensor,
+                need_weights: bool = False) -> (torch.Tensor, Optional[torch.Tensor]):
+        # masks: B*num_objects*(H0)*(W0)
+        # value: B*num_objects*value_dim*H*W
+        # -> B*num_objects*H*W*value_dim
+        h, w = value.shape[-2:]
+        masks = F.interpolate(masks, size=(h, w), mode='area')
+        masks = masks.unsqueeze(-1)
+        inv_masks = 1 - masks
+        repeated_masks = torch.cat([
+            masks.expand(-1, -1, -1, -1, self.num_summaries // 2),
+            inv_masks.expand(-1, -1, -1, -1, self.num_summaries // 2),
+        ],
+                                   dim=-1)
+
+        value = value.permute(0, 1, 3, 4, 2)
+        value = self.input_proj(value)
+        if self.add_pe:
+            pe = self.pos_enc(value)
+            value = value + pe
+
+        with safe_autocast(enabled=False):  # autocast disabled intentionally
+            value = value.float()
+            feature = self.feature_pred(value)
+            logits = self.weights_pred(value)
+            sums, area = _weighted_pooling(repeated_masks, feature, logits)
+
+        summaries = torch.cat([sums, area], dim=-1)
+
+        if need_weights:
+            return summaries, logits
+        else:
+            return summaries, None

vendor/MatAnyone2/matanyone2/model/transformer/object_transformer.py

@@ -0,0 +1,206 @@
+from typing import Dict, Optional
+from omegaconf import DictConfig
+
+import torch
+import torch.nn as nn
+from matanyone2.model.group_modules import GConv2d
+from matanyone2.utils.tensor_utils import aggregate
+from matanyone2.model.transformer.positional_encoding import PositionalEncoding
+from matanyone2.model.transformer.transformer_layers import CrossAttention, SelfAttention, FFN, PixelFFN
+
+
+class QueryTransformerBlock(nn.Module):
+    def __init__(self, model_cfg: DictConfig):
+        super().__init__()
+
+        this_cfg = model_cfg.object_transformer
+        self.embed_dim = this_cfg.embed_dim
+        self.num_heads = this_cfg.num_heads
+        self.num_queries = this_cfg.num_queries
+        self.ff_dim = this_cfg.ff_dim
+
+        self.read_from_pixel = CrossAttention(self.embed_dim,
+                                              self.num_heads,
+                                              add_pe_to_qkv=this_cfg.read_from_pixel.add_pe_to_qkv)
+        self.self_attn = SelfAttention(self.embed_dim,
+                                       self.num_heads,
+                                       add_pe_to_qkv=this_cfg.query_self_attention.add_pe_to_qkv)
+        self.ffn = FFN(self.embed_dim, self.ff_dim)
+        self.read_from_query = CrossAttention(self.embed_dim,
+                                              self.num_heads,
+                                              add_pe_to_qkv=this_cfg.read_from_query.add_pe_to_qkv,
+                                              norm=this_cfg.read_from_query.output_norm)
+        self.pixel_ffn = PixelFFN(self.embed_dim)
+
+    def forward(
+            self,
+            x: torch.Tensor,
+            pixel: torch.Tensor,
+            query_pe: torch.Tensor,
+            pixel_pe: torch.Tensor,
+            attn_mask: torch.Tensor,
+            need_weights: bool = False) -> (torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor):
+        # x: (bs*num_objects)*num_queries*embed_dim
+        # pixel: bs*num_objects*C*H*W
+        # query_pe: (bs*num_objects)*num_queries*embed_dim
+        # pixel_pe: (bs*num_objects)*(H*W)*C
+        # attn_mask: (bs*num_objects*num_heads)*num_queries*(H*W)
+
+        # bs*num_objects*C*H*W -> (bs*num_objects)*(H*W)*C
+        pixel_flat = pixel.flatten(3, 4).flatten(0, 1).transpose(1, 2).contiguous()
+        x, q_weights = self.read_from_pixel(x,
+                                            pixel_flat,
+                                            query_pe,
+                                            pixel_pe,
+                                            attn_mask=attn_mask,
+                                            need_weights=need_weights)
+        x = self.self_attn(x, query_pe)
+        x = self.ffn(x)
+
+        pixel_flat, p_weights = self.read_from_query(pixel_flat,
+                                                     x,
+                                                     pixel_pe,
+                                                     query_pe,
+                                                     need_weights=need_weights)
+        pixel = self.pixel_ffn(pixel, pixel_flat)
+
+        if need_weights:
+            bs, num_objects, _, h, w = pixel.shape
+            q_weights = q_weights.view(bs, num_objects, self.num_heads, self.num_queries, h, w)
+            p_weights = p_weights.transpose(2, 3).view(bs, num_objects, self.num_heads,
+                                                       self.num_queries, h, w)
+
+        return x, pixel, q_weights, p_weights
+
+
+class QueryTransformer(nn.Module):
+    def __init__(self, model_cfg: DictConfig):
+        super().__init__()
+
+        this_cfg = model_cfg.object_transformer
+        self.value_dim = model_cfg.value_dim
+        self.embed_dim = this_cfg.embed_dim
+        self.num_heads = this_cfg.num_heads
+        self.num_queries = this_cfg.num_queries
+
+        # query initialization and embedding
+        self.query_init = nn.Embedding(self.num_queries, self.embed_dim)
+        self.query_emb = nn.Embedding(self.num_queries, self.embed_dim)
+
+        # projection from object summaries to query initialization and embedding
+        self.summary_to_query_init = nn.Linear(self.embed_dim, self.embed_dim)
+        self.summary_to_query_emb = nn.Linear(self.embed_dim, self.embed_dim)
+
+        self.pixel_pe_scale = model_cfg.pixel_pe_scale
+        self.pixel_pe_temperature = model_cfg.pixel_pe_temperature
+        self.pixel_init_proj = GConv2d(self.embed_dim, self.embed_dim, kernel_size=1)
+        self.pixel_emb_proj = GConv2d(self.embed_dim, self.embed_dim, kernel_size=1)
+        self.spatial_pe = PositionalEncoding(self.embed_dim,
+                                             scale=self.pixel_pe_scale,
+                                             temperature=self.pixel_pe_temperature,
+                                             channel_last=False,
+                                             transpose_output=True)
+
+        # transformer blocks
+        self.num_blocks = this_cfg.num_blocks
+        self.blocks = nn.ModuleList(
+            QueryTransformerBlock(model_cfg) for _ in range(self.num_blocks))
+        self.mask_pred = nn.ModuleList(
+            nn.Sequential(nn.ReLU(), GConv2d(self.embed_dim, 1, kernel_size=1))
+            for _ in range(self.num_blocks + 1))
+
+        self.act = nn.ReLU(inplace=True)
+
+    def forward(self,
+                pixel: torch.Tensor,
+                obj_summaries: torch.Tensor,
+                selector: Optional[torch.Tensor] = None,
+                need_weights: bool = False,
+                seg_pass=False) -> (torch.Tensor, Dict[str, torch.Tensor]):
+        # pixel: B*num_objects*embed_dim*H*W
+        # obj_summaries: B*num_objects*T*num_queries*embed_dim
+        T = obj_summaries.shape[2]
+        bs, num_objects, _, H, W = pixel.shape
+
+        # normalize object values
+        # the last channel is the cumulative area of the object
+        obj_summaries = obj_summaries.view(bs * num_objects, T, self.num_queries,
+                                           self.embed_dim + 1)
+        # sum over time
+        # during inference, T=1 as we already did streaming average in memory_manager
+        obj_sums = obj_summaries[:, :, :, :-1].sum(dim=1)
+        obj_area = obj_summaries[:, :, :, -1:].sum(dim=1)
+        obj_values = obj_sums / (obj_area + 1e-4)
+        obj_init = self.summary_to_query_init(obj_values)
+        obj_emb = self.summary_to_query_emb(obj_values)
+
+        # positional embeddings for object queries
+        query = self.query_init.weight.unsqueeze(0).expand(bs * num_objects, -1, -1) + obj_init
+        query_emb = self.query_emb.weight.unsqueeze(0).expand(bs * num_objects, -1, -1) + obj_emb
+
+        # positional embeddings for pixel features
+        pixel_init = self.pixel_init_proj(pixel)
+        pixel_emb = self.pixel_emb_proj(pixel)
+        pixel_pe = self.spatial_pe(pixel.flatten(0, 1))
+        pixel_emb = pixel_emb.flatten(3, 4).flatten(0, 1).transpose(1, 2).contiguous()
+        pixel_pe = pixel_pe.flatten(1, 2) + pixel_emb
+
+        pixel = pixel_init
+
+        # run the transformer
+        aux_features = {'logits': []}
+
+        # first aux output
+        aux_logits = self.mask_pred[0](pixel).squeeze(2)
+        attn_mask = self._get_aux_mask(aux_logits, selector, seg_pass=seg_pass)
+        aux_features['logits'].append(aux_logits)
+        for i in range(self.num_blocks):
+            query, pixel, q_weights, p_weights = self.blocks[i](query,
+                                                                pixel,
+                                                                query_emb,
+                                                                pixel_pe,
+                                                                attn_mask,
+                                                                need_weights=need_weights)
+
+            if self.training or i <= self.num_blocks - 1 or need_weights:
+                aux_logits = self.mask_pred[i + 1](pixel).squeeze(2)
+                attn_mask = self._get_aux_mask(aux_logits, selector, seg_pass=seg_pass)
+                aux_features['logits'].append(aux_logits)
+
+        aux_features['q_weights'] = q_weights  # last layer only
+        aux_features['p_weights'] = p_weights  # last layer only
+
+        if self.training:
+            # no need to save all heads
+            aux_features['attn_mask'] = attn_mask.view(bs, num_objects, self.num_heads,
+                                                       self.num_queries, H, W)[:, :, 0]
+
+        return pixel, aux_features
+
+    def _get_aux_mask(self, logits: torch.Tensor, selector: torch.Tensor, seg_pass=False) -> torch.Tensor:
+        # logits: batch_size*num_objects*H*W
+        # selector: batch_size*num_objects*1*1
+        # returns a mask of shape (batch_size*num_objects*num_heads)*num_queries*(H*W)
+        # where True means the attention is blocked
+
+        if selector is None:
+            prob = logits.sigmoid()
+        else:
+            prob = logits.sigmoid() * selector
+        logits = aggregate(prob, dim=1)
+
+        is_foreground = (logits[:, 1:] >= logits.max(dim=1, keepdim=True)[0])
+        foreground_mask = is_foreground.bool().flatten(start_dim=2)
+        inv_foreground_mask = ~foreground_mask
+        inv_background_mask = foreground_mask
+
+        aux_foreground_mask = inv_foreground_mask.unsqueeze(2).unsqueeze(2).repeat(
+            1, 1, self.num_heads, self.num_queries // 2, 1).flatten(start_dim=0, end_dim=2)
+        aux_background_mask = inv_background_mask.unsqueeze(2).unsqueeze(2).repeat(
+            1, 1, self.num_heads, self.num_queries // 2, 1).flatten(start_dim=0, end_dim=2)
+
+        aux_mask = torch.cat([aux_foreground_mask, aux_background_mask], dim=1)
+
+        aux_mask[torch.where(aux_mask.sum(-1) == aux_mask.shape[-1])] = False
+
+        return aux_mask

vendor/MatAnyone2/matanyone2/model/transformer/positional_encoding.py

@@ -0,0 +1,110 @@
+# Reference:
+# https://github.com/facebookresearch/Mask2Former/blob/main/mask2former/modeling/transformer_decoder/position_encoding.py
+# https://github.com/tatp22/multidim-positional-encoding/blob/master/positional_encodings/torch_encodings.py
+
+import math
+
+import numpy as np
+import torch
+from torch import nn
+from matanyone2.utils.device import get_default_device
+
+
+def get_emb(sin_inp: torch.Tensor) -> torch.Tensor:
+    """
+    Gets a base embedding for one dimension with sin and cos intertwined
+    """
+    emb = torch.stack((sin_inp.sin(), sin_inp.cos()), dim=-1)
+    return torch.flatten(emb, -2, -1)
+
+
+class PositionalEncoding(nn.Module):
+    def __init__(self,
+                 dim: int,
+                 scale: float = math.pi * 2,
+                 temperature: float = 10000,
+                 normalize: bool = True,
+                 channel_last: bool = True,
+                 transpose_output: bool = False):
+        super().__init__()
+        dim = int(np.ceil(dim / 4) * 2)
+        self.dim = dim
+        inv_freq = 1.0 / (temperature**(torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq)
+        self.normalize = normalize
+        self.scale = scale
+        self.eps = 1e-6
+        self.channel_last = channel_last
+        self.transpose_output = transpose_output
+
+        self.cached_penc = None  # the cache is irrespective of the number of objects
+
+    def forward(self, tensor: torch.Tensor) -> torch.Tensor:
+        """
+        :param tensor: A 4/5d tensor of size 
+            channel_last=True: (batch_size, h, w, c) or (batch_size, k, h, w, c)
+            channel_last=False: (batch_size, c, h, w) or (batch_size, k, c, h, w)
+        :return: positional encoding tensor that has the same shape as the input if the input is 4d
+                 if the input is 5d, the output is broadcastable along the k-dimension
+        """
+        if len(tensor.shape) != 4 and len(tensor.shape) != 5:
+            raise RuntimeError(f'The input tensor has to be 4/5d, got {tensor.shape}!')
+
+        if len(tensor.shape) == 5:
+            # take a sample from the k dimension
+            num_objects = tensor.shape[1]
+            tensor = tensor[:, 0]
+        else:
+            num_objects = None
+
+        if self.channel_last:
+            batch_size, h, w, c = tensor.shape
+        else:
+            batch_size, c, h, w = tensor.shape
+
+        if self.cached_penc is not None and self.cached_penc.shape == tensor.shape:
+            if num_objects is None:
+                return self.cached_penc
+            else:
+                return self.cached_penc.unsqueeze(1)
+
+        self.cached_penc = None
+
+        pos_y = torch.arange(h, device=tensor.device, dtype=self.inv_freq.dtype)
+        pos_x = torch.arange(w, device=tensor.device, dtype=self.inv_freq.dtype)
+        if self.normalize:
+            pos_y = pos_y / (pos_y[-1] + self.eps) * self.scale
+            pos_x = pos_x / (pos_x[-1] + self.eps) * self.scale
+
+        sin_inp_y = torch.einsum("i,j->ij", pos_y, self.inv_freq)
+        sin_inp_x = torch.einsum("i,j->ij", pos_x, self.inv_freq)
+        emb_y = get_emb(sin_inp_y).unsqueeze(1)
+        emb_x = get_emb(sin_inp_x)
+
+        emb = torch.zeros((h, w, self.dim * 2), device=tensor.device, dtype=tensor.dtype)
+        emb[:, :, :self.dim] = emb_x
+        emb[:, :, self.dim:] = emb_y
+
+        if not self.channel_last and self.transpose_output:
+            # cancelled out
+            pass
+        elif (not self.channel_last) or (self.transpose_output):
+            emb = emb.permute(2, 0, 1)
+
+        self.cached_penc = emb.unsqueeze(0).repeat(batch_size, 1, 1, 1)
+        if num_objects is None:
+            return self.cached_penc
+        else:
+            return self.cached_penc.unsqueeze(1)
+
+
+if __name__ == '__main__':
+    device = get_default_device()
+    pe = PositionalEncoding(8).to(device)
+    input = torch.ones((1, 8, 8, 8), device=device)
+    output = pe(input)
+    # print(output)
+    print(output[0, :, 0, 0])
+    print(output[0, :, 0, 5])
+    print(output[0, 0, :, 0])
+    print(output[0, 0, 0, :])

vendor/MatAnyone2/matanyone2/model/transformer/transformer_layers.py

@@ -0,0 +1,161 @@
+# Modified from PyTorch nn.Transformer
+
+from typing import List, Callable
+
+import torch
+from torch import Tensor
+import torch.nn as nn
+import torch.nn.functional as F
+from matanyone2.model.channel_attn import CAResBlock
+
+
+class SelfAttention(nn.Module):
+    def __init__(self,
+                 dim: int,
+                 nhead: int,
+                 dropout: float = 0.0,
+                 batch_first: bool = True,
+                 add_pe_to_qkv: List[bool] = [True, True, False]):
+        super().__init__()
+        self.self_attn = nn.MultiheadAttention(dim, nhead, dropout=dropout, batch_first=batch_first)
+        self.norm = nn.LayerNorm(dim)
+        self.dropout = nn.Dropout(dropout)
+        self.add_pe_to_qkv = add_pe_to_qkv
+
+    def forward(self,
+                x: torch.Tensor,
+                pe: torch.Tensor,
+                attn_mask: bool = None,
+                key_padding_mask: bool = None) -> torch.Tensor:
+        x = self.norm(x)
+        if any(self.add_pe_to_qkv):
+            x_with_pe = x + pe
+            q = x_with_pe if self.add_pe_to_qkv[0] else x
+            k = x_with_pe if self.add_pe_to_qkv[1] else x
+            v = x_with_pe if self.add_pe_to_qkv[2] else x
+        else:
+            q = k = v = x
+
+        r = x
+        x = self.self_attn(q, k, v, attn_mask=attn_mask, key_padding_mask=key_padding_mask)[0]
+        return r + self.dropout(x)
+
+
+# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html#torch.nn.functional.scaled_dot_product_attention
+class CrossAttention(nn.Module):
+    def __init__(self,
+                 dim: int,
+                 nhead: int,
+                 dropout: float = 0.0,
+                 batch_first: bool = True,
+                 add_pe_to_qkv: List[bool] = [True, True, False],
+                 residual: bool = True,
+                 norm: bool = True):
+        super().__init__()
+        self.cross_attn = nn.MultiheadAttention(dim,
+                                                nhead,
+                                                dropout=dropout,
+                                                batch_first=batch_first)
+        if norm:
+            self.norm = nn.LayerNorm(dim)
+        else:
+            self.norm = nn.Identity()
+        self.dropout = nn.Dropout(dropout)
+        self.add_pe_to_qkv = add_pe_to_qkv
+        self.residual = residual
+
+    def forward(self,
+                x: torch.Tensor,
+                mem: torch.Tensor,
+                x_pe: torch.Tensor,
+                mem_pe: torch.Tensor,
+                attn_mask: bool = None,
+                *,
+                need_weights: bool = False) -> (torch.Tensor, torch.Tensor):
+        x = self.norm(x)
+        if self.add_pe_to_qkv[0]:
+            q = x + x_pe
+        else:
+            q = x
+
+        if any(self.add_pe_to_qkv[1:]):
+            mem_with_pe = mem + mem_pe
+            k = mem_with_pe if self.add_pe_to_qkv[1] else mem
+            v = mem_with_pe if self.add_pe_to_qkv[2] else mem
+        else:
+            k = v = mem
+        r = x
+        x, weights = self.cross_attn(q,
+                                     k,
+                                     v,
+                                     attn_mask=attn_mask,
+                                     need_weights=need_weights,
+                                     average_attn_weights=False)
+
+        if self.residual:
+            return r + self.dropout(x), weights
+        else:
+            return self.dropout(x), weights
+
+
+class FFN(nn.Module):
+    def __init__(self, dim_in: int, dim_ff: int, activation=F.relu):
+        super().__init__()
+        self.linear1 = nn.Linear(dim_in, dim_ff)
+        self.linear2 = nn.Linear(dim_ff, dim_in)
+        self.norm = nn.LayerNorm(dim_in)
+
+        if isinstance(activation, str):
+            self.activation = _get_activation_fn(activation)
+        else:
+            self.activation = activation
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        r = x
+        x = self.norm(x)
+        x = self.linear2(self.activation(self.linear1(x)))
+        x = r + x
+        return x
+
+
+class PixelFFN(nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.dim = dim
+        self.conv = CAResBlock(dim, dim)
+
+    def forward(self, pixel: torch.Tensor, pixel_flat: torch.Tensor) -> torch.Tensor:
+        # pixel: batch_size * num_objects * dim * H * W
+        # pixel_flat: (batch_size*num_objects) * (H*W) * dim
+        bs, num_objects, _, h, w = pixel.shape
+        pixel_flat = pixel_flat.view(bs * num_objects, h, w, self.dim)
+        pixel_flat = pixel_flat.permute(0, 3, 1, 2).contiguous()
+
+        x = self.conv(pixel_flat)
+        x = x.view(bs, num_objects, self.dim, h, w)
+        return x
+
+
+class OutputFFN(nn.Module):
+    def __init__(self, dim_in: int, dim_out: int, activation=F.relu):
+        super().__init__()
+        self.linear1 = nn.Linear(dim_in, dim_out)
+        self.linear2 = nn.Linear(dim_out, dim_out)
+
+        if isinstance(activation, str):
+            self.activation = _get_activation_fn(activation)
+        else:
+            self.activation = activation
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.linear2(self.activation(self.linear1(x)))
+        return x
+
+
+def _get_activation_fn(activation: str) -> Callable[[Tensor], Tensor]:
+    if activation == "relu":
+        return F.relu
+    elif activation == "gelu":
+        return F.gelu
+
+    raise RuntimeError("activation should be relu/gelu, not {}".format(activation))

vendor/MatAnyone2/matanyone2/model/utils/memory_utils.py

@@ -0,0 +1,107 @@
+import math
+import torch
+from typing import Optional, Union, Tuple
+
+
+# @torch.jit.script
+def get_similarity(mk: torch.Tensor,
+                   ms: torch.Tensor,
+                   qk: torch.Tensor,
+                   qe: torch.Tensor,
+                   add_batch_dim: bool = False,
+                   uncert_mask = None) -> torch.Tensor:
+    # used for training/inference and memory reading/memory potentiation
+    # mk: B x CK x [N]    - Memory keys
+    # ms: B x  1 x [N]    - Memory shrinkage
+    # qk: B x CK x [HW/P] - Query keys
+    # qe: B x CK x [HW/P] - Query selection
+    # Dimensions in [] are flattened
+    # Return: B*N*HW
+    if add_batch_dim:
+        mk, ms = mk.unsqueeze(0), ms.unsqueeze(0)
+        qk, qe = qk.unsqueeze(0), qe.unsqueeze(0)
+
+    CK = mk.shape[1]
+
+    mk = mk.flatten(start_dim=2)
+    ms = ms.flatten(start_dim=1).unsqueeze(2) if ms is not None else None
+    qk = qk.flatten(start_dim=2)
+    qe = qe.flatten(start_dim=2) if qe is not None else None
+    
+    # query token selection based on temporal sparsity
+    if uncert_mask is not None:
+        uncert_mask = uncert_mask.flatten(start_dim=2)
+        uncert_mask = uncert_mask.expand(-1, 64, -1)
+        qk = qk * uncert_mask
+        qe = qe * uncert_mask
+
+    if qe is not None:
+        # See XMem's appendix for derivation
+        mk = mk.transpose(1, 2)
+        a_sq = (mk.pow(2) @ qe)
+        two_ab = 2 * (mk @ (qk * qe))
+        b_sq = (qe * qk.pow(2)).sum(1, keepdim=True)
+        similarity = (-a_sq + two_ab - b_sq)
+    else:
+        # similar to STCN if we don't have the selection term
+        a_sq = mk.pow(2).sum(1).unsqueeze(2)
+        two_ab = 2 * (mk.transpose(1, 2) @ qk)
+        similarity = (-a_sq + two_ab)
+
+    if ms is not None:
+        similarity = similarity * ms / math.sqrt(CK)  # B*N*HW
+    else:
+        similarity = similarity / math.sqrt(CK)  # B*N*HW
+
+    return similarity
+
+
+def do_softmax(
+        similarity: torch.Tensor,
+        top_k: Optional[int] = None,
+        inplace: bool = False,
+        return_usage: bool = False) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    # normalize similarity with top-k softmax
+    # similarity: B x N x [HW/P]
+    # use inplace with care
+    if top_k is not None:
+        values, indices = torch.topk(similarity, k=top_k, dim=1)
+
+        x_exp = values.exp_()
+        x_exp /= torch.sum(x_exp, dim=1, keepdim=True)
+        if inplace:
+            similarity.zero_().scatter_(1, indices, x_exp)  # B*N*HW
+            affinity = similarity
+        else:
+            affinity = torch.zeros_like(similarity).scatter_(1, indices, x_exp)  # B*N*HW
+    else:
+        maxes = torch.max(similarity, dim=1, keepdim=True)[0]
+        x_exp = torch.exp(similarity - maxes)
+        x_exp_sum = torch.sum(x_exp, dim=1, keepdim=True)
+        affinity = x_exp / x_exp_sum
+        indices = None
+
+    if return_usage:
+        return affinity, affinity.sum(dim=2)
+
+    return affinity
+
+
+def get_affinity(mk: torch.Tensor, ms: torch.Tensor, qk: torch.Tensor,
+                 qe: torch.Tensor, uncert_mask = None) -> torch.Tensor:
+    # shorthand used in training with no top-k
+    similarity = get_similarity(mk, ms, qk, qe, uncert_mask=uncert_mask)
+    affinity = do_softmax(similarity)
+    return affinity
+
+def readout(affinity: torch.Tensor, mv: torch.Tensor, uncert_mask: torch.Tensor=None) -> torch.Tensor:
+    B, CV, T, H, W = mv.shape
+
+    mo = mv.view(B, CV, T * H * W)
+    mem = torch.bmm(mo, affinity)
+    if uncert_mask is not None:
+        uncert_mask = uncert_mask.flatten(start_dim=2).expand(-1, CV, -1)
+        mem = mem * uncert_mask
+    mem = mem.view(B, CV, H, W)
+
+    return mem

vendor/MatAnyone2/matanyone2/model/utils/parameter_groups.py

@@ -0,0 +1,72 @@
+import logging
+
+log = logging.getLogger()
+
+
+def get_parameter_groups(model, stage_cfg, print_log=False):
+    """
+    Assign different weight decays and learning rates to different parameters.
+    Returns a parameter group which can be passed to the optimizer.
+    """
+    weight_decay = stage_cfg.weight_decay
+    embed_weight_decay = stage_cfg.embed_weight_decay
+    backbone_lr_ratio = stage_cfg.backbone_lr_ratio
+    base_lr = stage_cfg.learning_rate
+
+    backbone_params = []
+    embed_params = []
+    other_params = []
+
+    embedding_names = ['summary_pos', 'query_init', 'query_emb', 'obj_pe']
+    embedding_names = [e + '.weight' for e in embedding_names]
+
+    # inspired by detectron2
+    memo = set()
+    for name, param in model.named_parameters():
+        if not param.requires_grad:
+            continue
+        # Avoid duplicating parameters
+        if param in memo:
+            continue
+        memo.add(param)
+
+        if name.startswith('module'):
+            name = name[7:]
+
+        inserted = False
+        if name.startswith('pixel_encoder.'):
+            backbone_params.append(param)
+            inserted = True
+            if print_log:
+                log.info(f'{name} counted as a backbone parameter.')
+        else:
+            for e in embedding_names:
+                if name.endswith(e):
+                    embed_params.append(param)
+                    inserted = True
+                    if print_log:
+                        log.info(f'{name} counted as an embedding parameter.')
+                    break
+
+        if not inserted:
+            other_params.append(param)
+
+    parameter_groups = [
+        {
+            'params': backbone_params,
+            'lr': base_lr * backbone_lr_ratio,
+            'weight_decay': weight_decay
+        },
+        {
+            'params': embed_params,
+            'lr': base_lr,
+            'weight_decay': embed_weight_decay
+        },
+        {
+            'params': other_params,
+            'lr': base_lr,
+            'weight_decay': weight_decay
+        },
+    ]
+
+    return parameter_groups

vendor/MatAnyone2/matanyone2/model/utils/resnet.py

@@ -0,0 +1,179 @@
+"""
+resnet.py - A modified ResNet structure
+We append extra channels to the first conv by some network surgery
+"""
+
+from collections import OrderedDict
+import math
+
+import torch
+import torch.nn as nn
+from torch.utils import model_zoo
+
+
+def load_weights_add_extra_dim(target, source_state, extra_dim=1):
+    new_dict = OrderedDict()
+
+    for k1, v1 in target.state_dict().items():
+        if 'num_batches_tracked' not in k1:
+            if k1 in source_state:
+                tar_v = source_state[k1]
+
+                if v1.shape != tar_v.shape:
+                    # Init the new segmentation channel with zeros
+                    # print(v1.shape, tar_v.shape)
+                    c, _, w, h = v1.shape
+                    pads = torch.zeros((c, extra_dim, w, h), device=tar_v.device)
+                    nn.init.orthogonal_(pads)
+                    tar_v = torch.cat([tar_v, pads], 1)
+
+                new_dict[k1] = tar_v
+
+    target.load_state_dict(new_dict)
+
+
+model_urls = {
+    'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
+    'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
+}
+
+
+def conv3x3(in_planes, out_planes, stride=1, dilation=1):
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     dilation=dilation,
+                     bias=False)
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, dilation=1):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(inplanes, planes, stride=stride, dilation=dilation)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes, stride=1, dilation=dilation)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, dilation=1):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes,
+                               planes,
+                               kernel_size=3,
+                               stride=stride,
+                               dilation=dilation,
+                               padding=dilation,
+                               bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * 4)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class ResNet(nn.Module):
+    def __init__(self, block, layers=(3, 4, 23, 3), extra_dim=0):
+        self.inplanes = 64
+        super(ResNet, self).__init__()
+        self.conv1 = nn.Conv2d(3 + extra_dim, 64, kernel_size=7, stride=2, padding=3, bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2. / n))
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilation=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes,
+                          planes * block.expansion,
+                          kernel_size=1,
+                          stride=stride,
+                          bias=False),
+                nn.BatchNorm2d(planes * block.expansion),
+            )
+
+        layers = [block(self.inplanes, planes, stride, downsample)]
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes, dilation=dilation))
+
+        return nn.Sequential(*layers)
+
+
+def resnet18(pretrained=True, extra_dim=0):
+    model = ResNet(BasicBlock, [2, 2, 2, 2], extra_dim)
+    if pretrained:
+        load_weights_add_extra_dim(model, model_zoo.load_url(model_urls['resnet18']), extra_dim)
+    return model
+
+
+def resnet50(pretrained=True, extra_dim=0):
+    model = ResNet(Bottleneck, [3, 4, 6, 3], extra_dim)
+    if pretrained:
+        load_weights_add_extra_dim(model, model_zoo.load_url(model_urls['resnet50']), extra_dim)
+    return model

vendor/MatAnyone2/matanyone2/utils/device.py

@@ -0,0 +1,33 @@
+import torch
+import functools
+
+def get_default_device():
+    if torch.cuda.is_available():
+        return torch.device("cuda")
+    elif torch.backends.mps.is_built() and torch.backends.mps.is_available():
+        return torch.device("mps")
+    else:
+        return torch.device("cpu")
+
+def safe_autocast_decorator(enabled=True):
+    def decorator(func):
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs):
+            device = get_default_device()
+            if device.type in ["cuda", "cpu"]:
+                with torch.amp.autocast(device_type=device.type, enabled=enabled):
+                    return func(*args, **kwargs)
+            else:
+                return func(*args, **kwargs)
+        return wrapper
+    return decorator
+
+import contextlib
+@contextlib.contextmanager
+def safe_autocast(enabled=True):
+    device = get_default_device()
+    if device.type in ["cuda", "cpu"]:
+        with torch.amp.autocast(device_type=device.type, enabled=enabled):
+            yield
+    else:
+        yield  # MPS or other unsupported backends skip autocast

vendor/MatAnyone2/matanyone2/utils/get_default_model.py

@@ -0,0 +1,27 @@
+"""
+A helper function to get a default model for quick testing
+"""
+from omegaconf import open_dict
+from hydra import compose, initialize
+
+import torch
+from matanyone2.model.matanyone2 import MatAnyone2
+
+def get_matanyone2_model(ckpt_path, device=None) -> MatAnyone2:
+    initialize(version_base='1.3.2', config_path="../config", job_name="eval_our_config")
+    cfg = compose(config_name="eval_matanyone_config")
+    
+    with open_dict(cfg):
+        cfg['weights'] = ckpt_path
+
+    # Load the network weights
+    if device is not None:
+        matanyone2 = MatAnyone2(cfg, single_object=True).to(device).eval()
+        model_weights = torch.load(cfg.weights, map_location=device)
+    else:  # if device is not specified, `.cuda()` by default
+        matanyone2 = MatAnyone2(cfg, single_object=True).cuda().eval()
+        model_weights = torch.load(cfg.weights)
+        
+    matanyone2.load_weights(model_weights)
+
+    return matanyone2

vendor/MatAnyone2/matanyone2/utils/inference_utils.py

@@ -0,0 +1,54 @@
+import os
+import cv2
+import random
+import numpy as np
+
+import torch
+import torchvision
+
+IMAGE_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.JPG', '.JPEG', '.PNG')
+VIDEO_EXTENSIONS = ('.mp4', '.mov', '.avi', '.MP4', '.MOV', '.AVI')
+
+def read_frame_from_videos(frame_root):
+    if frame_root.endswith(VIDEO_EXTENSIONS):  # Video file path
+        video_name = os.path.basename(frame_root)[:-4]
+        frames, _, info = torchvision.io.read_video(filename=frame_root, pts_unit='sec', output_format='TCHW') # RGB
+        fps = info['video_fps']
+    else:
+        video_name = os.path.basename(frame_root)
+        frames = []
+        fr_lst = sorted(os.listdir(frame_root))
+        for fr in fr_lst:
+            frame = cv2.imread(os.path.join(frame_root, fr))[...,[2,1,0]] # RGB, HWC
+            frames.append(frame)
+        fps = 24  # default
+        frames = torch.Tensor(np.array(frames)).permute(0, 3, 1, 2).contiguous() # TCHW
+    
+    length = frames.shape[0]
+
+    return frames, fps, length, video_name
+
+def get_video_paths(input_root):
+    video_paths = []
+    for root, _, files in os.walk(input_root):
+        for file in files:
+            if file.lower().endswith(VIDEO_EXTENSIONS):
+                video_paths.append(os.path.join(root, file))
+    return sorted(video_paths)
+
+def str_to_list(value):
+    return list(map(int, value.split(',')))
+
+def gen_dilate(alpha, min_kernel_size, max_kernel_size): 
+    kernel_size = random.randint(min_kernel_size, max_kernel_size)
+    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (kernel_size,kernel_size))
+    fg_and_unknown = np.array(np.not_equal(alpha, 0).astype(np.float32))
+    dilate = cv2.dilate(fg_and_unknown, kernel, iterations=1)*255
+    return dilate.astype(np.float32)
+
+def gen_erosion(alpha, min_kernel_size, max_kernel_size): 
+    kernel_size = random.randint(min_kernel_size, max_kernel_size)
+    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (kernel_size,kernel_size))
+    fg = np.array(np.equal(alpha, 255).astype(np.float32))
+    erode = cv2.erode(fg, kernel, iterations=1)*255
+    return erode.astype(np.float32)

vendor/MatAnyone2/matanyone2/utils/tensor_utils.py

@@ -0,0 +1,62 @@
+from typing import List, Iterable
+import torch
+import torch.nn.functional as F
+from matanyone2.utils.device import safe_autocast
+
+# STM
+def pad_divide_by(in_img: torch.Tensor, d: int) -> (torch.Tensor, Iterable[int]):
+    h, w = in_img.shape[-2:]
+
+    if h % d > 0:
+        new_h = h + d - h % d
+    else:
+        new_h = h
+    if w % d > 0:
+        new_w = w + d - w % d
+    else:
+        new_w = w
+    lh, uh = int((new_h - h) / 2), int(new_h - h) - int((new_h - h) / 2)
+    lw, uw = int((new_w - w) / 2), int(new_w - w) - int((new_w - w) / 2)
+    pad_array = (int(lw), int(uw), int(lh), int(uh))
+    out = F.pad(in_img, pad_array)
+    return out, pad_array
+
+
+def unpad(img: torch.Tensor, pad: Iterable[int]) -> torch.Tensor:
+    if len(img.shape) == 4:
+        if pad[2] + pad[3] > 0:
+            img = img[:, :, pad[2]:-pad[3], :]
+        if pad[0] + pad[1] > 0:
+            img = img[:, :, :, pad[0]:-pad[1]]
+    elif len(img.shape) == 3:
+        if pad[2] + pad[3] > 0:
+            img = img[:, pad[2]:-pad[3], :]
+        if pad[0] + pad[1] > 0:
+            img = img[:, :, pad[0]:-pad[1]]
+    elif len(img.shape) == 5:
+        if pad[2] + pad[3] > 0:
+            img = img[:, :, :, pad[2]:-pad[3], :]
+        if pad[0] + pad[1] > 0:
+            img = img[:, :, :, :, pad[0]:-pad[1]]
+    else:
+        raise NotImplementedError
+    return img
+
+
+# @torch.jit.script
+def aggregate(prob: torch.Tensor, dim: int) -> torch.Tensor:
+    with safe_autocast(enabled=False):
+        prob = prob.float()
+        new_prob = torch.cat([torch.prod(1 - prob, dim=dim, keepdim=True), prob],
+                             dim).clamp(1e-7, 1 - 1e-7)
+        logits = torch.log((new_prob / (1 - new_prob)))  # (0, 1) --> (-inf, inf)
+
+        return logits
+
+
+# @torch.jit.script
+def cls_to_one_hot(cls_gt: torch.Tensor, num_objects: int) -> torch.Tensor:
+    # cls_gt: B*1*H*W
+    B, _, H, W = cls_gt.shape
+    one_hot = torch.zeros(B, num_objects + 1, H, W, device=cls_gt.device).scatter_(1, cls_gt, 1)
+    return one_hot

vendor/segment-anything/.vendor_commit

@@ -0,0 +1 @@
+dca509fe793f601edb92606367a655c15ac00fdf

vendor/segment-anything/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.

vendor/segment-anything/README.md

@@ -0,0 +1,183 @@
+## Latest updates -- SAM 2: Segment Anything in Images and Videos
+
+Please check out our new release on [**Segment Anything Model 2 (SAM 2)**](https://github.com/facebookresearch/segment-anything-2).
+
+* SAM 2 code: https://github.com/facebookresearch/segment-anything-2
+* SAM 2 demo: https://sam2.metademolab.com/
+* SAM 2 paper: https://arxiv.org/abs/2408.00714
+
+ ![SAM 2 architecture](https://github.com/facebookresearch/segment-anything-2/blob/main/assets/model_diagram.png?raw=true)
+
+**Segment Anything Model 2 (SAM 2)** is a foundation model towards solving promptable visual segmentation in images and videos. We extend SAM to video by considering images as a video with a single frame. The model design is a simple transformer architecture with streaming memory for real-time video processing. We build a model-in-the-loop data engine, which improves model and data via user interaction, to collect [**our SA-V dataset**](https://ai.meta.com/datasets/segment-anything-video), the largest video segmentation dataset to date. SAM 2 trained on our data provides strong performance across a wide range of tasks and visual domains.
+
+# Segment Anything
+
+**[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>
+
+## 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
+
+```
+git clone git@github.com:facebookresearch/segment-anything.git
+cd segment-anything; 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
+
+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 SamPredictor, sam_model_registry
+sam = sam_model_registry["<model_type>"](checkpoint="<path/to/checkpoint>")
+predictor = SamPredictor(sam)
+predictor.set_image(<your_image>)
+masks, _, _ = predictor.predict(<input_prompts>)
+```
+
+or generate masks for an entire image:
+
+```
+from segment_anything import SamAutomaticMaskGenerator, sam_model_registry
+sam = sam_model_registry["<model_type>"](checkpoint="<path/to/checkpoint>")
+mask_generator = SamAutomaticMaskGenerator(sam)
+masks = mask_generator.generate(<your_image>)
+```
+
+Additionally, masks can be generated for images from the command line:
+
+```
+python scripts/amg.py --checkpoint <path/to/checkpoint> --model-type <model_type> --input <image_or_folder> --output <path/to/output>
+```
+
+See the examples notebooks on [using SAM with prompts](/notebooks/predictor_example.ipynb) and [automatically generating masks](/notebooks/automatic_mask_generator_example.ipynb) for more details.
+
+<p float="left">
+  <img src="assets/notebook1.png?raw=true" width="49.1%" />
+  <img src="assets/notebook2.png?raw=true" width="48.9%" />
+</p>
+
+## ONNX Export
+
+SAM's lightweight mask decoder can be exported to ONNX format so that it can be run in any environment that supports ONNX runtime, such as in-browser as showcased in the [demo](https://segment-anything.com/demo). Export the model with
+
+```
+python scripts/export_onnx_model.py --checkpoint <path/to/checkpoint> --model-type <model_type> --output <path/to/output>
+```
+
+See the [example notebook](https://github.com/facebookresearch/segment-anything/blob/main/notebooks/onnx_model_example.ipynb) for details on how to combine image preprocessing via SAM's backbone with mask prediction using the ONNX model. It is recommended to use the latest stable version of PyTorch for ONNX export.
+
+### Web demo
+
+The `demo/` folder has a simple one page React app which shows how to run mask prediction with the exported ONNX model in a web browser with multithreading. Please see [`demo/README.md`](https://github.com/facebookresearch/segment-anything/blob/main/demo/README.md) for more details.
+
+## <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["<model_type>"](checkpoint="<path/to/checkpoint>")
+```
+
+Click the links below to download the checkpoint for the corresponding model type.
+
+- **`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)
+
+## Dataset
+
+See [here](https://ai.facebook.com/datasets/segment-anything/) for an overview of the datastet. The dataset can be downloaded [here](https://ai.facebook.com/datasets/segment-anything-downloads/). By downloading the datasets you agree that you have read and accepted the terms of the SA-1B Dataset Research License.
+
+We save masks per image as a json file. It can be loaded as a dictionary in python in the below format.
+
+```python
+{
+    "image"                 : image_info,
+    "annotations"           : [annotation],
+}
+
+image_info {
+    "image_id"              : int,              # Image id
+    "width"                 : int,              # Image width
+    "height"                : int,              # Image height
+    "file_name"             : str,              # Image filename
+}
+
+annotation {
+    "id"                    : int,              # Annotation id
+    "segmentation"          : dict,             # Mask saved in COCO RLE format.
+    "bbox"                  : [x, y, w, h],     # 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
+    "stability_score"       : float,            # A measure of the mask's quality
+    "crop_box"              : [x, y, w, h],     # The crop of the image used to generate the mask, in XYWH format
+    "point_coords"          : [[x, y]],         # The point coordinates input to the model to generate the mask
+}
+```
+
+Image ids can be found in sa_images_ids.txt which can be downloaded using the above [link](https://ai.facebook.com/datasets/segment-anything-downloads/) as well.
+
+To decode a mask in COCO RLE format into binary:
+
+```
+from pycocotools import mask as mask_utils
+mask = mask_utils.decode(annotation["segmentation"])
+```
+
+See [here](https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocotools/mask.py) for more instructions to manipulate masks stored in RLE format.
+
+## License
+
+The model is licensed under the [Apache 2.0 license](LICENSE).
+
+## Contributing
+
+See [contributing](CONTRIBUTING.md) and the [code of conduct](CODE_OF_CONDUCT.md).
+
+## Contributors
+
+The Segment Anything project was made possible with the help of many contributors (alphabetical):
+
+Aaron Adcock, Vaibhav Aggarwal, Morteza Behrooz, Cheng-Yang Fu, Ashley Gabriel, Ahuva Goldstand, Allen Goodman, Sumanth Gurram, Jiabo Hu, Somya Jain, Devansh Kukreja, Robert Kuo, Joshua Lane, Yanghao Li, Lilian Luong, Jitendra Malik, Mallika Malhotra, William Ngan, Omkar Parkhi, Nikhil Raina, Dirk Rowe, Neil Sejoor, Vanessa Stark, Bala Varadarajan, Bram Wasti, Zachary Winstrom
+
+## Citing Segment Anything
+
+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}
+}
+```

vendor/segment-anything/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

vendor/segment-anything/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.
+          crop_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.
+          crop_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_like(data["boxes"][:, 0]),  # 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_like(data["boxes"][:, 0]),  # 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_like(boxes[:, 0]),  # 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

vendor/segment-anything/segment_anything/build_sam.py

@@ -0,0 +1,107 @@
+# 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
+
+
+def build_sam_vit_h(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1280,
+        encoder_depth=32,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[7, 15, 23, 31],
+        checkpoint=checkpoint,
+    )
+
+
+build_sam = build_sam_vit_h
+
+
+def build_sam_vit_l(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=1024,
+        encoder_depth=24,
+        encoder_num_heads=16,
+        encoder_global_attn_indexes=[5, 11, 17, 23],
+        checkpoint=checkpoint,
+    )
+
+
+def build_sam_vit_b(checkpoint=None):
+    return _build_sam(
+        encoder_embed_dim=768,
+        encoder_depth=12,
+        encoder_num_heads=12,
+        encoder_global_attn_indexes=[2, 5, 8, 11],
+        checkpoint=checkpoint,
+    )
+
+
+sam_model_registry = {
+    "default": build_sam_vit_h,
+    "vit_h": build_sam_vit_h,
+    "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,
+):
+    prompt_embed_dim = 256
+    image_size = 1024
+    vit_patch_size = 16
+    image_embedding_size = image_size // vit_patch_size
+    sam = Sam(
+        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,
+        ),
+        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

vendor/segment-anything/segment_anything/modeling/__init__.py

@@ -0,0 +1,11 @@
+# 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 .mask_decoder import MaskDecoder
+from .prompt_encoder import PromptEncoder
+from .transformer import TwoWayTransformer

vendor/segment-anything/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

vendor/segment-anything/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 (tuple(int, 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 (tuple(int, 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:
+        windows (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): 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

vendor/segment-anything/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
+        transformer 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 output
+        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

vendor/segment-anything/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

vendor/segment-anything/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: ImageEncoderViT,
+        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 prompts,
+                C is determined 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

vendor/segment-anything/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 attention 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

vendor/segment-anything/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[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        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_np = masks[0].detach().cpu().numpy()
+        iou_predictions_np = iou_predictions[0].detach().cpu().numpy()
+        low_res_masks_np = low_res_masks[0].detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    @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.
+          boxes (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

vendor/segment-anything/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.

vendor/segment-anything/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 unnecessary 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

vendor/segment-anything/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).to(torch.int64)
+        masks = masks[..., : prepadded_size[0], : prepadded_size[1]]  # type: ignore
+
+        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

vendor/segment-anything/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 the 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[2], image.shape[3], 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)

workflows/matanyone2_demo.json

@@ -16,7 +16,7 @@
       "outputs": [],
       "properties": {},
       "widgets_values": [
-        "MatAnyone2 — Video Matting (VHS)\n\n1. Load a video → Queue once → Editor auto-opens on first frame\n2. Left-click = positive point · Right-click = negative point → Apply\n3. Queue again → Full matte renders\n\nModels (auto-downloaded on first use):\n• MatAnyone 2     github.com/pq-yang/MatAnyone2/releases  →  models/matanyone/\n• SAM ViT-H       dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth  →  models/sams/\n\nRequires: ComfyUI-VideoHelperSuite  ·  github.com/Kosinkadink/ComfyUI-VideoHelperSuite"
+        "MatAnyone2 — Video Matting (VHS)\n\nBundled starter clip: matanyone2-demo-input.mp4\ninstall.py copies it into ComfyUI/input so this demo queues cleanly.\nSwap the Load Video node to your own footage any time.\n\n1. Queue once → Editor auto-opens on first frame\n2. Left-click = positive point · Right-click = negative point → Apply\n3. Queue again → Full matte renders\n\nModels (auto-downloaded on first use):\n• MatAnyone 2     github.com/pq-yang/MatAnyone2/releases  →  models/matanyone/\n• SAM ViT-H       dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth  →  models/sams/\n\nRequires: ComfyUI-VideoHelperSuite  ·  github.com/Kosinkadink/ComfyUI-VideoHelperSuite"
       ],
       "color": "#845b1c",
       "bgcolor": "#1d1309"
@@ -55,7 +55,7 @@
       ],
       "properties": { "Node name for S&R": "VHS_LoadVideo" },
       "widgets_values": {
-        "video": "",
+        "video": "matanyone2-demo-input.mp4",
         "force_rate": 0,
         "force_size": "Disabled",
         "custom_width": 512,
@@ -63,7 +63,20 @@
         "frame_load_cap": 0,
         "skip_first_frames": 0,
         "select_every_nth": 1,
-        "choose video to upload": "image"
+        "choose video to upload": "image",
+        "videopreview": {
+          "hidden": false,
+          "paused": false,
+          "params": {
+            "frame_load_cap": 0,
+            "skip_first_frames": 0,
+            "force_rate": 0,
+            "filename": "matanyone2-demo-input.mp4",
+            "type": "input",
+            "format": "video/mp4",
+            "select_every_nth": 1
+          }
+        }
       },
       "color": "#2e607c",
       "bgcolor": "#1a3a4c"