Deploy snapshot (LFS for demo images per .gitattributes)

.gitattributes

@@ -0,0 +1,5 @@
+# Hugging Face Hub stores these via Git LFS / Xet (plain PNG/JPG in git are rejected on push).
+demo_data/*.png filter=lfs diff=lfs merge=lfs -text
+demo_data/*.jpg filter=lfs diff=lfs merge=lfs -text
+demo_data/*.jpeg filter=lfs diff=lfs merge=lfs -text
+images/*.png filter=lfs diff=lfs merge=lfs -text

.github/workflows/check-headers.yml

@@ -0,0 +1,36 @@
+---
+    name: Check File Headers
+    
+    on:
+      push:
+        branches: [main]
+      pull_request:
+        branches: [main]
+    
+    jobs:
+      check-headers:
+        name: Check Python file headers
+        runs-on: ubuntu-latest
+        permissions:
+          contents: read
+    
+        steps:
+          - name: Checkout code
+            uses: actions/checkout@v3
+    
+          - name: Set up Python
+            uses: actions/setup-python@v4
+            with:
+              python-version: "3.10"
+    
+          - name: Check headers
+            run: |
+              python scripts/update_headers.py --check
+            continue-on-error: false
+    
+          - name: Provide fix instructions
+            if: failure()
+            run: |
+              echo "::error::Some files are missing proper headers."
+              echo "To fix this, run: python scripts/update_headers.py"
+              echo "Then commit the changes."

.github/workflows/codespell.yml

@@ -0,0 +1,21 @@
+---
+    name: Codespell
+    
+    on:
+      push:
+        branches: [main]
+      pull_request:
+        branches: [main]
+    
+    jobs:
+      codespell:
+        name: Check for spelling errors
+        runs-on: ubuntu-latest
+    
+        steps:
+          - name: Checkout
+            uses: actions/checkout@v3
+          - name: Codespell
+            uses: codespell-project/actions-codespell@v1
+            with:
+               ignore_words_list: prima-animal, mpjpe, uvd, xyz, hm36, cpn, dbb

.github/workflows/release-pypi.yml

@@ -0,0 +1,48 @@
+name: Update pypi release
+
+on:
+  push:
+    tags:
+      - 'v*.*.*'
+  pull_request:
+    branches:
+      - main
+    types:
+      - labeled
+      - opened
+      - edited
+      - synchronize
+      - reopened
+
+jobs:
+  release:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Cache dependencies
+        id: pip-cache
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/pip
+          key: ${{ runner.os }}-pip
+
+      - name: Install dependencies
+        run: |
+          pip install --upgrade pip
+          pip install wheel
+          # NOTE(stes) see https://github.com/pypa/twine/issues/1216#issuecomment-2629069669
+          pip install "packaging>=24.2"
+
+      - name: Checkout code
+        uses: actions/checkout@v3
+
+      - name: Build and publish to PyPI
+        if: ${{ github.event_name == 'push' }}
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.TWINE_API_KEY }}
+        run: |
+          pip install build twine
+          python3 -m build
+          ls dist/
+          python3 -m twine upload --verbose dist/*

.gitignore

@@ -0,0 +1,175 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+# Vscode
+.vscode/
+
+# Directory
+.gradio/
+demo_out/
+demo_out*/
+data/PRIMA*/
+data/backbone.pth
+logs/
+*.pth
+*.pkl
+datasets/

README.md

@@ -0,0 +1,220 @@
+---
+title: PRIMA Demo
+emoji: 🦮
+colorFrom: blue
+colorTo: green
+sdk: gradio
+python_version: "3.10"
+app_file: app.py
+startup_duration_timeout: 60m
+---
+
+# PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+
+This is the official implementation of the approach described in the preprint:
+
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation \
+Xiaohang Yu, Ti Wang, Mackenzie Weygandt Mathis
+
+![PRIMA teaser](images/teaser.png)
+
+
+---
+
+
+## 🚀 TL;DR
+PRIMA creates a 3D quadruped mesh from a single 2D image. It leverages BioCLIP-based biological priors for robust cross-species shape understanding, then applies test-time adaptation with 2D reprojection and auxiliary keypoint guidance to refine SMAL pose and shape predictions. 
+
+It further can be used to build Quadruped3D, a large-scale pseudo-3D dataset with diverse species and poses. 
+
+PRIMA achieves state-of-the-art results on Animal3D, CtrlAni3D, Quadruped2D, and Animal Kingdom datasets.
+
+## Installation
+
+### Install from PyPI
+
+> Recommended: Python 3.10 and a CUDA-enabled PyTorch installation.
+
+```bash
+conda create -n prima python=3.10 -y
+conda activate prima
+
+# Install PyTorch matching your CUDA (example: CUDA 11.8)
+pip install --index-url https://download.pytorch.org/whl/cu118 \
+    "torch==2.2.1" "torchvision==0.17.1" "torchaudio==2.2.1"
+
+# Install chumpy and PyTorch3D
+python -m pip install --no-build-isolation \
+      "git+https://github.com/mattloper/chumpy.git"
+python -m pip install --no-build-isolation \
+      "git+https://github.com/facebookresearch/pytorch3d.git"
+
+# Install PRIMA from PyPI
+pip install prima-animal
+```
+
+`prima-animal` includes demo runtime dependencies used by `demo.py`, `demo_tta.py`, and `app.py` (including Detectron2 and DeepLabCut).
+
+---
+
+## Demo
+
+### Checkpoints and data
+
+We provide an automated demo-download script for models hosted on Hugging Face.
+Use the helper script to download and place all demo assets automatically in `data/`:
+
+```bash
+python scripts/setup_demo_data.py --hf-repo-id MLAdaptiveIntelligence/PRIMA
+```
+
+Approximate download volume from Hugging Face is ~24 GB total
+(`s1ckpt.ckpt` ~10.2 GB + `s3ckpt.ckpt` ~10.2 GB + `amr_vitbb.pth` ~2.5 GB + SMAL files).
+Expected time is roughly:
+- 100 Mbps: ~35-45 minutes
+- 300 Mbps: ~12-18 minutes
+- 1 Gbps: ~4-8 minutes
+
+To avoid re-downloading completed assets, rerun without `--force`. The script now
+re-downloads only missing or invalid checkpoints.
+
+Expected files in that Hugging Face repo root:
+- `my_smpl_00781_4_all.pkl`
+- `my_smpl_data_00781_4_all.pkl`
+- `walking_toy_symmetric_pose_prior_with_cov_35parts.pkl`
+- `amr_vitbb.pth`
+- `config_s1_HYDRA.yaml`
+- `config_s3_HYDRA.yaml`
+- `s1ckpt.ckpt`
+- `s3ckpt.ckpt`
+
+### Demo (without TTA)
+
+Run animal detection + PRIMA 3D pose/shape inference:
+
+```bash
+python demo.py \
+  --checkpoint data/PRIMAS1/checkpoints/s1ckpt.ckpt \
+  --img_folder demo_data/ \
+  --out_folder demo_out/
+```
+
+Outputs are written to `demo_out/`.
+
+---
+
+### Demo (with TTA)
+
+`demo_tta.py` pipeline: specify learning rate and number of iterations:
+
+Example:
+
+```bash
+python demo_tta.py \
+  --checkpoint data/PRIMAS1/checkpoints/s1ckpt.ckpt \
+  --img_folder demo_data/ \
+  --out_folder demo_out_tta/ \
+  --tta_lr 1e-6 \
+  --tta_num_iters 30
+```
+
+Outputs are written to `demo_out_tta/` (before/after TTA renders, keypoints, and optional meshes).
+
+---
+
+### Gradio demo
+
+We also provide a simple Gradio-based web demo for interactive testing in the
+browser:
+
+```bash
+python app.py \
+  --checkpoint data/PRIMAS1/checkpoints/s1ckpt.ckpt \
+  --out_folder demo_out_tta_gradio/
+```
+
+This starts a local Gradio app (by default on http://127.0.0.1:7860), where
+you can upload images and visualize PRIMA predictions and adaptation results.
+
+#### Hugging Face Space (maintainers)
+
+Demo images under `demo_data/` and `images/teaser.png` are tracked with **Git LFS**
+(see `.gitattributes`) so they can be pushed to a Hugging Face Space under the Hub’s
+LFS / **Xet** bridge. Install tooling once:
+
+```bash
+brew install git-lfs git-xet
+git xet install
+git lfs install
+```
+
+Then from a clean checkout with LFS files present, deploy the Space repo:
+
+```bash
+./scripts/deploy_hf_space.sh
+```
+
+The script rsyncs the working tree (not `git archive`) so image files are materialized
+before `git add` turns them into LFS blobs.
+
+---
+
+
+## Training and Evaluation 
+
+### Dataset Setup
+
+Download datasets from [Animal3D](https://xujiacong.github.io/Animal3D/), [CtrlAni3D](https://github.com/luoxue-star/AniMer?tab=readme-ov-file#training), Quadruped2D, and [Animal Kingdom](https://drive.google.com/file/d/1dk2a0qB0fbVZ4X6eAgP6VJVXj0rxVfsJ/view?usp=drive_link). For Quadruped2D, download the images from [SuperAnimal-Quadruped80K](https://zenodo.org/records/14016777) and our processed annotations from [here](https://drive.google.com/drive/folders/1eBNboxVwl_eGPoC93zxf-U3hmE6e2f-f?usp=sharing). Put all the datasets under `datasets/`.
+
+### Training 
+
+Two-stage training script:
+
+```bash
+bash train.sh
+```
+
+Training outputs are written to `logs/train/runs/<exp_name>/`.
+
+
+### Evaluation
+
+```bash
+python eval.py \
+  --config data/PRIMAS1/.hydra/config.yaml \
+  --checkpoint data/PRIMAS1/checkpoints/s1ckpt.ckpt
+```
+
+Common values for `--dataset` are controlled by:
+- `configs_hydra/experiment/default_val.yaml`
+
+---
+
+
+## Acknowledgements
+
+This release builds on several open-source projects, including:
+- [Detectron2](https://github.com/facebookresearch/detectron2)
+- [BioCLIP](https://github.com/Imageomics/BioCLIP)
+- [AniMer](https://github.com/luoxue-star/AniMer)
+- [DeepLabCut](https://github.com/DeepLabCut/DeepLabCut)
+
+---
+
+## Citation
+
+If you use this code in your research, please cite our PRIMA paper.
+
+```bibtex
+@misc{yu_prima,
+  title={PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation},
+  author={Xiaohang Yu and Ti Wang and Mackenzie Weygandt Mathis},
+}
+```
+
+---
+
+## Contact
+
+For issues, please open a GitHub issue in this repository.

app.py

@@ -0,0 +1,561 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+"""Gradio demo for PRIMA + SuperAnimal + TTA.
+
+This script wraps the ``demo_tta.py`` pipeline into an interactive
+Gradio interface. The overall logic follows:
+
+1. Given an input image, run Detectron2 to detect animals.
+2. For each detected animal, run PRIMA for 3D pose/shape estimation.
+3. Run DeepLabCut SuperAnimal to obtain 2D keypoints.
+4. Map SuperAnimal 39 keypoints to the 26 PRIMA keypoints.
+5. Run test-time adaptation (TTA) with user-specified lr and iters.
+6. Render and save before/after TTA results and keypoint visualizations.
+
+"""
+
+import argparse
+import os
+import sys
+import tempfile
+import traceback
+from types import SimpleNamespace
+from typing import List, Tuple
+from pathlib import Path
+
+import cv2
+import gradio as gr
+import numpy as np
+import torch
+import torch.utils.data
+
+# Repo-local minimal ``chumpy`` shim (see ``chumpy/__init__.py``) so SMAL pickles load
+# without installing the full chumpy package in Space builds.
+_REPO_ROOT = Path(__file__).resolve().parent
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+
+# Default checkpoint path following README instructions
+DEFAULT_CHECKPOINT = "data/PRIMAS1/checkpoints/s1ckpt.ckpt"
+DEFAULT_HF_ASSET_REPO = "MLAdaptiveIntelligence/PRIMA"
+
+# Output folder for rendered images/meshes and keypoints
+DEFAULT_OUT_FOLDER = "demo_out_tta_gradio"
+
+
+def _is_truthy_env(var_name: str) -> bool:
+    return os.environ.get(var_name, "").strip().lower() in {"1", "true", "yes", "on"}
+
+
+def _running_on_space() -> bool:
+    return bool(os.environ.get("SPACE_ID") or os.environ.get("HF_SPACE_ID"))
+
+
+def _gradio_examples_for_interface() -> List[List]:
+    """Gradio prefetches example media at startup.
+
+    Demo images are tracked with Git LFS / Xet (see ``.gitattributes``) so they can live
+    in the Hugging Face Space repo. Use absolute paths only when files exist beside ``app.py``.
+    """
+    if _is_truthy_env("PRIMA_DISABLE_GRADIO_EXAMPLES"):
+        return []
+    rows: List[List] = []
+    template: List[Tuple[str, float, int, float, float, bool, bool]] = [
+        ("demo_data/000000015956_horse.png", 1e-6, 30, 0.7, 0.1, False, True),
+        ("demo_data/n02412080_12159.png", 1e-6, 30, 0.7, 0.1, False, True),
+        ("demo_data/000000315905_zebra.jpg", 1e-6, 30, 0.7, 0.1, False, True),
+        ("demo_data/beagle.jpg", 1e-6, 0, 0.7, 0.1, False, True),
+        ("demo_data/shepherd_hati.jpg", 1e-6, 0, 0.7, 0.1, False, True),
+    ]
+    for rel, *rest in template:
+        p = _REPO_ROOT / rel
+        if p.is_file():
+            rows.append([str(p), *rest])
+    return rows
+
+
+def _should_preload_assets() -> bool:
+    """Default to preload on Spaces; configurable via PRIMA_PRELOAD_ASSETS."""
+    preload_env = os.environ.get("PRIMA_PRELOAD_ASSETS")
+    if preload_env is not None:
+        return _is_truthy_env("PRIMA_PRELOAD_ASSETS")
+    return _running_on_space()
+
+
+def _ensure_demo_assets(checkpoint_path: str) -> None:
+    """Download required demo assets when running in a clean environment."""
+    from scripts.setup_demo_data import (
+        maybe_download_smal,
+        maybe_download_backbone,
+        maybe_download_stage,
+    )
+
+    checkpoint = Path(checkpoint_path)
+    data_dir = checkpoint.parents[2]
+    hf_repo_id = os.environ.get("PRIMA_HF_REPO_ID", DEFAULT_HF_ASSET_REPO)
+
+    maybe_download_smal(data_dir, force=False, hf_repo_id=hf_repo_id)
+    maybe_download_backbone(data_dir, force=False, hf_repo_id=hf_repo_id)
+    maybe_download_stage(
+        "PRIMAS1",
+        "config_s1_HYDRA.yaml",
+        "s1ckpt.ckpt",
+        "s1ckpt.ckpt",
+        data_dir,
+        force=False,
+        hf_repo_id=hf_repo_id,
+    )
+
+
+def _preload_assets_once(checkpoint_path: str) -> None:
+    checkpoint = Path(checkpoint_path)
+    cfg_path = checkpoint.parent.parent / ".hydra" / "config.yaml"
+    if checkpoint.exists() and cfg_path.exists():
+        print("[startup] Assets already present; skipping preload.")
+        return
+    print("[startup] Preloading demo assets from Hugging Face Hub...")
+    _ensure_demo_assets(checkpoint_path)
+    print("[startup] Asset preload complete.")
+
+
+def _load_prima_model(checkpoint_path: str = DEFAULT_CHECKPOINT):
+    """Load PRIMA model and renderer once for the Gradio app."""
+    from prima.models import load_prima
+    from prima.utils.renderer import Renderer
+
+    checkpoint = Path(checkpoint_path)
+    cfg_path = checkpoint.parent.parent / ".hydra" / "config.yaml"
+    if not checkpoint.exists() or not cfg_path.exists():
+        _ensure_demo_assets(checkpoint_path)
+    if not checkpoint.exists():
+        raise FileNotFoundError(
+            f"Missing checkpoint: {checkpoint}. Download demo checkpoints/data as described in README."
+        )
+    if not cfg_path.exists():
+        raise FileNotFoundError(
+            f"Missing model config: {cfg_path}. Ensure the full checkpoint folder layout from README is present."
+        )
+
+    model, model_cfg = load_prima(checkpoint_path)
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    model = model.to(device)
+    model.eval()
+
+    renderer = Renderer(model_cfg, faces=model.smal.faces)
+    return model, model_cfg, renderer, device
+
+
+def _build_detector():
+    """Build Detectron2 animal detector (same config as demo_tta/demo.py)."""
+    try:
+        import detectron2.config
+        import detectron2.engine
+        from detectron2 import model_zoo
+    except Exception as e:
+        print(f"[warn] Detectron2 unavailable ({type(e).__name__}: {e}); using full-image fallback bbox.")
+        return None
+
+    cfg = detectron2.config.get_cfg()
+    cfg.merge_from_file(
+        model_zoo.get_config_file("COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml")
+    )
+    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
+    cfg.MODEL.WEIGHTS = (
+        "https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/"
+        "faster_rcnn_X_101_32x8d_FPN_3x/139173657/model_final_68b088.pkl"
+    )
+    cfg.MODEL.DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+    detector = detectron2.engine.DefaultPredictor(cfg)
+    return detector
+
+# SuperAnimal defaults (same as in demo_tta parser)
+SUPER_ANIMAL_ARGS = SimpleNamespace(
+    superanimal_name="superanimal_quadruped",
+    superanimal_model_name="hrnet_w32",
+    superanimal_detector_name="fasterrcnn_resnet50_fpn_v2",
+    superanimal_max_individuals=1,
+)
+
+
+def _collect_animal_results(
+    model,
+    model_cfg,
+    renderer,
+    device,
+    detector,
+    out_folder: str,
+    img_rgb: np.ndarray,
+    tta_lr: float,
+    tta_num_iters: int,
+    det_thresh: float,
+    kp_conf_thresh: float,
+    side_view: bool,
+    save_mesh: bool,
+) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray], str | None, str | None]:
+    """Run detection + PRIMA + SuperAnimal + TTA on a single RGB image.
+
+    Returns:
+        before_imgs: list of HxWx3 RGB images (before TTA) for all animals
+        after_imgs: list of HxWx3 RGB images (after TTA) for all animals
+        kpt_imgs: list of HxWx3 RGB keypoint visualizations
+        first_before_mesh: path to first animal's before-TTA mesh (.obj) or None
+        first_after_mesh: path to first animal's after-TTA mesh (.obj) or None
+    """
+    from prima.utils import recursive_to
+    from prima.datasets.vitdet_dataset import ViTDetDataset
+    from demo_tta import (
+        ANIMAL_COCO_IDS,
+        denorm_patch_to_rgb,
+        map_superanimal_to_prima,
+        run_superanimal_on_patch,
+        save_keypoint_vis,
+        tta_optimize,
+    )
+
+    # Detect animals
+    img_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
+    if detector is None:
+        # Fallback for environments where Detectron2 is unavailable: process full image as one crop.
+        h, w = img_bgr.shape[:2]
+        boxes = np.array([[0.0, 0.0, float(max(1, w - 1)), float(max(1, h - 1))]], dtype=np.float32)
+    else:
+        det_out = detector(img_bgr)
+        det_instances = det_out["instances"]
+
+        valid_idx = [
+            i
+            for i, (c, s) in enumerate(zip(det_instances.pred_classes, det_instances.scores))
+            if (int(c) in ANIMAL_COCO_IDS) and (float(s) > float(det_thresh))
+        ]
+        if len(valid_idx) == 0:
+            return [], [], [], None, None
+
+        boxes = det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+    dataset = ViTDetDataset(model_cfg, img_bgr, boxes)
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=False, num_workers=0)
+
+    before_imgs: List[np.ndarray] = []
+    after_imgs: List[np.ndarray] = []
+    kpt_imgs: List[np.ndarray] = []
+    before_mesh_paths: List[str] = []
+    after_mesh_paths: List[str] = []
+
+    img_token = next(tempfile._get_candidate_names())
+
+    for batch in dataloader:
+        batch = recursive_to(batch, device)
+
+        with torch.no_grad():
+            out_before = model(batch)
+
+        animal_id = int(batch["animalid"][0])
+
+        # Save/render before TTA
+        img_fn = f"{img_token}"
+        from demo_tta import render_and_save  # imported lazily to avoid circular issues
+
+        render_and_save(
+            renderer,
+            out_before,
+            batch,
+            img_fn,
+            animal_id,
+            out_folder,
+            suffix="before_tta",
+            side_view=side_view,
+            save_mesh=save_mesh,
+        )
+
+        before_png_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_before_tta.png")
+        if os.path.exists(before_png_path):
+            before_bgr = cv2.imread(before_png_path)
+            if before_bgr is not None:
+                before_imgs.append(cv2.cvtColor(before_bgr, cv2.COLOR_BGR2RGB))
+
+        if save_mesh:
+            before_obj_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_before_tta.obj")
+            if os.path.exists(before_obj_path):
+                before_mesh_paths.append(before_obj_path)
+
+        if int(tta_num_iters) <= 0:
+            render_and_save(
+                renderer,
+                out_before,
+                batch,
+                img_fn,
+                animal_id,
+                out_folder,
+                suffix="after_tta",
+                side_view=side_view,
+                save_mesh=save_mesh,
+            )
+
+            after_png_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_after_tta.png")
+            if os.path.exists(after_png_path):
+                after_bgr = cv2.imread(after_png_path)
+                if after_bgr is not None:
+                    after_imgs.append(cv2.cvtColor(after_bgr, cv2.COLOR_BGR2RGB))
+
+            if save_mesh:
+                after_obj_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_after_tta.obj")
+                if os.path.exists(after_obj_path):
+                    after_mesh_paths.append(after_obj_path)
+            continue
+
+        # Prepare patch for SuperAnimal
+        patch_rgb = denorm_patch_to_rgb(batch["img"][0])
+        with tempfile.TemporaryDirectory(prefix=f"dlc_{img_fn}_{animal_id}_") as tmp_dir:
+            bodyparts_xyc = run_superanimal_on_patch(patch_rgb, SUPER_ANIMAL_ARGS, tmp_dir)
+
+        if bodyparts_xyc is None:
+            # No keypoints => skip TTA for this animal
+            continue
+
+        mapped_xyc = map_superanimal_to_prima(bodyparts_xyc)
+        mapped_xyc[mapped_xyc[:, 2] < float(kp_conf_thresh), 2] = 0.0
+
+        # Save keypoint visualization and npy
+        kpt_png_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_prima26_kpts.png")
+        save_keypoint_vis(patch_rgb, mapped_xyc, kpt_png_path)
+        npy_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_prima26_kpts.npy")
+        np.save(npy_path, mapped_xyc)
+
+        if os.path.exists(kpt_png_path):
+            kpt_bgr = cv2.imread(kpt_png_path)
+            if kpt_bgr is not None:
+                kpt_imgs.append(cv2.cvtColor(kpt_bgr, cv2.COLOR_BGR2RGB))
+
+        # Normalize keypoints to [-0.5, 0.5] as in demo_tta
+        patch_h, patch_w = patch_rgb.shape[:2]
+        mapped_norm = mapped_xyc.copy()
+        mapped_norm[:, 0] = mapped_norm[:, 0] / float(patch_w) - 0.5
+        mapped_norm[:, 1] = mapped_norm[:, 1] / float(patch_h) - 0.5
+        gt_kpts_norm = torch.from_numpy(mapped_norm[None]).to(device=device, dtype=batch["img"].dtype)
+
+        # Run TTA
+        out_after = tta_optimize(
+            model,
+            batch,
+            gt_kpts_norm,
+            num_iters=int(tta_num_iters),
+            lr=float(tta_lr),
+        )
+
+        render_and_save(
+            renderer,
+            out_after,
+            batch,
+            img_fn,
+            animal_id,
+            out_folder,
+            suffix="after_tta",
+            side_view=side_view,
+            save_mesh=save_mesh,
+        )
+
+        after_png_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_after_tta.png")
+        if os.path.exists(after_png_path):
+            after_bgr = cv2.imread(after_png_path)
+            if after_bgr is not None:
+                after_imgs.append(cv2.cvtColor(after_bgr, cv2.COLOR_BGR2RGB))
+
+        if save_mesh:
+            after_obj_path = os.path.join(out_folder, f"{img_fn}_{animal_id}_after_tta.obj")
+            if os.path.exists(after_obj_path):
+                after_mesh_paths.append(after_obj_path)
+
+    first_before_mesh = before_mesh_paths[0] if before_mesh_paths else None
+    first_after_mesh = after_mesh_paths[0] if after_mesh_paths else None
+
+    return before_imgs, after_imgs, kpt_imgs, first_before_mesh, first_after_mesh
+
+
+def build_demo(checkpoint_path: str = DEFAULT_CHECKPOINT, out_folder: str = DEFAULT_OUT_FOLDER) -> gr.Interface:
+    os.makedirs(out_folder, exist_ok=True)
+    runtime_cache = {
+        "model": None,
+        "model_cfg": None,
+        "renderer": None,
+        "device": None,
+        "detector": None,
+    }
+
+    def gradio_inference(
+        image: np.ndarray,
+        tta_lr: float,
+        tta_num_iters: int,
+        det_thresh: float,
+        kp_conf_thresh: float,
+        side_view: bool,
+        save_mesh: bool,
+    ):
+        """Wrapper for Gradio. ``image`` is an RGB numpy array.
+
+        Yields intermediate status so long first-run (Hub downloads + model load)
+        does not hit silent client/proxy timeouts.
+        """
+
+        if image is None:
+            yield None, None, None, "No image provided."
+            return
+
+        if image.dtype != np.uint8:
+            img_rgb = np.clip(image, 0, 255).astype(np.uint8)
+        else:
+            img_rgb = image
+
+        yield None, None, None, "Queued; preparing run…"
+
+        if runtime_cache["model"] is None:
+            yield (
+                None,
+                None,
+                None,
+                "First run: downloading demo assets from Hugging Face (large checkpoint) "
+                "and loading the model. This can take many minutes; status updates here "
+                "mean the session is still alive.",
+            )
+            try:
+                model, model_cfg, renderer, device = _load_prima_model(checkpoint_path)
+                detector = _build_detector()
+            except Exception:
+                yield None, None, None, f"Model initialization failed:\n{traceback.format_exc()}"
+                return
+            runtime_cache["model"] = model
+            runtime_cache["model_cfg"] = model_cfg
+            runtime_cache["renderer"] = renderer
+            runtime_cache["device"] = device
+            runtime_cache["detector"] = detector
+            yield None, None, None, "Model loaded. Running detection and inference…"
+
+        try:
+            before_imgs, after_imgs, kpt_imgs, mesh_before, mesh_after = _collect_animal_results(
+                runtime_cache["model"],
+                runtime_cache["model_cfg"],
+                runtime_cache["renderer"],
+                runtime_cache["device"],
+                runtime_cache["detector"],
+                out_folder,
+                img_rgb,
+                tta_lr=tta_lr,
+                tta_num_iters=tta_num_iters,
+                det_thresh=det_thresh,
+                kp_conf_thresh=kp_conf_thresh,
+                side_view=side_view,
+                save_mesh=save_mesh,
+            )
+        except Exception:
+            yield None, None, None, f"Inference failed:\n{traceback.format_exc()}"
+            return
+
+        first_before = before_imgs[0] if before_imgs else None
+        first_after = after_imgs[0] if after_imgs else None
+        first_kpts = kpt_imgs[0] if kpt_imgs else None
+        if first_before is None and first_after is None:
+            yield (
+                None,
+                None,
+                None,
+                "No output generated. Try an image with a clearly visible quadruped.",
+            )
+            return
+        yield first_before, first_after, first_kpts, "OK"
+
+    _gradio_examples = _gradio_examples_for_interface()
+    _iface_kw = dict(
+        fn=gradio_inference,
+        analytics_enabled=False,
+        cache_examples=False,
+        inputs=[
+            gr.Image(
+                label="Input image",
+                type="numpy",
+                sources=["upload", "clipboard"],
+            ),
+            gr.Slider(
+                label="TTA learning rate",
+                minimum=1e-7,
+                maximum=1e-4,
+                value=1e-6,
+                step=1e-7,
+            ),
+            gr.Slider(
+                label="TTA iterations",
+                minimum=0,
+                maximum=100,
+                value=30,
+                step=1,
+                info="Set to 0 to disable TTA and reuse the initial PRIMA prediction.",
+            ),
+            gr.Slider(
+                label="Detection threshold",
+                minimum=0.3,
+                maximum=0.9,
+                value=0.7,
+                step=0.05,
+            ),
+            gr.Slider(
+                label="Keypoint confidence threshold",
+                minimum=0.0,
+                maximum=1.0,
+                value=0.1,
+                step=0.05,
+            ),
+            gr.Checkbox(label="Render side view", value=False),
+            gr.Checkbox(label="Save meshes (.obj)", value=True),
+        ],
+        outputs=[
+            gr.Image(label="Before TTA"),
+            gr.Image(label="After TTA"),
+            gr.Image(label="PRIMA 26 keypoints"),
+            gr.Textbox(label="Status / Traceback", lines=12),
+        ],
+        title="PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation",
+        description=(
+            "Upload an animal image. The demo runs Detectron2 for animal detection, "
+            "PRIMA for 3D pose/shape, DeepLabCut SuperAnimal for 2D keypoints, and "
+            "test-time adaptation (TTA) with configurable learning rate and iterations. "
+            "Set TTA iterations to 0 to disable adaptation.\n\n"
+            "Results (PNG/OBJ and 26-keypoint visualizations) are saved under "
+            f"'{out_folder}'."
+        ),
+    )
+    if _gradio_examples:
+        _iface_kw["examples"] = _gradio_examples
+    demo = gr.Interface(**_iface_kw)
+    demo.queue(max_size=8, default_concurrency_limit=1)
+    return demo
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Gradio demo for PRIMA + SuperAnimal + TTA")
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        default=DEFAULT_CHECKPOINT,
+        help="Path to the pretrained PRIMA checkpoint",
+    )
+    parser.add_argument(
+        "--out_folder",
+        type=str,
+        default=DEFAULT_OUT_FOLDER,
+        help="Folder used to save rendered outputs and meshes",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    if _should_preload_assets():
+        _preload_assets_once(args.checkpoint)
+    demo = build_demo(checkpoint_path=args.checkpoint, out_folder=args.out_folder)
+    demo.launch()

chumpy/__init__.py

@@ -0,0 +1,16 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+"""Minimal ``chumpy`` compatibility for unpickling legacy SMAL model configs."""
+
+from __future__ import annotations
+
+from .ch import Ch, ChArray
+
+__all__ = ["Ch", "ChArray"]

chumpy/ch.py

@@ -0,0 +1,37 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+"""``chumpy.ch`` namespace expected by legacy SMAL pickles."""
+
+from __future__ import annotations
+
+import numpy as np
+
+
+class Ch:
+    """Minimal stand-in for ``chumpy.ch.Ch`` (unpickling only)."""
+
+    def __init__(self, *args, **kwargs):
+        self._data = None
+        if args:
+            self._data = np.asarray(args[0])
+
+    def r(self):
+        if self._data is None:
+            return np.zeros((), dtype=np.float32)
+        return np.asarray(self._data)
+
+
+class ChArray(np.ndarray):
+    """Minimal stand-in for ``chumpy.ch.ChArray``."""
+
+    pass
+
+
+__all__ = ["Ch", "ChArray"]

configs_hydra/experiment/default.yaml

@@ -0,0 +1,28 @@
+# @package _global_
+
+SMAL:
+  DATA_DIR: data/smal
+  MODEL_PATH: data/smal/my_smpl_00781_4_all.pkl
+  SHAPE_PRIOR_PATH: data/smal/my_smpl_data_00781_4_all.pkl
+  POSE_PRIOR_PATH: data/smal/walking_toy_symmetric_pose_prior_with_cov_35parts.pkl
+  NUM_JOINTS: 34
+
+EXTRA:
+  FOCAL_LENGTH: 1000
+  NUM_LOG_IMAGES: 4
+  NUM_LOG_SAMPLES_PER_IMAGE: 4
+  PELVIS_IND: 0
+
+DATASETS:
+  CONFIG:
+    SCALE_FACTOR: 0.3
+    ROT_FACTOR: 30
+    TRANS_FACTOR: 0.02
+    COLOR_SCALE: 0.2
+    ROT_AUG_RATE: 0.6
+    TRANS_AUG_RATE: 0.5
+    DO_FLIP: False
+    FLIP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_LEVEL: 1
+    

configs_hydra/experiment/default_val.yaml

@@ -0,0 +1,34 @@
+# @package _global_
+
+DATASETS:
+  ANIMAL3D:
+    ROOT_IMAGE: ./datasets/animal3d/
+    JSON_FILE:
+      TEST: ./datasets/animal3d/test.json
+  CONTROL_ANIMAL3D:
+    ROOT_IMAGE: ./datasets/control_animal3dlatest/  
+    JSON_FILE:
+      TEST: ./datasets/control_animal3dlatest/test.json
+  QUADRUPED2D:
+    ROOT_IMAGE: ./datasets/quadruped2d/  
+    JSON_FILE:
+      TEST: ./datasets/quadruped2d/test.json
+  ANIMAL_KINGDOM:
+    ROOT_IMAGE: ./datasets/Animal_Kingdom_test/  
+    JSON_FILE:
+      TEST: ./datasets/Animal_Kingdom_test/test.json
+  CONFIG:
+    SCALE_FACTOR: 0.0
+    ROT_FACTOR: 0
+    TRANS_FACTOR: 0.0
+    COLOR_SCALE: 0.0
+    ROT_AUG_RATE: 0.0
+    TRANS_AUG_RATE: 0.0
+    DO_FLIP: False
+    FLIP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_RATE: 0.0
+    EXTREME_CROP_AUG_LEVEL: 1
+
+METRIC:
+  PCK_THRESHOLD: [0.10, 0.15]
+    

configs_hydra/experiment/primaStage1.yaml

@@ -0,0 +1,83 @@
+# @package _global_
+
+defaults:
+  - default.yaml
+
+GENERAL:
+
+
+  TOTAL_STEPS: 63_000 
+  LOG_STEPS: 63 
+  VAL_STEPS: 63
+  VAL_EPOCHS: 1
+  CHECKPOINT_EPOCHS: 1
+  CHECKPOINT_SAVE_TOP_K: 2
+  NUM_WORKERS: 8
+  PREFETCH_FACTOR: 2
+
+LOSS_WEIGHTS:
+  KEYPOINTS_3D: 0.05  
+  KEYPOINTS_2D: 0.01
+  INTERMEDIATE_KP2D: 0.001
+  INTERMEDIATE_KP3D: 0.001  
+  GLOBAL_ORIENT: 0.005  
+  POSE: 0.001  
+  BETAS: 0.0005 
+  TRANSL: 0.0005
+  ADVERSARIAL: 0.0005
+  SUPCON: 0.0005
+
+
+TRAIN:
+  LR: 3.75e-6
+  WEIGHT_DECAY: 1e-4
+  BATCH_SIZE: 48
+  LOSS_REDUCTION: mean
+  NUM_TRAIN_SAMPLES: 2
+  NUM_TEST_SAMPLES: 64
+  POSE_2D_NOISE_RATIO: 0.01
+  SMPL_PARAM_NOISE_RATIO: 0.005
+
+MODEL:
+  IMAGE_SIZE: 256
+  IMAGE_MEAN: [0.485, 0.456, 0.406]
+  IMAGE_STD: [0.229, 0.224, 0.225]
+  BACKBONE:
+    TYPE: vith
+    PRETRAINED_WEIGHTS: ./data/amr_vitbb.pth
+    FREEZE: False
+
+  # Enable BioClip embedding
+  USE_BIOCLIP_EMBEDDING: True
+  BIOCLIP_EMBEDDING:
+    EMBED_DIM: 1280  # Match DINOv2 output dimension for token-wise concatenation
+    TYPE: bioclip1
+
+  # Enable 2D keypoint embedding for initialization; NewBioGuidedSMALPoseDecoder updates it dynamically
+  USE_KEYPOINT_EMBEDDING: False
+
+  SMAL_HEAD:
+    TYPE: new_bio_pose_transformer_decoder    # Use the newer version with SAM3D-style hierarchical updates
+    IN_CHANNELS: 1280
+    IEF_ITERS: 3
+    
+    # Pose Transformer Decoder configuration
+    DECODER_DIM: 1280
+    NUM_DECODER_LAYERS: 6
+    NUM_HEADS: 8
+    MLP_RATIO: 4.0
+    
+    # Keypoint token configuration specific to NewBioGuidedSMALPoseDecoder
+    USE_KEYPOINT_2D_TOKENS: True            # Enable 2D keypoint tokens with SAM3D-style dynamic updates
+    USE_KEYPOINT_3D_TOKENS: True            # Enable 3D keypoint tokens with pelvis normalization
+    KEYPOINT_TOKEN_UPDATE: True             # Enable hierarchical keypoint prediction and token updates
+    KP2D_INJECT_IMAGE_FEAT: True            # Key setting: inject image features via grid_sample
+    
+
+DATASETS:
+  ANIMAL3D:
+    ROOT_IMAGE: ./datasets/animal3d/ 
+    JSON_FILE:
+      TRAIN: ./datasets/animal3d/train.json
+      TEST: ./datasets/animal3d/test.json
+    WEIGHT: 1.0

configs_hydra/experiment/primaStage2.yaml

@@ -0,0 +1,113 @@
+# @package _global_
+
+defaults:
+  - default.yaml
+
+GENERAL:
+
+
+  TOTAL_STEPS: 450_000 
+  LOG_STEPS: 533
+  VAL_STEPS: 533  
+  VAL_EPOCHS: 1
+  CHECKPOINT_EPOCHS: 1
+  CHECKPOINT_SAVE_TOP_K: 2
+  NUM_WORKERS: 2
+  PREFETCH_FACTOR: 2
+
+LOSS_WEIGHTS:
+  KEYPOINTS_3D: 0.05  
+  KEYPOINTS_2D: 0.01
+  INTERMEDIATE_KP2D: 0.001
+  INTERMEDIATE_KP3D: 0.001
+  GLOBAL_ORIENT: 0.005  
+  POSE: 0.001  
+  BETAS: 0.0005 
+  TRANSL: 0.0005
+  ADVERSARIAL: 0.0
+  SUPCON: 0.0005
+
+
+TRAIN:
+  LR: 3.75e-6
+  WEIGHT_DECAY: 1e-4
+  BATCH_SIZE: 48
+  LOSS_REDUCTION: mean
+  NUM_TRAIN_SAMPLES: 2
+  NUM_TEST_SAMPLES: 64
+  POSE_2D_NOISE_RATIO: 0.01
+  SMPL_PARAM_NOISE_RATIO: 0.005
+
+MODEL:
+  IMAGE_SIZE: 256
+  IMAGE_MEAN: [0.485, 0.456, 0.406]
+  IMAGE_STD: [0.229, 0.224, 0.225]
+  BACKBONE:
+    TYPE: vith
+    PRETRAINED_WEIGHTS: ./data/amr_vitbb.pth
+    FREEZE: False
+
+  # Enable BioClip embedding
+  USE_BIOCLIP_EMBEDDING: True
+  BIOCLIP_EMBEDDING:
+    EMBED_DIM: 1280  # Match vit output dimension for token-wise concatenation
+    TYPE: bioclip1
+
+  # Enable 2D keypoint embedding
+  USE_KEYPOINT_EMBEDDING: False
+  KEYPOINT_EMBEDDING:
+    NUM_KEYPOINTS: 26        # Number of SMAL keypoints
+    KEYPOINT_DIM: 2          # 2D coordinates (x, y)
+    EMBED_DIM: 1280          # Match vit output dimension
+    HIDDEN_DIM: 512          # Hidden layer dimension in MLP
+    TYPE: 'token'            # Use token-based embedding (recommended)
+  
+  SMAL_HEAD:
+    TYPE: new_bio_pose_transformer_decoder    # Use the newer version with SAM3D-style hierarchical updates
+    IN_CHANNELS: 1280
+    IEF_ITERS: 1
+    
+    # Pose Transformer Decoder configuration
+    DECODER_DIM: 1280
+    NUM_DECODER_LAYERS: 6
+    NUM_HEADS: 8
+    MLP_RATIO: 4.0
+    
+    # Keypoint token configuration specific to NewBioGuidedSMALPoseDecoder
+    USE_KEYPOINT_2D_TOKENS: True            # Enable 2D keypoint tokens with SAM3D-style dynamic updates
+    USE_KEYPOINT_3D_TOKENS: True            # Enable 3D keypoint tokens with pelvis normalization
+    KEYPOINT_TOKEN_UPDATE: True             # Enable hierarchical keypoint prediction and token updates
+    KP2D_INJECT_IMAGE_FEAT: True            # Key setting: inject image features via grid_sample
+    
+    # Legacy transformer config (kept for compatibility)
+    TRANSFORMER_DECODER:
+      depth: 6
+      heads: 8
+      mlp_dim: 1024
+      dim_head: 64
+      dropout: 0.0
+      emb_dropout: 0.0
+      norm: layer
+      context_dim: 1280
+
+    
+
+DATASETS:
+  ANIMAL3D:
+    ROOT_IMAGE: ./datasets/animal3d/ 
+    JSON_FILE:
+      TRAIN: ./datasets/animal3d/train.json
+      TEST: ./datasets/animal3d/test.json
+    WEIGHT: 1.0
+  CONTROL_ANIMAL3D:
+    ROOT_IMAGE: ./datasets/control_animal3dlatest/  
+    JSON_FILE:
+      TRAIN: ./datasets/control_animal3dlatest/train.json
+      TEST: ./datasets/control_animal3dlatest/test.json
+    WEIGHT: 0.5
+  QUADRUPED2D:
+    ROOT_IMAGE: ./datasets/quadruped2d/  
+    JSON_FILE:
+      TRAIN: ./datasets/quadruped2d/train.json
+      TEST: ./datasets/quadruped2d/test.json
+    WEIGHT: 0.15

configs_hydra/extras/default.yaml

@@ -0,0 +1,8 @@
+# disable python warnings if they annoy you
+ignore_warnings: False
+
+# ask user for tags if none are provided in the config
+enforce_tags: True
+
+# pretty print config tree at the start of the run using Rich library
+print_config: True

configs_hydra/hydra/default.yaml

@@ -0,0 +1,26 @@
+# @package _global_
+# https://hydra.cc/docs/configure_hydra/intro/
+
+# enable color logging
+defaults:
+  - override /hydra/hydra_logging: colorlog
+  - override /hydra/job_logging: colorlog
+
+# exp_name: ovrd_${hydra:job.override_dirname}
+exp_name: ${now:%Y-%m-%d}_${now:%H-%M-%S}
+
+hydra:
+  run:
+    dir: ${paths.log_dir}/${task_name}/runs/${exp_name}
+  sweep:
+    dir: ${paths.log_dir}/${task_name}/multiruns/${exp_name}
+    subdir: ${hydra.job.num}
+  job:
+    config:
+      override_dirname:
+        exclude_keys:
+          - trainer
+          - trainer.devices
+          - trainer.num_nodes
+          - callbacks
+          - debug

configs_hydra/launcher/local.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+defaults:
+  - override /hydra/launcher: submitit_local
+
+hydra:
+  launcher:
+    timeout_min: 10_080   # 7 days
+    nodes: 1
+    tasks_per_node: ${trainer.devices}
+    cpus_per_task: 8
+    gpus_per_node: ${trainer.devices}
+    name: amr

configs_hydra/launcher/slurm.yaml

@@ -0,0 +1,22 @@
+# @package _global_
+
+defaults:
+  - override /hydra/launcher: submitit_slurm
+
+hydra:
+  launcher:
+    timeout_min: 10_080   # 7 days
+    max_num_timeout: 3
+    partition: g40
+    qos: idle
+    nodes: 1
+    tasks_per_node: ${trainer.devices}
+    gpus_per_task: null
+    cpus_per_task: 12
+    gpus_per_node: ${trainer.devices}
+    cpus_per_gpu: null
+    comment: prima
+    name: prima
+    setup:
+      - module load cuda openmpi libfabric-aws
+      - export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

configs_hydra/paths/default.yaml

@@ -0,0 +1,18 @@
+# path to root directory
+# this requires PROJECT_ROOT environment variable to exist
+# PROJECT_ROOT is inferred and set by pyrootutils package in `train.py` and `eval.py`
+root_dir: ${oc.env:PROJECT_ROOT}
+
+# path to data directory
+data_dir: ${paths.root_dir}/data/
+
+# path to logging directory
+log_dir: logs/
+
+# path to output directory, created dynamically by hydra
+# path generation pattern is specified in `configs/hydra/default.yaml`
+# use it to store all files generated during the run, like ckpts and metrics
+output_dir: ${hydra:runtime.output_dir}
+
+# path to working directory
+work_dir: ${hydra:runtime.cwd}

configs_hydra/train.yaml

@@ -0,0 +1,46 @@
+# @package _global_
+
+# specify here default configuration
+# order of defaults determines the order in which configs override each other
+defaults:
+  - _self_
+  - trainer: ddp.yaml
+  - paths: default.yaml
+  - extras: default.yaml
+  - hydra: default.yaml
+
+  # experiment configs allow for version control of specific hyperparameters
+  # e.g. best hyperparameters for given model and datamodule
+  - experiment: null
+  - texture_exp: null
+
+  # optional local config for machine/user specific settings
+  # it's optional since it doesn't need to exist and is excluded from version control
+  - optional launcher: local.yaml
+  # - optional launcher: slurm.yaml
+
+  # debugging config (enable through command line, e.g. `python train.py debug=default)
+  - debug: null
+
+# task name, determines output directory path
+task_name: "train"
+
+# tags to help you identify your experiments
+# you can overwrite this in experiment configs
+# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
+# appending lists from command line is currently not supported :(
+# https://github.com/facebookresearch/hydra/issues/1547
+tags: ["dev"]
+
+# set False to skip model training
+train: True
+
+# evaluate on test set, using best model weights achieved during training
+# lightning chooses best weights based on the metric specified in checkpoint callback
+test: False
+
+# simply provide checkpoint path to resume training
+ckpt_path: True
+
+# seed for random number generators in pytorch, numpy and python.random
+seed: null

configs_hydra/trainer/cpu.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_amr.yaml
+
+accelerator: cpu
+devices: 1

configs_hydra/trainer/ddp.yaml

@@ -0,0 +1,14 @@
+defaults:
+  - default.yaml
+  - default_amr.yaml
+
+# use "ddp_spawn" instead of "ddp",
+# it's slower but normal "ddp" currently doesn't work ideally with hydra
+# https://github.com/facebookresearch/hydra/issues/2070
+# https://pytorch-lightning.readthedocs.io/en/latest/accelerators/gpu_intermediate.html#distributed-data-parallel-spawn
+strategy: ddp_spawn
+
+accelerator: gpu
+devices: 2
+num_nodes: 1
+sync_batchnorm: True

configs_hydra/trainer/default.yaml

@@ -0,0 +1,10 @@
+_target_: pytorch_lightning.Trainer
+
+default_root_dir: ${paths.output_dir}
+
+accelerator: gpu
+devices: 1
+
+# set True to to ensure deterministic results
+# makes training slower but gives more reproducibility than just setting seeds
+deterministic: False

configs_hydra/trainer/default_amr.yaml

@@ -0,0 +1,9 @@
+num_sanity_val_steps: 0
+log_every_n_steps: ${GENERAL.LOG_STEPS}
+val_check_interval: ${GENERAL.VAL_STEPS}  # How often within one training epoch to check the validation set. 
+check_val_every_n_epoch: ${GENERAL.VAL_EPOCHS}  # Check val every n train epochs.
+precision: 16-mixed  # 16-mixed, 32
+max_steps: ${GENERAL.TOTAL_STEPS}
+# move_metrics_to_cpu: True
+limit_val_batches: 80  # How much of validation dataset to check. 
+# track_grad_norm: -1

configs_hydra/trainer/gpu.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_amr.yaml
+
+accelerator: gpu
+devices: 1

configs_hydra/trainer/mps.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default.yaml
+  - default_amr.yaml
+
+accelerator: mps
+devices: 1

demo.py

@@ -0,0 +1,144 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+from pathlib import Path
+import detectron2.config
+import detectron2.engine
+import torch
+import argparse
+import os
+import cv2
+import numpy as np
+from tqdm import tqdm
+import torch.utils
+import torch.utils.data
+from prima.models import load_prima
+from prima.utils import recursive_to
+from prima.datasets.vitdet_dataset import ViTDetDataset, DEFAULT_MEAN, DEFAULT_STD
+from prima.utils.renderer import Renderer, cam_crop_to_full
+import detectron2
+from detectron2 import model_zoo
+import warnings
+warnings.filterwarnings("ignore")
+
+LIGHT_BLUE = (0.65098039, 0.74117647, 0.85882353)
+GREEN = (0.65, 0.86, 0.74)
+
+
+
+def main():
+    parser = argparse.ArgumentParser(description='prima demo code')
+    parser.add_argument('--checkpoint', type=str,
+                        help='Path to pretrained model checkpoint')
+    parser.add_argument('--img_folder', type=str, default='demo_data/', help='Folder with input images')
+    parser.add_argument('--out_folder', type=str, default='demo_out', help='Output folder to save rendered results')
+    parser.add_argument('--side_view', dest='side_view', action='store_true', default=False,
+                        help='If set, render side view also')
+    parser.add_argument('--save_mesh', dest='save_mesh', action='store_true', default=False,
+                        help='If set, save meshes to disk also')
+    parser.add_argument('--batch_size', type=int, default=1, help='Batch size for inference/fitting')
+    parser.add_argument('--file_type', nargs='+', default=['*.jpg', '*.png', '*.jpeg', '*.JPEG'],
+                        help='List of file extensions to consider')
+
+    args = parser.parse_args()
+
+    model, model_cfg = load_prima(args.checkpoint)
+
+    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+    model = model.to(device)
+    model.eval()
+
+    # Setup the renderer
+    renderer = Renderer(model_cfg, faces=model.smal.faces)
+
+    # Make output directory if it does not exist
+    os.makedirs(args.out_folder, exist_ok=True)
+
+    # Load detector
+    cfg = detectron2.config.get_cfg()
+    cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"))
+    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5 
+    cfg.MODEL.WEIGHTS = "https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x/139173657/model_final_68b088.pkl" 
+    detector = detectron2.engine.DefaultPredictor(cfg)
+
+    img_paths = sorted([img for end in args.file_type for img in Path(args.img_folder).glob(end)])
+    for img_path in img_paths:
+        img_bgr = cv2.imread(str(img_path))
+        if img_bgr is None:
+            print(f"[WARN] Cannot read image: {img_path}")
+            continue
+        # Detect animals in image
+        det_out = detector(img_bgr)
+
+        det_instances = det_out['instances']
+        valid_idx = [i for i, (c, s) in enumerate(zip(det_instances.pred_classes, det_instances.scores)) if ((c in [15, 16, 17, 18, 19, 21, 22]) & (s > 0.7))]
+        boxes = det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+        # Run PRIMA on detected animals
+        dataset = ViTDetDataset(model_cfg, img_bgr, boxes)
+        dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, shuffle=False, num_workers=0)
+        for batch in tqdm(dataloader):
+            batch = recursive_to(batch, device)
+            with torch.no_grad():
+                out = model(batch)
+
+            pred_cam = out['pred_cam']
+            box_center = batch["box_center"].float()
+            box_size = batch["box_size"].float()
+            img_size = batch["img_size"].float()
+            scaled_focal_length = model_cfg.EXTRA.FOCAL_LENGTH / model_cfg.MODEL.IMAGE_SIZE * img_size.max()
+            pred_cam_t_full = cam_crop_to_full(pred_cam, box_center, box_size, img_size,
+                                               scaled_focal_length).detach().cpu().numpy()
+
+            # Render the result
+            batch_size = batch['img'].shape[0]
+            for n in range(batch_size):
+                # Get filename from path img_path
+                img_fn, _ = os.path.splitext(os.path.basename(img_path))
+                animal_id = int(batch['animalid'][n])
+                white_img = (torch.ones_like(batch['img'][n]).cpu() - DEFAULT_MEAN[:, None, None] / 255) / (
+                            DEFAULT_STD[:, None, None] / 255)
+                input_patch = (batch['img'][n].cpu() * (DEFAULT_STD[:, None, None]) + (
+                            DEFAULT_MEAN[:, None, None])) / 255.
+                input_patch = input_patch.permute(1, 2, 0).numpy()
+
+                regression_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                        out['pred_cam_t'][n].detach().cpu().numpy(),
+                                        batch['img'][n],
+                                        mesh_base_color=GREEN,
+                                        scene_bg_color=(1, 1, 1),
+                                            )
+
+                final_img = np.concatenate([input_patch, regression_img], axis=1)
+
+                if args.side_view:
+                    side_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                        out['pred_cam_t'][n].detach().cpu().numpy(),
+                                        white_img,
+                                        mesh_base_color=GREEN,
+                                        scene_bg_color=(1, 1, 1),
+                                        side_view=True)
+                    final_img = np.concatenate([final_img, side_img], axis=1)
+
+                cv2.imwrite(os.path.join(args.out_folder, f'{img_fn}_{animal_id}.png'), 
+                            cv2.cvtColor((255 * final_img).astype(np.uint8), cv2.COLOR_RGB2BGR))
+
+                # Add all verts and cams to list
+                verts = out['pred_vertices'][n].detach().cpu().numpy()
+                cam_t = pred_cam_t_full[n]
+
+                # Save all meshes to disk
+                if args.save_mesh:
+                    camera_translation = cam_t.copy()
+                    tmesh = renderer.vertices_to_trimesh(verts, camera_translation, LIGHT_BLUE)
+                    tmesh.export(os.path.join(args.out_folder, f'{img_fn}_{animal_id}.obj'))
+
+
+if __name__ == '__main__':
+    main()

demo_tta.py

@@ -0,0 +1,340 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+"""
+demo_tta.py: PRIMA inference with DeepLabCut SuperAnimal TTA
+
+Pipeline:
+1. Run Detectron2 to detect animals in the input image.
+2. Run PRIMA on each detected animal to obtain 3D pose/shape estimation.
+3. Run DeepLabCut SuperAnimal to obtain 2D keypoint estimation.
+4. Map the 39 SuperAnimal keypoints to the 26 PRIMA keypoints.
+5. Run test-time adaptation (TTA) with user-specified lr and num_iters
+   to further optimize the 3D pose and shape estimation.
+6. Render and save before/after TTA results (PNG + OBJ) and the
+   26-keypoint visualization (PNG).
+
+Reference code:
+- Test-time adaptation: prima/../eval_with_tta.py
+- DeepLabCut: https://github.com/AdaptiveMotorControlLab/FMPose3D/blob/main/animals/demo/vis_animals.py
+- Keypoint mapping (SuperAnimal 39 → PRIMA 26):
+    keypoint_mapping = {"quadruped80k":[10, 5, -1, 26, 29, 30, 35, 22, 24, 27, 31, 32, -1, -1,
+                                     25, 28, 33, 34, 15, 23, 11, 6, 4, 3, 0, -1]}
+"""
+
+
+from pathlib import Path
+import argparse
+import copy
+import os
+import tempfile
+import warnings
+
+import cv2
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.data
+from tqdm import tqdm
+
+from prima.models import load_prima
+from prima.utils import recursive_to
+from prima.datasets.vitdet_dataset import ViTDetDataset, DEFAULT_MEAN, DEFAULT_STD
+from prima.utils.renderer import Renderer, cam_crop_to_full
+
+warnings.filterwarnings("ignore")
+
+LIGHT_BLUE = (0.65098039, 0.74117647, 0.85882353)
+GREEN = (0.65, 0.86, 0.74)
+
+ANIMAL_COCO_IDS = [15, 16, 17, 18, 19, 21, 22]
+keypoint_mapping = {
+    "quadruped80k": [10, 5, -1, 26, 29, 30, 35, 22, 24, 27, 31, 32, -1, -1, 25, 28, 33, 34, 15, 23, 11, 6, 4, 3, 0, -1]
+}
+
+
+def denorm_patch_to_rgb(img_tensor: torch.Tensor) -> np.ndarray:
+    patch = (img_tensor.detach().cpu() * (DEFAULT_STD[:, None, None]) + DEFAULT_MEAN[:, None, None]) / 255.0
+    patch = patch.permute(1, 2, 0).numpy()
+    return np.clip(patch, 0.0, 1.0)
+
+
+def map_superanimal_to_prima(bodyparts_xyc: np.ndarray) -> np.ndarray:
+    mapping = keypoint_mapping["quadruped80k"]
+    num_src = bodyparts_xyc.shape[0]
+    mapped = np.zeros((len(mapping), 3), dtype=np.float32)
+
+    for tgt_i, src_i in enumerate(mapping):
+        if src_i >= 0 and src_i < num_src:
+            mapped[tgt_i] = bodyparts_xyc[src_i]
+
+    return mapped
+
+
+def save_keypoint_vis(patch_rgb: np.ndarray, kpts_xyc: np.ndarray, save_path: str) -> None:
+    vis = cv2.cvtColor((patch_rgb * 255).astype(np.uint8), cv2.COLOR_RGB2BGR).copy()
+    num_kpts = len(kpts_xyc)
+
+    for i, (x, y, c) in enumerate(kpts_xyc):
+        if c <= 0:
+            continue
+
+        # Use distinct color for each keypoint (OpenCV uses BGR)
+        hue = int(179 * i / max(1, num_kpts - 1))
+        color_bgr = cv2.cvtColor(np.uint8([[[hue, 255, 255]]]), cv2.COLOR_HSV2BGR)[0, 0]
+        color_bgr = (int(color_bgr[0]), int(color_bgr[1]), int(color_bgr[2]))
+
+        cx, cy = int(round(float(x))), int(round(float(y)))
+        cv2.circle(vis, (cx, cy), 3, color_bgr, -1)
+        cv2.putText(vis, str(i), (cx + 3, cy - 3), cv2.FONT_HERSHEY_SIMPLEX, 0.35, (255, 255, 255), 1, cv2.LINE_AA)
+
+    cv2.imwrite(save_path, vis)
+
+
+def run_superanimal_on_patch(patch_rgb: np.ndarray, args, tmp_dir: str):
+    try:
+        from deeplabcut.pose_estimation_pytorch.apis import superanimal_analyze_images
+    except Exception as e:
+        raise RuntimeError(
+            "Cannot import DeepLabCut SuperAnimal API. Please install deeplabcut with pose_estimation_pytorch support."
+        ) from e
+
+    patch_path = os.path.join(tmp_dir, "patch.png")
+    cv2.imwrite(patch_path, cv2.cvtColor((patch_rgb * 255).astype(np.uint8), cv2.COLOR_RGB2BGR))
+
+    preds = superanimal_analyze_images(
+        args.superanimal_name,
+        args.superanimal_model_name,
+        args.superanimal_detector_name,
+        patch_path,
+        args.superanimal_max_individuals,
+        out_folder=tmp_dir,
+    )
+
+    payload = preds.get(patch_path, None)
+    if payload is None:
+        return None
+    bodyparts = payload.get("bodyparts", None)
+    if bodyparts is None or len(bodyparts) == 0:
+        return None
+
+    best_idx = int(np.argmax(bodyparts[..., 2].mean(axis=1)))
+    return bodyparts[best_idx]
+
+
+def render_and_save(renderer, out, batch, img_fn, animal_id, out_folder, suffix, side_view, save_mesh):
+    pred_cam = out['pred_cam']
+    box_center = batch['box_center'].float()
+    box_size = batch['box_size'].float()
+    img_size = batch['img_size'].float()
+    scaled_focal_length = batch['focal_length'][0, 0] / batch['img'].shape[-1] * img_size.max()
+    pred_cam_t_full = cam_crop_to_full(pred_cam, box_center, box_size, img_size, scaled_focal_length)
+
+    white_img = (torch.ones_like(batch['img'][0]).cpu() - DEFAULT_MEAN[:, None, None] / 255) / (
+        DEFAULT_STD[:, None, None] / 255
+    )
+    input_patch = denorm_patch_to_rgb(batch['img'][0])
+
+    regression_img = renderer(
+        out['pred_vertices'][0].detach().cpu().numpy(),
+        out['pred_cam_t'][0].detach().cpu().numpy(),
+        batch['img'][0],
+        mesh_base_color=GREEN,
+        scene_bg_color=(1, 1, 1),
+    )
+
+    final_img = np.concatenate([input_patch, regression_img], axis=1)
+    if side_view:
+        side_img = renderer(
+            out['pred_vertices'][0].detach().cpu().numpy(),
+            out['pred_cam_t'][0].detach().cpu().numpy(),
+            white_img,
+            mesh_base_color=GREEN,
+            scene_bg_color=(1, 1, 1),
+            side_view=True,
+        )
+        final_img = np.concatenate([final_img, side_img], axis=1)
+
+    cv2.imwrite(
+        os.path.join(out_folder, f'{img_fn}_{animal_id}_{suffix}.png'),
+        cv2.cvtColor((255 * final_img).astype(np.uint8), cv2.COLOR_RGB2BGR),
+    )
+
+    if save_mesh:
+        verts = out['pred_vertices'][0].detach().cpu().numpy()
+        cam_t = pred_cam_t_full[0].detach().cpu().numpy()
+        tmesh = renderer.vertices_to_trimesh(verts, cam_t.copy(), LIGHT_BLUE)
+        tmesh.export(os.path.join(out_folder, f'{img_fn}_{animal_id}_{suffix}.obj'))
+
+
+def tta_optimize(model, batch, gt_kpts_norm, num_iters, lr):
+    model.eval()
+
+    if hasattr(model, 'backbone'):
+        for p in model.backbone.parameters():
+            p.requires_grad = False
+
+    orig_smal_head_state = copy.deepcopy(model.smal_head.state_dict())
+    model.smal_head.freeze_except_regression_heads()
+    tta_params = model.smal_head.get_tta_parameters(mode='all')
+    optimizer = torch.optim.Adam(tta_params, lr=lr)
+
+    valid_mask = (gt_kpts_norm[..., 2] > 0).float().unsqueeze(-1)
+    gt_xy = gt_kpts_norm[..., :2]
+
+    for _ in range(num_iters):
+        optimizer.zero_grad()
+        out = model(batch)
+        pred_xy = out['pred_keypoints_2d']
+        loss = F.mse_loss(pred_xy * valid_mask, gt_xy * valid_mask, reduction='sum') / (valid_mask.sum() + 1e-6)
+        loss.backward()
+        optimizer.step()
+
+    with torch.no_grad():
+        out_after = model(batch)
+
+    model.smal_head.load_state_dict(orig_smal_head_state)
+    model.smal_head.unfreeze_all()
+
+    return out_after
+
+
+def main():
+    parser = argparse.ArgumentParser(description='PRIMA + SuperAnimal + TTA demo')
+    parser.add_argument('--checkpoint', type=str, required=True, help='Path to pretrained PRIMA checkpoint')
+    parser.add_argument('--img_path', type=str, default=None, help='Single image path')
+    parser.add_argument('--img_folder', type=str, default='demo_data/', help='Folder with input images')
+    parser.add_argument('--out_folder', type=str, default='demo_out_tta', help='Output folder')
+    parser.add_argument('--side_view', dest='side_view', action='store_true', default=False, help='Render side view')
+    parser.add_argument('--save_mesh', dest='save_mesh', action='store_true', default=False, help='Save meshes')
+    parser.add_argument('--file_type', nargs='+', default=['*.jpg', '*.png', '*.jpeg', '*.JPEG'], help='Image globs')
+    parser.add_argument('--det_thresh', type=float, default=0.7, help='Detectron2 score threshold for animals')
+
+    parser.add_argument('--tta_lr', type=float, default=1e-6, help='TTA learning rate')
+    parser.add_argument('--tta_num_iters', type=int, default=30, help='TTA iterations')
+    parser.add_argument('--kp_conf_thresh', type=float, default=0.1, help='Keypoint confidence threshold')
+
+    parser.add_argument('--superanimal_name', type=str, default='superanimal_quadruped')
+    parser.add_argument('--superanimal_model_name', type=str, default='hrnet_w32')
+    parser.add_argument('--superanimal_detector_name', type=str, default='fasterrcnn_resnet50_fpn_v2')
+    parser.add_argument('--superanimal_max_individuals', type=int, default=1)
+
+    args = parser.parse_args()
+
+    model, model_cfg = load_prima(args.checkpoint)
+    device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+    model = model.to(device)
+    model.eval()
+
+    renderer = Renderer(model_cfg, faces=model.smal.faces)
+    os.makedirs(args.out_folder, exist_ok=True)
+
+    import detectron2.config
+    import detectron2.engine
+    from detectron2 import model_zoo
+
+    cfg = detectron2.config.get_cfg()
+    cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x.yaml"))
+    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
+    cfg.MODEL.WEIGHTS = "https://dl.fbaipublicfiles.com/detectron2/COCO-Detection/faster_rcnn_X_101_32x8d_FPN_3x/139173657/model_final_68b088.pkl"
+    detector = detectron2.engine.DefaultPredictor(cfg)
+
+    if args.img_path is not None:
+        img_paths = [Path(args.img_path)]
+    else:
+        img_paths = sorted([img for end in args.file_type for img in Path(args.img_folder).glob(end)])
+
+    for img_path in img_paths:
+        img_bgr = cv2.imread(str(img_path))
+        if img_bgr is None:
+            print(f"[WARN] Cannot read image: {img_path}")
+            continue
+        det_out = detector(img_bgr)
+        det_instances = det_out['instances']
+        valid_idx = [
+            i for i, (c, s) in enumerate(zip(det_instances.pred_classes, det_instances.scores))
+            if (int(c) in ANIMAL_COCO_IDS) and (float(s) > args.det_thresh)
+        ]
+        boxes = det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+        if len(boxes) == 0:
+            print(f"[INFO] No animal detected in {img_path}")
+            continue
+
+        dataset = ViTDetDataset(model_cfg, img_bgr, boxes)
+        dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=False, num_workers=0)
+
+        for batch in tqdm(dataloader, desc=f"{img_path.name}"):
+            batch = recursive_to(batch, device)
+            with torch.no_grad():
+                out_before = model(batch)
+
+            img_fn = img_path.stem
+            animal_id = int(batch['animalid'][0])
+
+            render_and_save(
+                renderer,
+                out_before,
+                batch,
+                img_fn,
+                animal_id,
+                args.out_folder,
+                suffix='before_tta',
+                side_view=args.side_view,
+                save_mesh=args.save_mesh,
+            )
+
+            patch_rgb = denorm_patch_to_rgb(batch['img'][0])
+            with tempfile.TemporaryDirectory(prefix=f"dlc_{img_fn}_{animal_id}_") as tmp_dir:
+                bodyparts_xyc = run_superanimal_on_patch(patch_rgb, args, tmp_dir)
+
+            if bodyparts_xyc is None:
+                print(f"[WARN] No SuperAnimal keypoints for {img_fn}_{animal_id}, skip TTA")
+                continue
+
+            mapped_xyc = map_superanimal_to_prima(bodyparts_xyc)
+            mapped_xyc[mapped_xyc[:, 2] < args.kp_conf_thresh, 2] = 0.0
+
+            save_keypoint_vis(
+                patch_rgb,
+                mapped_xyc,
+                os.path.join(args.out_folder, f"{img_fn}_{animal_id}_prima26_kpts.png"),
+            )
+            np.save(os.path.join(args.out_folder, f"{img_fn}_{animal_id}_prima26_kpts.npy"), mapped_xyc)
+
+            patch_h, patch_w = patch_rgb.shape[:2]
+            mapped_norm = mapped_xyc.copy()
+            mapped_norm[:, 0] = mapped_norm[:, 0] / float(patch_w) - 0.5
+            mapped_norm[:, 1] = mapped_norm[:, 1] / float(patch_h) - 0.5
+            gt_kpts_norm = torch.from_numpy(mapped_norm[None]).to(device=device, dtype=batch['img'].dtype)
+
+            out_after = tta_optimize(
+                model,
+                batch,
+                gt_kpts_norm,
+                num_iters=args.tta_num_iters,
+                lr=args.tta_lr,
+            )
+
+            render_and_save(
+                renderer,
+                out_after,
+                batch,
+                img_fn,
+                animal_id,
+                args.out_folder,
+                suffix='after_tta',
+                side_view=args.side_view,
+                save_mesh=args.save_mesh,
+            )
+
+
+if __name__ == '__main__':
+    main()
+

eval.py

@@ -0,0 +1,102 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+import numpy as np
+from tqdm import tqdm
+import torch
+from prima.utils import recursive_to
+from prima.utils.evaluate_metric import Evaluator
+from prima.datasets.datasets import EvaluationDataset
+import argparse
+from torch.utils.data import DataLoader
+from prima.models.prima import PRIMA
+from prima.configs import get_config
+torch.multiprocessing.set_sharing_strategy('file_system')
+
+
+def main(args):
+    cfg = get_config(args.config)
+    default_cfg = get_config(args.default_eval_config)
+    model = PRIMA.load_from_checkpoint(args.checkpoint, cfg=cfg, strict=False)
+    model.eval()
+
+    smal_evaluator = Evaluator(smal_model=model.smal, image_size=cfg.MODEL.IMAGE_SIZE)
+    cfg_eval_dataset = dict(default_cfg.DATASETS)
+    aug_cfg = cfg_eval_dataset.pop("CONFIG", None)  # augmentation config is not used in evaluation
+
+    if args.dataset.upper() == "ALL":
+        for key in cfg_eval_dataset.keys():
+            print(f"-------- Evaluate {key} dataset ------------")
+            eval_one_dataset(cfg_eval_dataset[key], default_cfg, cfg, model, 
+                             evaluator=smal_evaluator, 
+                             aug_cfg=aug_cfg, 
+                             key=key,
+                             device=args.device)
+            print(f"-------{key} Dataset evaluate finish ------")
+    else:
+        print(f"-------- Evaluate {args.dataset} dataset ------------")
+        eval_one_dataset(cfg_eval_dataset[args.dataset], default_cfg, cfg, model, 
+                         evaluator=smal_evaluator, 
+                         aug_cfg=aug_cfg, 
+                         key=args.dataset,
+                         device=args.device)
+        print(f"-------{args.dataset} Dataset evaluate finish ------")
+
+
+def eval_one_dataset(dataset_cfg, default_cfg, cfg, model, evaluator, aug_cfg, key, device='cuda'):
+    dataset = EvaluationDataset(root_image=dataset_cfg['ROOT_IMAGE'], 
+                                json_file=dataset_cfg['JSON_FILE']['TEST'], 
+                                augm_config=aug_cfg, focal_length=cfg.SMAL.get("FOCAL_LENGTH", 1000),
+                                image_size=cfg.MODEL.IMAGE_SIZE,
+                                )
+    dataloader = DataLoader(dataset, batch_size=1, num_workers=cfg.GENERAL.NUM_WORKERS)
+
+    bar = tqdm(dataloader)
+    pa_mpjpe_list, pck_list, auc_list, pa_mpvpe_list = [], [], [], []
+    for i, batch in enumerate(bar):
+        batch = recursive_to(batch, device)
+        with torch.no_grad():
+            output = model(batch)
+
+        if key in ["ANIMAL3D", "CONTROL_ANIMAL3D"]:
+            pa_mpjpe, pa_mpvpe = evaluator.eval_3d(output, batch)
+        else:
+            pa_mpjpe, pa_mpvpe = 0., 0.
+        pck, auc = evaluator.eval_2d(output, batch, pck_threshold=default_cfg.METRIC.PCK_THRESHOLD)
+
+        pa_mpjpe_list.append(pa_mpjpe)
+        pa_mpvpe_list.append(pa_mpvpe)
+        auc_list.append(auc)
+        pck_list.append(pck)
+
+        bar.set_postfix(PA_MPJPE=pa_mpjpe,
+                        PA_MPVPE=pa_mpvpe,
+                        AUC=auc,
+                        pck=pck,)
+
+    print("---------------- 3D metric -----------------")
+    print(f"Avg PA-MPJPE: {np.mean(pa_mpjpe_list)}")
+    print(f"Avg PA-MPVPE: {np.mean(pa_mpvpe_list)}")
+    
+    print("--------------- 2D metric ------------------")
+    print(f"AUC: {np.mean(auc_list)}")
+    pck_list = np.array(pck_list)
+    for _, th in enumerate(default_cfg.METRIC.PCK_THRESHOLD):
+        print(f"PCK@{th}: {np.mean(pck_list[:, _])}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--config", type=str, help="Path to config file", required=True)
+    parser.add_argument("--checkpoint", type=str, help="Path to checkpoint file", required=True)
+    parser.add_argument("--default_eval_config", type=str, default="./configs_hydra/experiment/default_val.yaml")
+    parser.add_argument("--dataset", type=str, default="ALL")
+    parser.add_argument("--device", type=str, default="cuda", help="Device to use for evaluation")
+    args = parser.parse_args()
+    main(args)

packages.txt

@@ -0,0 +1,4 @@
+libosmesa6
+libgl1
+libegl1
+libgles2

prima/__init__.py

@@ -0,0 +1,25 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+"""Top-level package for PRIMA.
+
+This package contains models, datasets and utilities for
+3D animal pose and shape estimation.
+"""
+
+from importlib.metadata import PackageNotFoundError, version
+
+
+try:  # pragma: no cover - best effort during development
+	__version__ = version("prima-animal")
+except PackageNotFoundError:  # pragma: no cover
+	__version__ = "0.0.0"
+
+
+__all__ = ["__version__"]

prima/configs/__init__.py

@@ -0,0 +1,99 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+from typing import Dict
+from yacs.config import CfgNode as CN
+
+def to_lower(x: Dict) -> Dict:
+    """
+    Convert all dictionary keys to lowercase
+    Args:
+      x (dict): Input dictionary
+    Returns:
+      dict: Output dictionary with all keys converted to lowercase
+    """
+    return {k.lower(): v for k, v in x.items()}
+
+
+_C = CN(new_allowed=True)
+
+_C.GENERAL = CN(new_allowed=True)
+_C.GENERAL.RESUME = True
+_C.GENERAL.TIME_TO_RUN = 3300
+_C.GENERAL.VAL_STEPS = 100
+_C.GENERAL.LOG_STEPS = 100
+_C.GENERAL.CHECKPOINT_STEPS = 20000
+_C.GENERAL.CHECKPOINT_DIR = "checkpoints"
+_C.GENERAL.SUMMARY_DIR = "tensorboard"
+_C.GENERAL.NUM_GPUS = 1
+_C.GENERAL.NUM_WORKERS = 4
+_C.GENERAL.MIXED_PRECISION = True
+_C.GENERAL.ALLOW_CUDA = True
+_C.GENERAL.PIN_MEMORY = False
+_C.GENERAL.DISTRIBUTED = False
+_C.GENERAL.LOCAL_RANK = 0
+_C.GENERAL.USE_SYNCBN = False
+_C.GENERAL.WORLD_SIZE = 1
+_C.GENERAL.PREFETCH_FACTOR = 2
+
+_C.TRAIN = CN(new_allowed=True)
+_C.TRAIN.NUM_EPOCHS = 100
+_C.TRAIN.SHUFFLE = True
+_C.TRAIN.WARMUP = False
+_C.TRAIN.NORMALIZE_PER_IMAGE = False
+_C.TRAIN.CLIP_GRAD = False
+_C.TRAIN.CLIP_GRAD_VALUE = 1.0
+_C.LOSS_WEIGHTS = CN(new_allowed=True)
+
+_C.DATASETS = CN(new_allowed=True)
+
+_C.MODEL = CN(new_allowed=True)
+_C.MODEL.IMAGE_SIZE = 224
+
+_C.EXTRA = CN(new_allowed=True)
+_C.EXTRA.FOCAL_LENGTH = 5000
+
+_C.DATASETS.CONFIG = CN(new_allowed=True)
+_C.DATASETS.CONFIG.SCALE_FACTOR = 0.3
+_C.DATASETS.CONFIG.ROT_FACTOR = 30
+_C.DATASETS.CONFIG.TRANS_FACTOR = 0.02
+_C.DATASETS.CONFIG.COLOR_SCALE = 0.2
+_C.DATASETS.CONFIG.ROT_AUG_RATE = 0.6
+_C.DATASETS.CONFIG.TRANS_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.DO_FLIP = False
+_C.DATASETS.CONFIG.FLIP_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.EXTREME_CROP_AUG_RATE = 0.10
+
+
+def default_config() -> CN:
+    """
+    Get a yacs CfgNode object with the default config values.
+    """
+    # Return a clone so that the defaults will not be altered
+    # This is for the "local variable" use pattern
+    return _C.clone()
+
+
+def get_config(config_file: str, merge: bool = True) -> CN:
+    """
+    Read a config file and optionally merge it with the default config file.
+    Args:
+      config_file (str): Path to config file.
+      merge (bool): Whether to merge with the default config or not.
+    Returns:
+      CfgNode: Config as a yacs CfgNode object.
+    """
+    if merge:
+        cfg = default_config()
+    else:
+        cfg = CN(new_allowed=True)
+    cfg.merge_from_file(config_file)
+
+    cfg.freeze()
+    return cfg

prima/models/__init__.py

@@ -0,0 +1,54 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+from .prima import PRIMA
+
+
+def load_prima(checkpoint_path):
+    from pathlib import Path
+    from ..configs import get_config
+    model_cfg = str(Path(checkpoint_path).parent.parent / '.hydra/config.yaml')
+    model_cfg = get_config(model_cfg)
+
+    # Override some config values, to crop bbox correctly
+    if (model_cfg.MODEL.BACKBONE.TYPE == 'vit') and ('BBOX_SHAPE' not in model_cfg.MODEL):
+        model_cfg.defrost()
+        assert model_cfg.MODEL.IMAGE_SIZE == 256, f"MODEL.IMAGE_SIZE ({model_cfg.MODEL.IMAGE_SIZE}) should be 256 for ViT backbone"
+        model_cfg.MODEL.BBOX_SHAPE = [192, 256]
+        model_cfg.freeze()
+    if (model_cfg.MODEL.BACKBONE.TYPE == 'dinov3') and ('BBOX_SHAPE' not in model_cfg.MODEL):
+        model_cfg.defrost()
+        assert model_cfg.MODEL.IMAGE_SIZE == 256, f"MODEL.IMAGE_SIZE ({model_cfg.MODEL.IMAGE_SIZE}) should be 256 for dino backbone"
+        model_cfg.MODEL.BBOX_SHAPE = [256, 256]
+        model_cfg.freeze()
+
+    if (model_cfg.MODEL.BACKBONE.TYPE == 'dinov2') and ('BBOX_SHAPE' not in model_cfg.MODEL):
+        model_cfg.defrost()
+        assert model_cfg.MODEL.IMAGE_SIZE == 252, f"MODEL.IMAGE_SIZE ({model_cfg.MODEL.IMAGE_SIZE}) should be 252 for dino backbone"
+        model_cfg.MODEL.BBOX_SHAPE = [252, 252]
+        model_cfg.freeze()
+        
+ 
+
+    # Update config to be compatible with demo
+    if ('PRETRAINED_WEIGHTS' in model_cfg.MODEL.BACKBONE):
+        model_cfg.defrost()
+        model_cfg.MODEL.BACKBONE.pop('PRETRAINED_WEIGHTS')
+        model_cfg.freeze()
+
+    # Offscreen training renderer is not needed for demo/inference startup and
+    # can fail on some local OpenGL backends.
+    model = PRIMA.load_from_checkpoint(
+        checkpoint_path,
+        strict=False,
+        cfg=model_cfg,
+        map_location='cpu',
+        init_renderer=False,
+    )
+    return model, model_cfg

prima/models/backbones/__init__.py

@@ -0,0 +1,19 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+from .vit import vith
+
+
+
+
+def create_backbone(cfg):
+    if cfg.MODEL.BACKBONE.TYPE in ['vith','concat','aa']:   # vit bb will be used in these three cases - animal feature extractor 
+        return vith(cfg)
+    else:
+        raise NotImplementedError('Backbone type is not implemented')

prima/models/backbones/vit.py

@@ -0,0 +1,375 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+# Copyright (c) OpenMMLab. All rights reserved.
+import math
+
+import torch
+from functools import partial
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.layers import drop_path, to_2tuple, trunc_normal_
+
+
+def vith(cfg):
+    return ViT(
+        img_size=(256, 192),
+        patch_size=16,
+        embed_dim=1280,
+        depth=32,
+        num_heads=16,
+        ratio=1,
+        use_checkpoint=False,
+        # use_checkpoint=True,
+        mlp_ratio=4,
+        qkv_bias=True,
+        drop_path_rate=0.55,
+        use_cls=True, # cls for animal family classification 
+    )
+
+
+def get_abs_pos(abs_pos, h, w, ori_h, ori_w, has_cls_token=True):
+    """
+    Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token
+        dimension for the original embeddings.
+    Args:
+        abs_pos (Tensor): absolute positional embeddings with (1, num_position, C).
+        has_cls_token (bool): If true, has 1 embedding in abs_pos for cls token.
+        hw (Tuple): size of input image tokens.
+
+    Returns:
+        Absolute positional embeddings after processing with shape (1, H, W, C)
+    """
+    cls_token = None
+    B, L, C = abs_pos.shape
+    if has_cls_token:
+        cls_token = abs_pos[:, 0:1]
+        abs_pos = abs_pos[:, 1:]
+
+    if ori_h != h or ori_w != w:
+        new_abs_pos = F.interpolate(
+            abs_pos.reshape(1, ori_h, ori_w, -1).permute(0, 3, 1, 2),
+            size=(h, w),
+            mode="bicubic",
+            align_corners=False,
+        ).permute(0, 2, 3, 1).reshape(B, -1, C)
+
+    else:
+        new_abs_pos = abs_pos
+
+    if cls_token is not None:
+        new_abs_pos = torch.cat([cls_token, new_abs_pos], dim=1)
+    return new_abs_pos
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+    def extra_repr(self):
+        return 'p={}'.format(self.drop_prob)
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., attn_head_dim=None):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.dim = dim
+
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=qkv_bias)
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x)
+        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None,
+                 drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU,
+                 norm_layer=nn.LayerNorm, attn_head_dim=None, 
+                 ):
+        super().__init__()
+
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, attn_head_dim=attn_head_dim
+        )
+
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.drop_path(self.attn(self.norm1(x)))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, ratio=1):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) * (ratio ** 2)
+        self.patch_shape = (int(img_size[0] // patch_size[0] * ratio), int(img_size[1] // patch_size[1] * ratio))
+        self.origin_patch_shape = (int(img_size[0] // patch_size[0]), int(img_size[1] // patch_size[1]))
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=(patch_size[0] // ratio),
+                              padding=4 + 2 * (ratio // 2 - 1))
+
+    def forward(self, x, **kwargs):
+        B, C, H, W = x.shape
+        x = self.proj(x)
+        Hp, Wp = x.shape[2], x.shape[3]
+
+        x = x.flatten(2).transpose(1, 2)
+        return x, (Hp, Wp)
+
+
+class HybridEmbed(nn.Module):
+    """ CNN Feature Map Embedding
+    Extract feature map from CNN, flatten, project to embedding dim.
+    """
+
+    def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768):
+        super().__init__()
+        assert isinstance(backbone, nn.Module)
+        img_size = to_2tuple(img_size)
+        self.img_size = img_size
+        self.backbone = backbone
+        if feature_size is None:
+            with torch.no_grad():
+                training = backbone.training
+                if training:
+                    backbone.eval()
+                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))[-1]
+                feature_size = o.shape[-2:]
+                feature_dim = o.shape[1]
+                backbone.train(training)
+        else:
+            feature_size = to_2tuple(feature_size)
+            feature_dim = self.backbone.feature_info.channels()[-1]
+        self.num_patches = feature_size[0] * feature_size[1]
+        self.proj = nn.Linear(feature_dim, embed_dim)
+
+    def forward(self, x):
+        x = self.backbone(x)[-1]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.proj(x)
+        return x
+
+
+class ViT(nn.Module):
+
+    def __init__(self,
+                 img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., hybrid_backbone=None, norm_layer=None, use_checkpoint=False,
+                 frozen_stages=-1, ratio=1, last_norm=True, use_cls=False,
+                 patch_padding='pad', freeze_attn=False, freeze_ffn=False,
+                 ):
+        # Protect mutable default arguments
+        super(ViT, self).__init__()
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.frozen_stages = frozen_stages
+        self.use_checkpoint = use_checkpoint
+        self.patch_padding = patch_padding
+        self.freeze_attn = freeze_attn
+        self.freeze_ffn = freeze_ffn
+        self.depth = depth
+
+        if hybrid_backbone is not None:
+            self.patch_embed = HybridEmbed(
+                hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim)
+        else:
+            self.patch_embed = PatchEmbed(
+                img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, ratio=ratio)
+        num_patches = self.patch_embed.num_patches
+
+        # since the pretraining model has class token
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+            )
+            for i in range(depth)])
+
+        self.last_norm = norm_layer(embed_dim) if last_norm else nn.Identity()
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+
+        self.use_cls = use_cls
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        nn.init.normal_(self.cls_token, std=1e-6)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        """Freeze parameters."""
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        for i in range(1, self.frozen_stages + 1):
+            m = self.blocks[i]
+            m.eval()
+            for param in m.parameters():
+                param.requires_grad = False
+
+        if self.freeze_attn:
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.attn.eval()
+                m.norm1.eval()
+                for param in m.attn.parameters():
+                    param.requires_grad = False
+                for param in m.norm1.parameters():
+                    param.requires_grad = False
+
+        if self.freeze_ffn:
+            self.pos_embed.requires_grad = False
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.mlp.eval()
+                m.norm2.eval()
+                for param in m.mlp.parameters():
+                    param.requires_grad = False
+                for param in m.norm2.parameters():
+                    param.requires_grad = False
+
+    def init_weights(self):
+        """Initialize the weights in backbone.
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+
+        def _init_weights(m):
+            if isinstance(m, nn.Linear):
+                trunc_normal_(m.weight, std=.02)
+                if isinstance(m, nn.Linear) and m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.LayerNorm):
+                nn.init.constant_(m.bias, 0)
+                nn.init.constant_(m.weight, 1.0)
+
+        self.apply(_init_weights)
+
+    def get_num_layers(self):
+        return len(self.blocks)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def forward_features(self, x):
+        B, C, H, W = x.shape
+        x, (Hp, Wp) = self.patch_embed(x)
+
+        if self.pos_embed is not None:
+            # fit for multiple GPU training
+            # since the first element for pos embed (sin-cos manner) is zero, it will cause no difference
+            x = x + self.pos_embed[:, 1:] + self.pos_embed[:, :1]
+
+        x = torch.cat((self.cls_token.expand(B, -1, -1), x), dim=1) if self.use_cls else x
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+
+        x = self.last_norm(x)
+
+        cls = x[:, 0] if self.use_cls else None
+        x = x[:, 1:] if self.use_cls else x
+        xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp).contiguous()
+
+        return xp, cls # shape [B, D, Hp, Wp], [B, D]
+
+    def forward(self, x):
+        x, cls = self.forward_features(x)
+        return x, cls
+
+    def train(self, mode=True):
+        """Convert the model into training mode."""
+        super().train(mode)
+        self._freeze_stages()

prima/models/bioclip_embedding.py

@@ -0,0 +1,70 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+"""
+bioclip Embedding Module
+Converts image batch to embeddings that can be concatenated with image features
+"""
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class BioClipEmbedding(nn.Module):
+    """
+    Embeds images into a feature space using BioClip model that can be combined with image features.
+    
+    Args:
+        embed_dim: Output embedding dimension, should match the dimension of image features for concatenation
+    """
+    
+    def __init__(self, cfg, embed_dim: int = 1024):
+        super().__init__()
+        
+        self.embed_dim = embed_dim
+        
+        import open_clip
+        
+        if cfg.MODEL.BIOCLIP_EMBEDDING.TYPE == 'bioclip2':
+            print("[BioClipEmbedding] Using BioClip2 model from Hugging Face Hub")
+            self.species_model, _,_ = open_clip.create_model_and_transforms('hf-hub:imageomics/bioclip-2')
+        else:   
+            self.species_model, _,_ = open_clip.create_model_and_transforms('hf-hub:imageomics/bioclip')
+        # tokenizer = open_clip.get_tokenizer('hf-hub:imageomics/bioclip')
+
+
+        self.species_model.eval()
+        
+        # Get the output dimension from the model
+        bioclip_output_dim = self.species_model.visual.output_dim
+        
+        # Project to target dimension
+        self.projection = nn.Sequential(
+            nn.Linear(bioclip_output_dim, embed_dim),
+            nn.LayerNorm(embed_dim),
+        )
+        
+    def forward(self, images: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            images: Tensor of shape (B, C, H, W) representing a batch of images
+        Returns:
+            Tensor of shape (B, embed_dim) representing the embedded features
+        """
+        # BioClip expects 224x224 input, resize if needed
+        if images.shape[-2:] != (224, 224):
+            images_resized = F.interpolate(images, size=(224, 224), mode='bilinear', align_corners=False)
+        else:
+            images_resized = images
+            
+        with torch.no_grad():
+            image_features = self.species_model.encode_image(images_resized)
+        
+        projected_features = self.projection(image_features)
+        
+        return projected_features

prima/models/components/model_utils.py

@@ -0,0 +1,160 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+# 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 copy
+from typing import Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def select_closest_cond_frames(frame_idx, cond_frame_outputs, max_cond_frame_num):
+    """
+    Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
+    that are temporally closest to the current frame at `frame_idx`. Here, we take
+    - a) the closest conditioning frame before `frame_idx` (if any);
+    - b) the closest conditioning frame after `frame_idx` (if any);
+    - c) any other temporally closest conditioning frames until reaching a total
+         of `max_cond_frame_num` conditioning frames.
+
+    Outputs:
+    - selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
+    - unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
+    """
+    if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
+        selected_outputs = cond_frame_outputs
+        unselected_outputs = {}
+    else:
+        assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
+        selected_outputs = {}
+
+        # the closest conditioning frame before `frame_idx` (if any)
+        idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
+        if idx_before is not None:
+            selected_outputs[idx_before] = cond_frame_outputs[idx_before]
+
+        # the closest conditioning frame after `frame_idx` (if any)
+        idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
+        if idx_after is not None:
+            selected_outputs[idx_after] = cond_frame_outputs[idx_after]
+
+        # add other temporally closest conditioning frames until reaching a total
+        # of `max_cond_frame_num` conditioning frames.
+        num_remain = max_cond_frame_num - len(selected_outputs)
+        inds_remain = sorted(
+            (t for t in cond_frame_outputs if t not in selected_outputs),
+            key=lambda x: abs(x - frame_idx),
+        )[:num_remain]
+        selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
+        unselected_outputs = {
+            t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
+        }
+
+    return selected_outputs, unselected_outputs
+
+
+def get_1d_sine_pe(pos_inds, dim, temperature=10000):
+    """
+    Get 1D sine positional embedding as in the original Transformer paper.
+    """
+    pe_dim = dim // 2
+    dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
+    dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)
+
+    pos_embed = pos_inds.unsqueeze(-1) / dim_t
+    pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
+    return pos_embed
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+class DropPath(nn.Module):
+    # adapted from https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/drop.py
+    def __init__(self, drop_prob=0.0, scale_by_keep=True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        if self.drop_prob == 0.0 or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+
+# 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,
+        activation: nn.Module = nn.ReLU,
+        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
+        self.act = activation()
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = self.act(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return 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

prima/models/components/pose_transformer.py

@@ -0,0 +1,366 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+from inspect import isfunction
+from typing import Callable, Optional
+
+import torch
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from torch import nn
+
+from .t_cond_mlp import (
+    AdaptiveLayerNorm1D,
+    FrequencyEmbedder,
+    normalization_layer,
+)
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+class PreNorm(nn.Module):
+    def __init__(self, dim: int, fn: Callable, norm: str = "layer", norm_cond_dim: int = -1):
+        super().__init__()
+        self.norm = normalization_layer(norm, dim, norm_cond_dim)
+        self.fn = fn
+
+    def forward(self, x: torch.Tensor, *args, **kwargs):
+        if isinstance(self.norm, AdaptiveLayerNorm1D):
+            return self.fn(self.norm(x, *args), **kwargs)
+        else:
+            return self.fn(self.norm(x), **kwargs)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout=0.0):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x):
+        qkv = self.to_qkv(x).chunk(3, dim=-1)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), qkv)
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        context_dim = default(context_dim, dim)
+        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias=False)
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x, context=None):
+        context = default(context, x)
+        k, v = self.to_kv(context).chunk(2, dim=-1)
+        q = self.to_q(x)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), [q, k, v])
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args):
+        for attn, ff in self.layers:
+            x = attn(x, *args) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class TransformerCrossAttn(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ca = CrossAttention(
+                dim, context_dim=context_dim, heads=heads, dim_head=dim_head, dropout=dropout
+            )
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ca, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args, context=None, context_list=None):
+        if context_list is None:
+            context_list = [context] * len(self.layers)
+        if len(context_list) != len(self.layers):
+            raise ValueError(f"len(context_list) != len(self.layers) ({len(context_list)} != {len(self.layers)})")
+
+        for i, (self_attn, cross_attn, ff) in enumerate(self.layers):
+            x = self_attn(x, *args) + x
+            x = cross_attn(x, *args, context=context_list[i]) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class DropTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()
+
+            if zero_mask.any():
+                x = x[:, ~zero_mask, :]
+        return x
+
+
+class ZeroTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()
+            # Zero-out the masked tokens
+            x[zero_mask, :] = 0
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = "drop",
+        emb_dropout_loc: str = "token",
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        token_pe_numfreq: int = -1,
+    ):
+        super().__init__()
+        if token_pe_numfreq > 0:
+            token_dim_new = token_dim * (2 * token_pe_numfreq + 1)
+            self.to_token_embedding = nn.Sequential(
+                Rearrange("b n d -> (b n) d", n=num_tokens, d=token_dim),
+                FrequencyEmbedder(token_pe_numfreq, token_pe_numfreq - 1),
+                Rearrange("(b n) d -> b n d", n=num_tokens, d=token_dim_new),
+                nn.Linear(token_dim_new, dim),
+            )
+        else:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        else:
+            raise ValueError(f"Unknown emb_dropout_type: {emb_dropout_type}")
+        self.emb_dropout_loc = emb_dropout_loc
+
+        self.transformer = Transformer(
+            dim, depth, heads, dim_head, mlp_dim, dropout, norm=norm, norm_cond_dim=norm_cond_dim
+        )
+
+    def forward(self, inp: torch.Tensor, *args, **kwargs):
+        x = inp
+
+        if self.emb_dropout_loc == "input":
+            x = self.dropout(x)
+        x = self.to_token_embedding(x)
+
+        if self.emb_dropout_loc == "token":
+            x = self.dropout(x)
+        b, n, _ = x.shape
+        x += self.pos_embedding[:, :n]
+
+        if self.emb_dropout_loc == "token_afterpos":
+            x = self.dropout(x)
+        x = self.transformer(x, *args)
+        return x
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = 'drop',
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+        skip_token_embedding: bool = False,
+    ):
+        super().__init__()
+        if not skip_token_embedding:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        else:
+            self.to_token_embedding = nn.Identity()
+            if token_dim != dim:
+                raise ValueError(
+                    f"token_dim ({token_dim}) != dim ({dim}) when skip_token_embedding is True"
+                )
+
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        elif emb_dropout_type == "normal":
+            self.dropout = nn.Dropout(emb_dropout)
+
+        self.transformer = TransformerCrossAttn(
+            dim,
+            depth,
+            heads,
+            dim_head,
+            mlp_dim,
+            dropout,
+            norm=norm,
+            norm_cond_dim=norm_cond_dim,
+            context_dim=context_dim,
+        )
+
+    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None):
+        x = self.to_token_embedding(inp)
+        b, n, _ = x.shape
+
+        x = self.dropout(x)
+        x += self.pos_embedding[:, :n]
+
+        x = self.transformer(x, *args, context=context, context_list=context_list)
+        return x
+

prima/models/components/position_encoding.py

@@ -0,0 +1,84 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Any, Optional, Tuple
+
+import numpy as np
+
+import torch
+from torch import nn
+
+# Rotary Positional Encoding, adapted from:
+# 1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
+# 2. https://github.com/naver-ai/rope-vit
+# 3. https://github.com/lucidrains/rotary-embedding-torch
+
+
+def init_t_xy(end_x: int, end_y: int):
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    t_x = (t % end_x).float()
+    t_y = torch.div(t, end_x, rounding_mode="floor").float()
+    return t_x, t_y
+
+
+def compute_axial_cis(dim: int, end_x: int, end_y: int, theta: float = 10000.0):
+    freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+    freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+
+    t_x, t_y = init_t_xy(end_x, end_y)
+    freqs_x = torch.outer(t_x, freqs_x)
+    freqs_y = torch.outer(t_y, freqs_y)
+    freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
+    freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
+    return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
+
+
+def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+    assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
+    shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis.view(*shape)
+
+
+def apply_rotary_enc(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    repeat_freqs_k: bool = False,
+):
+    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_ = (
+        torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+        if xk.shape[-2] != 0
+        else None
+    )
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
+    if xk_ is None:
+        # no keys to rotate, due to dropout
+        return xq_out.type_as(xq).to(xq.device), xk
+    # repeat freqs along seq_len dim to match k seq_len
+    if repeat_freqs_k:
+        r = xk_.shape[-2] // xq_.shape[-2]
+        if freqs_cis.is_cuda:
+            freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
+        else:
+            # torch.repeat on complex numbers may not be supported on non-CUDA devices
+            # (freqs_cis has 4 dims and we repeat on dim 2) so we use expand + flatten
+            freqs_cis = freqs_cis.unsqueeze(2).expand(-1, -1, r, -1, -1).flatten(2, 3)
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+    return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)

prima/models/components/t_cond_mlp.py

@@ -0,0 +1,204 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+import copy
+from typing import List, Optional
+
+import torch
+
+
+class AdaptiveLayerNorm1D(torch.nn.Module):
+    def __init__(self, data_dim: int, norm_cond_dim: int):
+        super().__init__()
+        if data_dim <= 0:
+            raise ValueError(f"data_dim must be positive, but got {data_dim}")
+        if norm_cond_dim <= 0:
+            raise ValueError(f"norm_cond_dim must be positive, but got {norm_cond_dim}")
+        self.norm = torch.nn.LayerNorm(data_dim)
+        self.linear = torch.nn.Linear(norm_cond_dim, 2 * data_dim)
+        torch.nn.init.zeros_(self.linear.weight)
+        torch.nn.init.zeros_(self.linear.bias)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        # x: (batch, ..., data_dim)
+        # t: (batch, norm_cond_dim)
+        # return: (batch, data_dim)
+        x = self.norm(x)
+        alpha, beta = self.linear(t).chunk(2, dim=-1)
+
+        # Add singleton dimensions to alpha and beta
+        if x.dim() > 2:
+            alpha = alpha.view(alpha.shape[0], *([1] * (x.dim() - 2)), alpha.shape[1])
+            beta = beta.view(beta.shape[0], *([1] * (x.dim() - 2)), beta.shape[1])
+
+        return x * (1 + alpha) + beta
+
+
+class SequentialCond(torch.nn.Sequential):
+    def forward(self, input, *args, **kwargs):
+        for module in self:
+            if isinstance(module, (AdaptiveLayerNorm1D, SequentialCond, ResidualMLPBlock)):
+                input = module(input, *args, **kwargs)
+            else:
+                input = module(input)
+        return input
+
+
+def normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):
+    if norm == "batch":
+        return torch.nn.BatchNorm1d(dim)
+    elif norm == "layer":
+        return torch.nn.LayerNorm(dim)
+    elif norm == "ada":
+        assert norm_cond_dim > 0, f"norm_cond_dim must be positive, got {norm_cond_dim}"
+        return AdaptiveLayerNorm1D(dim, norm_cond_dim)
+    elif norm is None:
+        return torch.nn.Identity()
+    else:
+        raise ValueError(f"Unknown norm: {norm}")
+
+
+def linear_norm_activ_dropout(
+    input_dim: int,
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    layers.append(torch.nn.Linear(input_dim, output_dim, bias=bias))
+    if norm is not None:
+        layers.append(normalization_layer(norm, output_dim, norm_cond_dim))
+    layers.append(copy.deepcopy(activation))
+    if dropout > 0.0:
+        layers.append(torch.nn.Dropout(dropout))
+    return SequentialCond(*layers)
+
+
+def create_simple_mlp(
+    input_dim: int,
+    hidden_dims: List[int],
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    prev_dim = input_dim
+    for hidden_dim in hidden_dims:
+        layers.extend(
+            linear_norm_activ_dropout(
+                prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            )
+        )
+        prev_dim = hidden_dim
+    layers.append(torch.nn.Linear(prev_dim, output_dim, bias=bias))
+    return SequentialCond(*layers)
+
+
+class ResidualMLPBlock(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        if not (input_dim == output_dim == hidden_dim):
+            raise NotImplementedError(
+                f"input_dim {input_dim} != output_dim {output_dim} is not implemented"
+            )
+
+        layers = []
+        prev_dim = input_dim
+        for i in range(num_hidden_layers):
+            layers.append(
+                linear_norm_activ_dropout(
+                    prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+                )
+            )
+            prev_dim = hidden_dim
+        self.model = SequentialCond(*layers)
+        self.skip = torch.nn.Identity()
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return x + self.model(x, *args, **kwargs)
+
+
+class ResidualMLP(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        num_blocks: int = 1,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.input_dim = input_dim
+        self.model = SequentialCond(
+            linear_norm_activ_dropout(
+                input_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            ),
+            *[
+                ResidualMLPBlock(
+                    hidden_dim,
+                    hidden_dim,
+                    num_hidden_layers,
+                    hidden_dim,
+                    activation,
+                    bias,
+                    norm,
+                    dropout,
+                    norm_cond_dim,
+                )
+                for _ in range(num_blocks)
+            ],
+            torch.nn.Linear(hidden_dim, output_dim, bias=bias),
+        )
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return self.model(x, *args, **kwargs)
+
+
+class FrequencyEmbedder(torch.nn.Module):
+    def __init__(self, num_frequencies, max_freq_log2):
+        super().__init__()
+        frequencies = 2 ** torch.linspace(0, max_freq_log2, steps=num_frequencies)
+        self.register_buffer("frequencies", frequencies)
+
+    def forward(self, x):
+        # x should be of size (N,) or (N, D)
+        N = x.size(0)
+        if x.dim() == 1:  # (N,)
+            x = x.unsqueeze(1)  # (N, D) where D=1
+        x_unsqueezed = x.unsqueeze(-1)  # (N, D, 1)
+        scaled = self.frequencies.view(1, 1, -1) * x_unsqueezed  # (N, D, num_frequencies)
+        s = torch.sin(scaled)
+        c = torch.cos(scaled)
+        embedded = torch.cat([s, c, x_unsqueezed], dim=-1).view(
+            N, -1
+        )  # (N, D * 2 * num_frequencies + D)
+        return embedded
+

prima/models/components/transformer.py

@@ -0,0 +1,400 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+# 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 contextlib
+import math
+import warnings
+from functools import partial
+from typing import Tuple, Type
+
+import torch
+import torch.nn.functional as F
+from torch import nn, Tensor
+
+from .position_encoding import apply_rotary_enc, compute_axial_cis
+from .model_utils import MLP
+
+warnings.simplefilter(action="ignore", category=FutureWarning)
+
+
+def get_sdpa_settings():
+    if torch.cuda.is_available():
+        old_gpu = torch.cuda.get_device_properties(0).major < 7
+        # only use Flash Attention on Ampere (8.0) or newer GPUs
+        use_flash_attn = torch.cuda.get_device_properties(0).major >= 8
+        if not use_flash_attn:
+            warnings.warn(
+                "Flash Attention is disabled as it requires a GPU with Ampere (8.0) CUDA capability.",
+                category=UserWarning,
+                stacklevel=2,
+            )
+        # keep math kernel for PyTorch versions before 2.2 (Flash Attention v2 is only
+        # available on PyTorch 2.2+, while Flash Attention v1 cannot handle all cases)
+        pytorch_version = tuple(int(v) for v in torch.__version__.split(".")[:2])
+        if pytorch_version < (2, 2):
+            warnings.warn(
+                f"You are using PyTorch {torch.__version__} without Flash Attention v2 support. "
+                "Consider upgrading to PyTorch 2.2+ for Flash Attention v2 (which could be faster).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+        math_kernel_on = pytorch_version < (2, 2) or not use_flash_attn
+    else:
+        old_gpu = True
+        use_flash_attn = False
+        math_kernel_on = True
+
+    return old_gpu, use_flash_attn, math_kernel_on
+
+
+# Check whether Flash Attention is available (and use it by default)
+OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = get_sdpa_settings()
+# A fallback setting to allow all available kernels if Flash Attention fails
+ALLOW_ALL_KERNELS = False
+
+
+def sdp_kernel_context(dropout_p):
+    """
+    Get the context for the attention scaled dot-product kernel. We use Flash Attention
+    by default, but fall back to all available kernels if Flash Attention fails.
+    """
+    if ALLOW_ALL_KERNELS:
+        return contextlib.nullcontext()
+
+    return torch.backends.cuda.sdp_kernel(
+        enable_flash=USE_FLASH_ATTN,
+        # if Flash attention kernel is off, then math kernel needs to be enabled
+        enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
+        enable_mem_efficient=OLD_GPU,
+    )
+
+
+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 = MLP(
+            embedding_dim, mlp_dim, embedding_dim, num_layers=2, activation=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,
+        dropout: float = 0.0,
+        kv_in_dim: int = None,
+    ) -> None:
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.kv_in_dim = kv_in_dim if kv_in_dim is not None else 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(self.kv_in_dim, self.internal_dim)
+        self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+        self.dropout_p = dropout
+
+    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).contiguous()  # 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).contiguous()
+        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)
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        try:
+            with sdp_kernel_context(dropout_p):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        except Exception as e:
+            # Fall back to all kernels if the Flash attention kernel fails
+            warnings.warn(
+                f"Flash Attention kernel failed due to: {e}\nFalling back to all available "
+                f"kernels for scaled_dot_product_attention (which may have a slower speed).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+            global ALLOW_ALL_KERNELS
+            ALLOW_ALL_KERNELS = True
+            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out
+
+
+class RoPEAttention(Attention):
+    """Attention with rotary position encoding."""
+
+    def __init__(
+        self,
+        *args,
+        rope_theta=10000.0,
+        # whether to repeat q rope to match k length
+        # this is needed for cross-attention to memories
+        rope_k_repeat=False,
+        feat_sizes=(32, 32),  # [w, h] for stride 16 feats at 512 resolution
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+
+        self.compute_cis = partial(
+            compute_axial_cis, dim=self.internal_dim // self.num_heads, theta=rope_theta
+        )
+        freqs_cis = self.compute_cis(end_x=feat_sizes[0], end_y=feat_sizes[1])
+        self.freqs_cis = freqs_cis
+        self.rope_k_repeat = rope_k_repeat
+
+    def forward(
+        self, q: Tensor, k: Tensor, v: Tensor, num_k_exclude_rope: int=0,
+    ) -> 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)
+
+        # Apply rotary position encoding
+        w = h = math.sqrt(q.shape[-2])
+        self.freqs_cis = self.freqs_cis.to(q.device)
+        if self.freqs_cis.shape[0] != q.shape[-2]:
+            self.freqs_cis = self.compute_cis(end_x=w, end_y=h).to(q.device)
+        if q.shape[-2] != k.shape[-2]:
+            assert self.rope_k_repeat
+
+        num_k_rope = k.size(-2) - num_k_exclude_rope
+        q, k[:, :, :num_k_rope] = apply_rotary_enc(
+            q,
+            k[:, :, :num_k_rope],
+            freqs_cis=self.freqs_cis,
+            repeat_freqs_k=self.rope_k_repeat,
+        )
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        try:
+            with sdp_kernel_context(dropout_p):
+                out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+        except Exception as e:
+            # Fall back to all kernels if the Flash attention kernel fails
+            warnings.warn(
+                f"Flash Attention kernel failed due to: {e}\nFalling back to all available "
+                f"kernels for scaled_dot_product_attention (which may have a slower speed).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+            global ALLOW_ALL_KERNELS
+            ALLOW_ALL_KERNELS = True
+            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out

prima/models/discriminator.py

@@ -0,0 +1,129 @@
+"""
+PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation
+
+Official implementation of the paper:
+"PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation"
+by Xiaohang Yu, Ti Wang, and Mackenzie Weygandt Mathis
+Licensed under a modified MIT license
+"""
+
+import torch
+import torch.nn as nn
+
+
+class Discriminator(nn.Module):
+
+    def __init__(self):
+        """
+        Pose + Shape discriminator proposed in HMR
+        """
+        super(Discriminator, self).__init__()
+
+        self.num_joints = 34
+        # poses_alone
+        self.D_conv1 = nn.Conv2d(9, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv1.weight)
+        nn.init.zeros_(self.D_conv1.bias)
+        self.relu = nn.ReLU(inplace=True)
+        self.D_conv2 = nn.Conv2d(32, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv2.weight)
+        nn.init.zeros_(self.D_conv2.bias)
+        pose_out = []
+        for i in range(self.num_joints):
+            pose_out_temp = nn.Linear(32, 1)
+            nn.init.xavier_uniform_(pose_out_temp.weight)
+            nn.init.zeros_(pose_out_temp.bias)
+            pose_out.append(pose_out_temp)
+        self.pose_out = nn.ModuleList(pose_out)
+
+        # betas
+        self.betas_fc1 = nn.Linear(41, 10)  # SMAL betas is 41
+        nn.init.xavier_uniform_(self.betas_fc1.weight)
+        nn.init.zeros_(self.betas_fc1.bias)
+        self.betas_fc2 = nn.Linear(10, 5)
+        nn.init.xavier_uniform_(self.betas_fc2.weight)
+        nn.init.zeros_(self.betas_fc2.bias)
+        self.betas_out = nn.Linear(5, 1)
+        nn.init.xavier_uniform_(self.betas_out.weight)
+        nn.init.zeros_(self.betas_out.bias)
+
+        # bones
+        self.bone_fc1 = nn.Linear(24, 10)  # SMAL betas is 41
+        nn.init.xavier_uniform_(self.bone_fc1.weight)
+        nn.init.zeros_(self.bone_fc1.bias)
+        self.bone_fc2 = nn.Linear(10, 5)
+        nn.init.xavier_uniform_(self.bone_fc2.weight)
+        nn.init.zeros_(self.bone_fc2.bias)
+        self.bone_out = nn.Linear(5, 1)
+        nn.init.xavier_uniform_(self.bone_out.weight)
+        nn.init.zeros_(self.bone_out.bias)
+
+        # poses_joint
+        self.D_alljoints_fc1 = nn.Linear(32 * self.num_joints, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc1.weight)
+        nn.init.zeros_(self.D_alljoints_fc1.bias)
+        self.D_alljoints_fc2 = nn.Linear(1024, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc2.weight)
+        nn.init.zeros_(self.D_alljoints_fc2.bias)
+        self.D_alljoints_out = nn.Linear(1024, 1)
+        nn.init.xavier_uniform_(self.D_alljoints_out.weight)
+        nn.init.zeros_(self.D_alljoints_out.bias)
+
+    def forward(self, poses: torch.Tensor, betas: torch.Tensor, bone=None) -> torch.Tensor:
+        """
+        Forward pass of the discriminator.
+        Args:
+            poses (torch.Tensor): Tensor of shape (B, 23, 3, 3) containing a batch of poses (excluding the global orientation).
+            betas (torch.Tensor): Tensor of shape (B, 41) containing a batch of SMAL beta coefficients.
+        Returns:
+            torch.Tensor: Discriminator output with shape (B, 25)
+        """
+        # bn = poses.shape[0]
+        # poses B x 207
+        # poses = poses.reshape(bn, -1)
+        # poses B x num_joints x 1 x 9
+        poses = poses.reshape(-1, self.num_joints, 1, 9)
+        bn = poses.shape[0]
+        # poses B x 9 x num_joints x 1
+        poses = poses.permute(0, 3, 1, 2).contiguous()
+
+        # poses_alone
+        poses = self.D_conv1(poses)
+        poses = self.relu(poses)
+        poses = self.D_conv2(poses)
+        poses = self.relu(poses)
+
+        poses_out = []
+        for i in range(self.num_joints):
+            poses_out_ = self.pose_out[i](poses[:, :, i, 0])
+            poses_out.append(poses_out_)
+        poses_out = torch.cat(poses_out, dim=1)
+
+        # betas
+        betas = self.betas_fc1(betas)
+        betas = self.relu(betas)
+        betas = self.betas_fc2(betas)
+        betas = self.relu(betas)
+        betas_out = self.betas_out(betas)
+
+        # bone
+        if bone is not None:
+            bone = self.bone_fc1(bone)
+            bone = self.relu(bone)
+            bone = self.bone_fc2(bone)
+            bone = self.relu(bone)
+            bone_out = self.bone_out(bone)
+
+        # poses_joint
+        poses = poses.reshape(bn, -1)
+        poses_all = self.D_alljoints_fc1(poses)
+        poses_all = self.relu(poses_all)
+        poses_all = self.D_alljoints_fc2(poses_all)
+        poses_all = self.relu(poses_all)
+        poses_all_out = self.D_alljoints_out(poses_all)
+
+        if bone is not None:
+            disc_out = torch.cat((poses_out, betas_out, poses_all_out, bone_out), 1)
+        else:
+            disc_out = torch.cat((poses_out, betas_out, poses_all_out), 1)
+        return disc_out

prima/models/heads/__init__.py

@@ -0,0 +1 @@
+from .smal_head import build_smal_head