facetorch

API documentation, extension guide, v0.6.x migration guide, model compatibility

Docker Hub (GPU)

Facetorch is a Python library designed for facial detection and analysis, leveraging the power of deep neural networks. Its primary aim is to curate open-source face analysis models from the community, package them as portable torch.export models, and integrate them into a versatile face analysis toolkit. The library offers the following key features:

1. Customizable Configuration: Easily configure your setup using Hydra and its powerful OmegaConf capabilities.

2. Reproducible Environments: Ensure reproducibility with uv for fast Python package management, conda-lock for conda-forge dependency management, and Docker for containerization.

3. Portable Models: Models are serialized with torch.export (.pt2 format) — no model source code needed at inference time, with dynamic batch support and torch.compile compatibility.

4. Governed Extensibility: Custom readers, processors, and configurations can be supplied directly. Adding a hosted model to the built-in defaults also requires an immutable manifest entry, validation evidence, provenance, and model-rights approval.

5. Deterministic Input: Accepts local paths, tensors, NumPy arrays, PIL images, and bytes through one canonical pipeline. Remote input requires an explicit, bounded URLReader configuration.

Facetorch provides an efficient, scalable, and user-friendly solution for facial analysis tasks, catering to developers and researchers looking for flexibility and performance.

Requirements

• Python >= 3.10 and < 3.13
• PyTorch 2.6.x through 2.13.x with the matching torchvision minor
• Linux x86-64 is the official v1 candidate platform; Windows, macOS, ARM, and Apple MPS are experimental

The exact candidate matrix, named CUDA pairs, experimental platforms, and current model-rights gates are documented in Model compatibility and governance.

Please use this library responsibly and with caution. Adhere to the European Commission's Ethics Guidelines for Trustworthy AI to ensure ethical and fair usage. Keep in mind that the models may have limitations and potential biases, so it is crucial to evaluate their outputs critically and consider their impact.

Install

[!IMPORTANT] This branch prepares 1.0.0rc4 (Beta), which is not yet published. Use the versioned installation and container commands below only after its approved publication on PyPI and Docker. Bare pip install facetorch and Docker latest remain on the stable 0.6.2 line during the RC soak. Conda-forge is asynchronous and must be verified separately. None of those unpinned routes can be assumed to provide the v1 API shown below.

Torch 2.13.0 / torchvision 0.28.0 is the recommended pair for new installations within the current supported range. Older supported integrations remain welcome; see the Torch support policy.

Use a virtual environment and install the supported CPU PyTorch cohort first. This avoids pip selecting a multi-gigabyte CUDA dependency graph on a CPU host:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install --index-url https://download.pytorch.org/whl/cpu \
  "torch==2.13.0+cpu" "torchvision==0.28.0+cpu"
python -m pip install "facetorch==1.0.0rc4"

For the validated CUDA 13.0 cohort, use a compatible NVIDIA host and replace the PyTorch install above with:

python -m pip install --index-url https://download.pytorch.org/whl/cu130 \
  "torch==2.13.0+cu130" "torchvision==0.28.0+cu130"

Every matching pair from Torch 2.6/torchvision 0.21 through Torch 2.13/torchvision 0.28 is supported; see the exact compatibility matrix. The default model selection then needs approximately 1.2 GB of cache data and at least 2 GB of free cache space while downloads are staged.

Conda-forge remains an asynchronous channel. Do not use its unversioned install command for the RC; wait until the feedstock displays 1.0.0rc4, then pin that exact version.

Usage

Docker option

Docker is optional for the Python API. With Docker and Docker Compose, use the immutable RC image tag rather than latest:

Run docker example

The production image contains the example script. Compose mounts data/input read-only and keeps generated images in the facetorch-output volume, so the non-root container never needs write access to the source checkout.

CPU:

FACETORCH_DOCKER_TAG=1.0.0-rc.4 docker compose run --rm facetorch \
  python /opt/facetorch/example.py /workspace/data/input/test.jpg \
  --output /workspace/data/output/test.png

GPU:

FACETORCH_DOCKER_TAG=1.0.0-rc.4 docker compose run --rm facetorch-gpu \
  python /opt/facetorch/example.py /workspace/data/input/test.jpg \
  --profile gpu --output /workspace/data/output/test-gpu.png

Copy a result from the persistent volume into the checkout when needed:

FACETORCH_DOCKER_TAG=1.0.0-rc.4 docker compose run --rm -T facetorch \
  cat /workspace/data/output/test.png > data/output/test.png

Check data/output for the copied image with bounding boxes and facial 3D landmarks.

(Apple Mac M1) Use Rosetta 2 emulator in Docker Desktop to run the CPU version.

Python API

This copy-paste smoke checks an installed wheel, packaged configuration, reader, and result contract without downloading models or requiring a repository checkout:

import torch

from facetorch import FaceAnalyzer, load_config

cfg = load_config()
analyzer = FaceAnalyzer(cfg.analyzer)
result = analyzer.run(
    image_source=torch.zeros((3, 32, 32), dtype=torch.uint8),
    skip_detector=True,
    include_predictors=[],
    include_tensors=True,
)

assert len(result.faces) == 1
assert result.image is not None
assert not analyzer.detector_loaded
assert analyzer.loaded_predictors == ()

For normal analysis, provide one image. The first call downloads only the selected compatible model artifacts; subsequent calls reuse the local cache:

from facetorch import FaceAnalyzer, InputSpec, load_config

cfg = load_config()  # packaged CPU profile; independent of the working directory
analyzer = FaceAnalyzer(cfg.analyzer)

# Analyze one local path, tensor, NumPy array, PIL image, or bytes object
result = analyzer.run(image_source="path/to/image.jpg")

# Run only specific predictors
result = analyzer.run(image_source="image.jpg", include_predictors=["fer", "embed"])

# Inspect configured names without loading their models
print(analyzer.configured_predictors)

# Batch faces detected within this image; source-image batches are not supported
result = analyzer.run(image_source="image.jpg", face_batch_size=8)

# Strictly describe a normalized pre-cropped CHW face tensor
result = analyzer.run(
    image_source=face_tensor,
    input_policy="strict",
    input_spec=InputSpec(layout="CHW", value_range="0_1", color_space="RGB"),
    skip_detector=True,
    include_tensors=True,
)

print(result.faces)
print(result.image)  # retained because include_tensors=True

# FaceAnalyzer is also callable
result = analyzer("image.jpg")

input_policy="coerce" is the default. It applies documented, deterministic range and channel conversions and emits InputCoercionWarning when a conversion may surprise the caller. input_policy="strict" accepts source-specific uint8 RGB conventions by default and requires InputSpec for declared layout, range, color, or alpha conversions. Torch defaults to CHW/BCHW; NumPy defaults to HWC/BHWC; a four-dimensional input must have B=1.

Every successful call returns AnalysisResult. Set include_tensors=True to retain the optional image, tensor, and detection fields. The v0.x batch_size name is a warning alias for face_batch_size throughout v1.x; supplying both is an error. face_batch_size is an upper bound: the shipped predictor artifacts automatically split requests into batches of at most 64, while custom predictors may declare a different max_batch_size. Code that temporarily needs the old Response/ImageData union can call the explicit, warning-emitting analyzer.run_legacy(...) adapter.

Detector and predictor models are loaded only when execution requests them, then cached for the lifetime of that FaceAnalyzer. Selection-linked utilizers are lazy too. Their predictor requirements are declared explicitly by analyzer.utilizer_dependencies, so component names do not have to match and excluding a required predictor cannot trigger a utilizer metadata download. include_predictors=None runs every configured predictor, while include_predictors=[] runs none. Exclusions follow the same configured ordering. Unknown or duplicate names and any simultaneous use of include_predictors and exclude_predictors are rejected before the image is read or a model is loaded. skip_detector=True never constructs the detector. Selected predictors still require a configured face unifier; use include_predictors=[] when no unifier is configured and predictor-free processing is intended.

Use configured_predictors to inspect names without loading models; loaded_predictors, loaded_utilizers, and detector_loaded expose current cache state. Accessing analyzer.detector or a value in analyzer.predictors or analyzer.utilizers explicitly loads and caches that component. Lazy initialization is protected against concurrent construction, but concurrent run() calls are not guaranteed safe because configured custom readers and processors may be stateful. Use one analyzer per worker or synchronize calls externally. Timing instrumentation supports overlapping calls on separate analyzers, retains no measurement history, and only measures calls while debug logging is enabled.

Detector and predictor configs may set compile_model: true and a compile_options mapping; the options are passed unchanged to torch.compile when that selected component is first loaded. Model wrapper constructors accept only declared options now—custom extensions should expose their own constructor parameters instead of relying on arbitrary attributes.

Network access is never inferred from a string. To accept a remote image, configure facetorch.analyzer.reader.URLReader explicitly with allowed HTTP schemes, a timeout, maximum redirects, and a maximum response size. The reader resolves and rejects loopback, private, link-local, and other non-public targets before every request and redirect; it is not a general-purpose internal-network fetcher.

Configure

load_config() composes configuration resources installed inside the facetorch package. It is the supported default for library users and works without a cloned repository:

from facetorch import load_config, load_config_from_path

cpu_cfg = load_config()
gpu_cfg = load_config("gpu")
customized = load_config(
    overrides=[
        "analyzer.optimize_transforms=false",
        "analyzer/predictor/fer=efficientnet_b0_7",
    ]
)

# Advanced deployments may compose an explicit external Hydra tree.
external = load_config_from_path(
    "/srv/my-app/conf/config.yaml",
    overrides=["analyzer.device=cpu"],
)

The no-argument loader deliberately selects CPU because it is the portable, device-independent baseline. GPU support has not been removed: use load_config("gpu"), the --profile gpu script option, or the GPU container to select CUDA explicitly. Release validation covers both devices, and predictor batches remain multiple faces detected within one source image—not multiple source images.

Overrides use standard Hydra override strings and are applied after the selected CPU/GPU profile. load_config_from_path() resolves a relative path against the caller's working directory and composes the file's defaults list using its parent directory as the Hydra configuration root.

The old source-checkout pattern OmegaConf.load("conf/config.yaml") remains a low-level OmegaConf operation, but it does not compose Hydra defaults and is not the installed-library configuration API. Source and managed deployments that maintain an external Hydra tree should migrate to load_config_from_path() during v1.x.

Runtime paths

Models and generated model metadata use a versioned, OS-appropriate user cache. Configuration loading does not create it; directories are created only when a selected artifact is first needed. The following environment variables provide explicit deployment overrides:

• FACETORCH_CACHE_DIR: root for all facetorch caches.
• FACETORCH_MODEL_DIR: optional model-only override.
• FACETORCH_METADATA_DIR: optional metadata-only override.
• FACETORCH_OFFLINE: 1, true, yes, or on forbids every model and metadata network request; invalid values fail configuration loading.

Linux follows XDG_CACHE_HOME and otherwise uses ~/.cache/facetorch; macOS uses ~/Library/Caches/facetorch; Windows uses the local application-data cache. File logging and image output are disabled by default. Enable them only with an explicit writable analyzer.logger.path_file override or path_output argument.

For containers, mount a persistent volume and set the same cache root explicitly, for example FACETORCH_CACHE_DIR=/var/cache/facetorch with a volume mounted at /var/cache/facetorch. A cache populated by an online run can be mounted at that same location for later no-network execution; keep its directory layout intact.

Components

FaceAnalyzer is the main class of facetorch as it is the orchestrator responsible for initializing and running the following components:

1. Reader - reads the image and returns an ImageData object containing the image tensor.
2. Detector - wrapper around a neural network that detects faces.
3. Unifier - processor that unifies sizes of all faces and normalizes them between 0 and 1.
4. Predictor dict - set of wrappers around neural networks trained to analyze facial features.
5. Utilizer dict - set of wrappers around any functionality that requires the output of neural networks e.g. drawing bounding boxes or facial landmarks.

Structure

analyzer
    ├── reader
    ├── detector
    ├── unifier
    └── predictor
            ├── embed
            ├── verify
            ├── fer
            ├── au
            ├── va
            ├── deepfake
            └── align
    └── utilizer
            ├── align
            ├── draw
            └── save

Models

The source links below are the original repositories already used by Facetorch. The weight-license column reflects the artifact-specific review approved on 2026-08-23 and recorded, with checkpoint hashes and mapping methods, in facetorch/models/governance.json. MIT and Apache-2.0 are preserved as received; neither license was converted into the other. These licenses do not grant rights to upstream training datasets.

Detector

|     model     |   source  |   params  | weight license | version |
| ------------- | --------- | --------- | ----------- | ------- |
|   RetinaFace  |  biubug6  |   27.3M   | MIT license |    1    |

1. biubug6
   • code: Pytorch_Retinaface
   • paper: Deng et al. - RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild

Predictor

Facial Representation Learning (embed)

|       model       |   source   |  params | weight license | version |
| ----------------- | ---------- | ------- | ----------- | ------- |
|  ResNet-50 VGG 1M |  1adrianb  |  28.4M  | MIT license |    1    |

1. 1adrianb
   • code: unsupervised-face-representation
   • paper: Bulat et al. - Pre-training strategies and datasets for facial representation learning
   • Note: include_tensors needs to be True in order to include the model prediction in Prediction.logits

Face Verification (verify)

|       model      |   source    |  params  |   weight license   | version |
| ---------------- | ----------- | -------- | ------------------ | ------- |
|    MagFace+UNPG  | Jung-Jun-Uk |   65.2M  | Apache License 2.0 |    1    |
|  AdaFaceR100W12M |  mk-minchul |    -     |     MIT License    |    2    |

1. Jung-Jun-Uk
   • code: UNPG and MagFace
   • paper: Jung et al. - Unified Negative Pair Generation toward Well-discriminative Feature Space for Face Recognition
   • Note: include_tensors needs to be True in order to include the model prediction in Prediction.logits
2. mk-minchul
   • code: AdaFace
   • paper: Kim et al. - AdaFace: Quality Adaptive Margin for Face Recognition
   • Note: include_tensors needs to be True in order to include the model prediction in Prediction.logits

Facial Expression Recognition (fer)

|       model       |      source    |  params  |   weight license   | version |
| ----------------- | -------------- | -------- | ------------------ | ------- |
| EfficientNet B0 7 |   sb-ai-lab    |    4M    | Apache License 2.0 |    1    |
| EfficientNet B2 8 |   sb-ai-lab    |   7.7M   | Apache License 2.0 |    2    |

1. sb-ai-lab
   • code and checkpoints: EmotiEffLib
   • paper: Savchenko et al. - Classifying Emotions and Engagement in Online Learning Based on a Single Facial Expression Recognition Neural Network

Facial Action Unit Detection (au)

|        model        |   source   |  params |   weight license   | version |
| ------------------- | --------- | ------- | ------------------ | ------- |
| OpenGraph Swin Base | lingjivoo  |   94M   | Apache License 2.0 |    1    |

1. lingjivoo
   • checkpoint and primary code: OpenGraphAU
   • related code with preserved MIT attribution: ME-GraphAU
   • paper: Luo et al. - Learning Multi-dimensional Edge Feature-based AU Relation Graph for Facial Action Unit Recognition
   • Note: The v1 candidate reuses two explicit torch.export artifact cohorts across the validated Torch 2.6-2.13 runtime range; release CPU/CUDA evidence is tracked separately.

Facial Valence Arousal (va)

|       model       |   source   |  params | weight license | version |
| ----------------- | ---------- | ------- | ----------- | ------- |
|   ELIM AlexNet    | kdhht2334  |  2.3M   | MIT license |    1    |

1. kdhht2334
   • code: ELIM
   • paper: Kim et al. - Optimal Transport-based Identity Matching for Identity-invariant Facial Expression Recognition

Deepfake Detection (deepfake)

|         model        |      source      |  params  | weight license | version |
| -------------------- | ---------------- | -------- | ----------- | ------- |
|    EfficientNet B7   |     selimsef     |   66.4M  | MIT license |    1    |

1. selimsef
   • code: dfdc_deepfake_challenge
   • challenge: Seferbekov - Deepfake Detection Challenge 1st place solution

Face Alignment (align)

|       model       |      source      |  params  | weight license | version |
| ----------------- | ---------------- | -------- | ----------- | ------- |
|    MobileNet v2   |     choyingw     |   4.1M   | MIT license |    1    |

1. choyingw
   • code: SynergyNet
   • challenge: Wu et al. - Synergy between 3DMM and 3D Landmarks for Accurate 3D Facial Geometry
   • Note: include_tensors needs to be True in order to include the model prediction in Prediction.logits

Model download

Models are downloaded on first selected use into facetorch's versioned user cache using Hugging Face Hub. Each request is pinned to an immutable repository commit. The packaged manifest selects one artifact for the active PyTorch/device pair, and records its real format, byte size, SHA-256 digest, schema cohort, and validation metadata. Downloads are staged beside the destination, verified without executing them, and atomically promoted under a process-safe cache lock. Existing files are verified again before load; corrupt entries are quarantined rather than executed.

Full verification on use is the secure default. A deployment with a separately protected, read-only cache may add verify_on_use: false to selected downloader configs to skip the repeat digest pass; release validation must keep it enabled.

Set FACETORCH_CACHE_DIR before configuration loading to choose an explicit writable or persistent location. The current default model selection occupies approximately 1.2 GB; allow at least 2 GB of free cache space for download staging and metadata. Selecting fewer predictors reduces the download.

Plan first, then explicitly confirm a multi-artifact prefetch:

from facetorch import plan_model_prefetch, prefetch_models

plan = plan_model_prefetch(
    "cpu",
    include_predictors=["fer", "embed"],
    skip_detector=False,
)
print(plan.download_bytes, plan.items)

prefetch_models(
    "cpu",
    include_predictors=["fer", "embed"],
    skip_detector=False,
    confirm=True,
)

include_predictors=[] and skip_detector=True produce an empty plan and make no model-network request. Alignment metadata is included only when the selected predictor set satisfies the composed analyzer.utilizer_dependencies requirements for the align utilizer. Bulk prefetch refuses to start before confirm=True and reports the estimated missing bytes first.

For deployment, populate the same cache online and then compose with load_config(offline=True) or set FACETORCH_OFFLINE=1. Offline mode verifies and uses the cache, or raises OfflineCacheError before inference; it never attempts a network fallback.

Legacy TorchScript is disabled by default. An eligible, manifest-pinned legacy artifact may be selected only with load_config(allow_legacy_models=True), emits LegacyModelWarning, retains its real .pt extension, and is never selected for CUDA. A missing or incompatible .pt2 does not trigger a filename/download cascade. The Google Drive model configs are deprecated and fail closed unless an operator supplies immutable size and digest metadata; the verified 3D alignment metadata remains supported.

Inspect and recover old caches without executing them:

from facetorch import (
    cleanup_quarantined_cache,
    inspect_incompatible_cache,
    inspect_legacy_cache,
    inspect_quarantined_cache,
    migrate_legacy_artifact,
    reset_incompatible_cache,
)

entries = inspect_legacy_cache("/path/to/v0-cache")
print(entries)  # includes mislabeled TorchScript stored as .pt2

migrate_legacy_artifact(
    "/path/to/v0-cache/model.pt2",
    "detector-retinaface-legacy",
    "/path/to/v1-cache/model.pt",
)

print(inspect_quarantined_cache())
cleanup_quarantined_cache(confirm=True)  # explicit, versioned-cache files only
print(inspect_incompatible_cache())
reset_incompatible_cache(confirm=True)  # after runtime/artifact remediation

Migration succeeds only for an exact manifest hash/format match and copies rather than changes the old file. Facetorch never automatically rewrites or deletes a v0.x cache. For rollback, keep v0.6.x and v1 model roots separate (for example with different FACETORCH_MODEL_DIR values); do not point v0.6.x at the v1 manifest layout. Quarantine inspection is non-destructive, and cleanup is restricted to the versioned facetorch model/metadata roots.

Models are available on the Hugging Face Hub. The legacy Google Drive folder is retained for manual backward compatibility only. The packaged manifest is approved and pins the published, digest-bound validation metadata and legal documents produced by the clean RC1 model-export commit. Model governance is also approved: all ten records are release-eligible under their pinned upstream licenses, verified checkpoint mappings, preserved attribution, and owner-approved redistribution policy. That approval does not grant rights to upstream training datasets or remove the deployment-specific consent, privacy, performance, and legal obligations listed in each model record. The coordinated RC1 release retains its separate exact-final-commit dry-run and protected-environment gates.

Maintainers export and validate every requested model locally before any remote write. Inline --upload is disabled. Publication requires a deterministic plan, an approval bound to that plan's digest, immutable parent revisions, and a resumable receipt. Each model's artifact and metadata are committed together to a candidate branch; the initial immutable manifest commit is created only after every model repository succeeds. Deterministically rendered legal documents are then committed and a final manifest binds those resulting immutable revisions. See the model publication runbook.

Why exported models?

Facetorch v1 moved default model artifacts from TorchScript (.pt) to torch.export (.pt2) so inference no longer depends on bundled model source code, custom class definitions, or TorchScript-specific runtime behavior. This makes the hosted models easier to validate, redistribute, and load across normal Python package installations. TorchScript artifacts are still useful as legacy fallbacks, but v1 workflows should prefer Hugging Face .pt2 artifacts.

torch.export serialization is tied to PyTorch's exported-program schema, so one .pt2 file is not guaranteed to load across future or older PyTorch minors. The approved manifest has artifact cohorts exported with Torch 2.6 and 2.11. The 2.6 artifacts serve runtimes 2.6-2.8, while the 2.11 artifacts serve runtimes 2.9-2.13. Package metadata accepts that contiguous, bounded range. Torch 2.3-2.5 and 2.14 or newer fail before download; no schema-major or numeric fallback is attempted. Torch 2.3 was dropped because its affected torch.load(weights_only=True) path has a critical remote-code-execution advisory. The exception policy retains nine earlier records and adds 76 exact Torch profile/version records approved on September 6 for 18 reviewed findings. All expire on 2026-11-20. The dependency gate passes under this scoped policy; affected runtimes require trusted models and metadata. The full candidate qualification and release approval remain required. See the risk treatment. Validation uses immutable CPU golden references for both CPU and CUDA artifacts, with TensorFloat-32 disabled and the numeric policy recorded. Predictor batch sizes refer only to faces from one input image; multi-image batching is not supported in v1. See model compatibility and governance for the exact candidate evidence and remaining blockers.

Execution time

Reference GPU benchmark (AU included, face_batch_size=8, utilizers disabled, default runtime):

• test.jpg (4 faces): pass1 688 ms, pass2 216 ms, pass3 311 ms, warm avg (pass2+pass3) 263 ms
• test3.jpg (25 faces): pass1 904 ms, pass2 745 ms, pass3 728 ms, warm avg (pass2+pass3) 737 ms

Full FaceAnalyzer artifact comparison (detector + predictors embed, verify, fer, au, va, deepfake, align):

• benchmark method: 12 passes/image, report warm median over passes 3-12

CPU full-stack comparison (8 passes/image, warm median over passes 3-8, torch_num_threads=16):

Environment used for this benchmark:

• GPU: NVIDIA GeForce RTX 3090
• Python: 3.10.12
• Torch: 2.11.0+cu130

Method notes:

• Reference GPU timings include pass-level values; artifact comparisons use warm medians to avoid first-pass warm-up effects.
• Timings above include all predictors, including AU.
• Utilizers were disabled for this benchmark run.
• Full TorchScript CUDA comparison required disabling TorchScript profiling/fuser paths for runtime stability (_jit_set_profiling_executor(False), _jit_set_profiling_mode(False), _jit_override_can_fuse_on_gpu(False)).
• In this setup, TorchScript is only slightly faster in warm median (about 2-4 ms, <2%); treat this as near-parity rather than a large format-level difference.
• On CPU, TorchScript was faster by about 8-9% for this full-stack run.
• TorchScript shows much slower first pass on some runs (runtime graph specialization), so use warm metrics for fair comparison.
• One can monitor component timings in logs using DEBUG level.

Development

Run the Docker container:

• CPU: docker compose -f docker-compose.dev.yml run facetorch-dev
• GPU: docker compose -f docker-compose.dev.yml run facetorch-dev-gpu

Extend facetorch

Custom predictors and detectors remain first-class in v1. The packaged model manifest secures the models shipped by facetorch; it is not an allow-list for application extensions.

There are three deliberately separate paths:

1. Install an already constructed predictor or detector directly on a FaceAnalyzer.
2. Use an external Hydra tree with a custom component and a digest-pinned model.
3. Contribute a model to facetorch's built-in defaults and complete the full artifact, compatibility, provenance, rights, and release-evidence process.

Private and third-party Hugging Face models use direct external mode: omit manifest_id and declare the exact filename, immutable Hub commit, SHA-256, byte size, format, and device in the application configuration. Only an officially shipped model belongs in facetorch/models/manifest.json.

The complete custom predictor and detector guide contains a runnable no-download example, the predictor and detector contracts, a direct immutable Hugging Face YAML example, external configuration guidance, Torch cohort responsibilities, testing guidance, and the separate checklist for contributing an officially governed built-in model.

Update environment

Dependency ownership and release channels

• pyproject.toml is the packaging source of truth for PyPI releases and pip/uv installs (including Docker build paths using uv).
• Conda package publishing (conda-forge/facetorch) is maintained outside this repository in conda-forge feedstock workflows.
• environment.yml and gpu.environment.yml are conda environment baselines for conda users.
• environments/ contains exact CPU and CUDA release profiles for every supported Torch line from 2.6 through 2.13. Each profile has its own pyproject.toml, uv.lock, and explicit official PyTorch index.
• The production CPU image uses the exact Torch 2.6 CPU profile. The production GPU image uses the exact Torch 2.6/CUDA 12.4 profile; neither image upgrades Torch after resolution.
• The GPU conda baseline is deliberately only a Python 3.12/CUDA 12.4 system layer. Its Python packages come from environments/torch-2.6-cu124/uv.lock, because conda-forge's current Torch 2.6 GPU solve requires a newer CUDA line than the validated v1 pair.
• uv uses PyPI for normal packages and explicit named PyTorch indexes only for torch and torchvision, avoiding global extra-index resolution drift.
• Overlapping dependencies between pyproject and conda env files are intentionally kept aligned.
• CI enforces alignment with python scripts/check_dependency_sync.py, audits every exact profile with python scripts/audit_dependencies.py, and emits hashed requirements plus CycloneDX SBOMs. Advisory exceptions must be approved, justified, and no longer than 90 days.

uv (used by Docker dev/test images)

• Add packages with corresponding versions to pyproject.toml dependencies
• Lock the environment: uv lock
• Sync the environment: uv sync --extra dev
• Check every exact profile: for profile in environments/*; do uv lock --check --project "$profile"; done

conda (for conda-forge users)

CPU:

• Add packages with corresponding versions to environment.yml file
• Lock the environment: conda-lock -p linux-64 -f environment.yml --lockfile conda-lock.yml
• (Alternative Docker) Lock the environment: docker compose -f docker-compose.dev.yml run facetorch-lock
• Install the locked environment: conda-lock install --name env conda-lock.yml

GPU:

• Add packages with corresponding versions to gpu.environment.yml file
• Lock the system layer: conda-lock --with-cuda 12.4 -p linux-64 -f gpu.environment.yml --lockfile gpu.conda-lock.yml
• Sync the Python layer: uv sync --project environments/torch-2.6-cu124 --frozen
• (Alternative Docker) Lock the environment: docker compose -f docker-compose.dev.yml run facetorch-lock-gpu
• Install the locked environment: conda-lock install --name env gpu.conda-lock.yml

Run tests + coverage

• Run tests and generate coverage: pytest tests --verbose --cov-report html:coverage --cov facetorch

The source distribution includes the tests, configuration trees, workflow and container definitions needed by the suite. To validate an archive independently of a checkout (Python 3.12 and uv 0.9.14):

tar -xzf facetorch-1.0.0rc4.tar.gz
cd facetorch-1.0.0rc4
uv sync --frozen --extra dev --extra release --python 3.12.12
uv run --frozen --no-sync python -m pytest -q

The suite downloads pinned model fixtures into .pytest_cache/facetorch-models on first use. Set FACETORCH_TEST_MODEL_ROOT to an already populated, verified fixture cache to reuse it. The tensor fixture is generated in pytest's temporary directory from the bundled JPEG; no separate binary tensor is needed. Two tests marked checkout require Git metadata and are explicitly skipped in an archive. CI runs this command in a fresh environment outside the checkout before accepting the distributions. Running the suite also requires Git for temporary repository fixtures and OpenSSL for the local HTTPS deadline tests.

Generate documentation

• Generate documentation from docstrings using pdoc3: pdoc --html facetorch --output-dir docs --force --template-dir pdoc/templates/

Profiling

1. Run profiling of the example script: python -m cProfile -o profiling/example.prof scripts/example.py
2. Open profiling file in the browser: snakeviz profiling/example.prof

Research Highlights Leveraging facetorch

Sharma et al. (2024)

Sharma, Paritosh, Camille Challant, and Michael Filhol. "Facial Expressions for Sign Language Synthesis using FACSHuman and AZee." Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages, pp. 354–360, 2024.

Liang et al. (2023)

Liang, Cong, Jiahe Wang, Haofan Zhang, Bing Tang, Junshan Huang, Shangfei Wang, and Xiaoping Chen. "Unifarn: Unified transformer for facial reaction generation." Proceedings of the 31st ACM International Conference on Multimedia, pp. 9506–9510, 2023.

Gue et al. (2023)

Gue, Jia Xuan, Chun Yong Chong, and Mei Kuan Lim. "Facial Expression Recognition as markers of Depression." 2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pp. 674–680, 2023.

Acknowledgements

I would like to thank the open-source community and the researchers who have shared their work and published models. This project would not have been possible without their contributions.

Citing

If you use facetorch in your work, please make sure to appropriately credit the original authors of the models it employs. Additionally, you may consider citing the facetorch library itself. Below is an example citation for facetorch:

@misc{facetorch,
    author = {Gajarsky, Tomas},
    title = {Facetorch: A Python Library for Analyzing Faces Using PyTorch},
    year = {2024},
    publisher = {GitHub},
    journal = {GitHub Repository},
    howpublished = {\url{https://github.com/tomas-gajarsky/facetorch}}
}