Pixano-Inference is an inference library for Pixano.

Under active development, subject to API change

Pixano-Inference

A Ray Serve inference server for the Pixano annotation tool, with a REST API and a Python client.

Models are independent Python packages with their own code, dependencies, and environments. The server discovers installed models automatically; the core requires no ML framework.

Quickstart

Requires Python 3.10–3.13. Install Pixano-Inference with the Grounding DINO model:

pip install "pixano-inference[grounding-dino]"

Create models.py:

from pixano_inference.configs import DeploymentConfig, ModelConfig
from pixano_inference_grounding_dino import GroundingDINOParams

models = [
    ModelConfig(
        name="grounding-dino",
        model_class="GroundingDINOModel",
        model_params=GroundingDINOParams(path="IDEA-Research/grounding-dino-tiny"),
        deployment=DeploymentConfig(num_gpus=0),  # Set to 1 to use a GPU
    ),
]

Start the server. The first run downloads the model weights.

pixano-inference --config models.py

Check readiness, then save this request as predict.py:

from pixano_inference_client import DetectionRequest, SyncPixanoInferenceClient

client = SyncPixanoInferenceClient("http://localhost:7463")
result = client.detection(
    DetectionRequest(
        model="grounding-dino",
        image="https://raw.githubusercontent.com/pixano/pixano-inference/main/docs/assets/examples/sam2/bedroom/00000.jpg",
        classes=["bed", "lamp"],
        box_threshold=0.3,
        text_threshold=0.25,
    )
)
print(result.data.classes, result.data.boxes, result.data.scores)

Run it in another terminal, using the same Python environment:

python predict.py

Applications calling an existing server only need pixano-inference-client. See the documentation for the API, Docker deployment, and autoscaling.

Model packages

Choose a model with an extra, for example pip install "pixano-inference[sam]":

Each extra installs an independent model package. SAM also needs the upstream sam-2 library; follow its package's installation instructions.

Development

For source development, clone the repository and use uv:

git clone https://github.com/pixano/pixano-inference.git
cd pixano-inference
uv sync
uv run pytest -m "not integration" tests/  # Core unit tests
uv run --project packages/pixano-inference-sam pytest packages/pixano-inference-sam/tests

Develop and test each model in its own package, with its own pyproject.toml, uv.lock, and tests. To build a custom model, start with the numpy detector and follow the guide and package specification.

License

Pixano-Inference is released under the terms of the CeCILL-C license.