pxl-pipeline init

The init command bootstraps a new pipeline project (either processing or training) with the standard folder structure, dependencies, and metadata required to run pipelines in Picsellia.

Usage

pxl-pipeline init PIPELINE_NAME \
  --type [processing|training] \
  --template TEMPLATE_NAME \
  [OPTIONS]

Arguments

Argument	Description	Required
PIPELINE_NAME	Name of the pipeline project (and its folder).	✅

---

Options

Option	Description	Default
--type	Pipeline type: processing or training.	✅ Required
--template	Template name.	✅ Required
--output-dir	Target directory where the pipeline will be created.	. (current)
--use-pyproject	Generate a pyproject.toml for dependency management (via uv).	True
--run-config-file	Prefill config.toml from an existing run_config.toml (non-interactive).	None

Note: --run-config-file is also available on other commands (like test, smoke-test, and launch) for both training and processing pipelines. The training-only restriction applies specifically to pxl-pipeline init.

Templates

Processing

• dataset_version: process a dataset version with custom logic, inputs, and parameters.

• datalake: process data in a datalake (tagging, filtering, etc.).

• model_version: process a model version (conversion, compression, etc.).

Training

• simple: A minimal, framework-agnostic training pipeline scaffold that you can extend with any ML framework.
• yolov8: A YOLOv8-oriented training pipeline scaffold (Ultralytics).

Behavior

Processing pipelines

• Generate the full pipeline scaffold:

  • config.toml
  • steps.py
  • utils/parameters.py
  • utils/inputs.py (for processing pipelines that declare inputs)
  • .venv/ (with dependencies installed via uv)

Training pipelines

• Prompt for organization and environment if not set via env vars:

export PICSELLIA_ORGANIZATION=my-org
export PICSELLIA_ENV=STAGING

• Prompt for model version:

  • Reuse an existing private/public model version

  • Or create a new one (define model name, version, framework, inference type)

• Save deploy targets into config.toml (required for pxl-pipeline deploy):

[model_version]
origin_name = "MyModel"
name = "v1"
framework = "ONNX"
inference_type = "OBJECT_DETECTION"

To deploy the same training image to multiple model versions, use [[model_versions]] instead — see deploy — Training pipelines.

With --run-config-file, init runs non-interactively and may only store [model_version].id from the run config. That id is for local testing; deploy still requires origin_name, name, framework, and inference_type in config.toml.

Examples

Create a dataset processing pipeline

pxl-pipeline init my_dataset_pipeline \
  --type processing \
  --template dataset_version

Create a training pipeline

pxl-pipeline init my_training_pipeline \
  --type training \
  --template simple

👉 During setup, the CLI will prompt:

• Organization name (if not set in env vars)
• Picsellia environment (prod/staging/local)
• Model version reuse/creation

Project Structure

my_pipeline/
├── config.toml
├── pyproject.toml
├── uv.lock
├── Dockerfile

├── pipeline.py
├── steps.py
├── utils/
│   ├── parameters.py
│   └── inputs.py

├── runs/
│   └── run1/
│       └── run_config.toml

└── .venv/