Skip to content

Model Version Processing Pipeline

This guide explains how to create, customize, test, and deploy a processing pipeline that runs on a Model Version using the pxl-pipeline cli with the model_version template.

These pipelines receive a model version as their target and can apply conversions, compressions, or any custom logic to the model.

1. Initialize your pipeline

pxl-pipeline init my_model_pipeline --type processing --template model_version

This generates a pipeline folder with standard files. See project structure for details.

2. Customize your pipeline logic

steps.py

Contains the process() step where your core logic lives. The context gives you access to the target model version, parameters, and any inputs you've declared.

@step
def process():
    context: PicselliaModelProcessingContext = Pipeline.get_active_context()
    parameters = context.processing_parameters
    model_version = context.target

    # Your logic goes here ...

utils/parameters.py

Define custom parameters using a class that inherits from Parameters:

class ProcessingParameters(Parameters):
    def __init__(self, log_data):
        super().__init__(log_data)
        self.example_parameter = self.extract_parameter(
            ["example_parameter"], expected_type=str, default="default"
        )

See Working with pipeline parameters for more.

utils/inputs.py

Define the inputs your processing expects. Inputs are registered on the Picsellia platform when you deploy and are validated at launch time.

from picsellia.types.enums import ProcessingInputType
from picsellia_pipelines_cli.utils.inputs import PipelineInputs


class ProcessingInputs(PipelineInputs):
    def __init__(self):
        super().__init__()
        self.define_input(
            name="example_input",
            input_type=ProcessingInputType.TEXT,
            required=True,
        )

See Working with pipeline inputs for the full guide.

3. Configure run_config.toml for local testing

When you run pxl-pipeline init, a run_config.toml is generated:

override_outputs = true

target_id = ""

[job]
type = "MODEL_CONVERSION"

[inputs]
example_input = "example_value"

[parameters]
example_parameter = "default"

Fill in the target_id with the UUID of the model version you want to process.

4. Manage dependencies with uv

uv add onnx --project my_model_pipeline

Dependencies are declared in pyproject.toml. See dependency management with uv.

5. Test your pipeline locally

pxl-pipeline test my_model_pipeline

This will:

  • Prompt for the target model version if not set in the run config
  • Scaffold any missing inputs with empty defaults
  • Run the pipeline via pipeline.py --mode local
  • Save everything under runs/runX/

To reuse the same folder and avoid re-downloading assets:

pxl-pipeline test my_model_pipeline --reuse-dir

See how runs/ work for more details.

6. Deploy to Picsellia

pxl-pipeline deploy my_model_pipeline

This will:

  • Build and push the Docker image
  • Register the pipeline in Picsellia
  • Sync the declared inputs to the platform (add new inputs, update existing ones, remove stale ones)

See deployment lifecycle.