Export as a model

At this stage, you should have already created several Experiment, executed the related trainings, and analyzed their results. You should now have selected the Experiment you consider a success and that suits your needs. The next step is to turn this Experiment into a ModelVersion stored in your Private Registry and ready for deployment.

1. Export as a Model

After accessing the Experiment you want to create a ModelVersion from, click on the Export as a Model button:

A modal will then open, offering two options:

Create a brand-new Model and its very first ModelVersion.
Create a new ModelVersion of an existing Model.

📘
Model Versionning system
It is important to keep the Picsellia Model Versionning system in mind: as is the case for Dataset, a Model on Picsellia is composed of one or several ModelVersion.
The ModelVersion is what can actually be considered a model in the data science ecosystem, as the model files and all model properties are contained under a ModelVersion.
This Model Versionning system mainly aims at ensuring the model's traceability throughout its whole lifecycle and any potential evolution or improvement.

A. New Model

To create a brand-new Model and its first ModelVersion from the current Experiment, use the Create Model tab.

Model and ModelVersion creation through Experiment export

In this modal, you'll first be asked to initialize your new Model with a name and description, then initialize its first ModelVersion with a name and description. The Model and ModelVersion are created in your Private Registry once you click Export.

A good practice is to name and describe your Model after its actual detection purpose, and the ModelVersion with more specific details that might change in a further version — but naming and description conventions are ultimately up to you.

B. New ModelVersion

To export the current Experiment as a new ModelVersion of a Model that already exists in your Private Registry, switch to the Create Model Version tab.

ModelVersion creation of an existing Model through Experiment export

In this modal, select the existing Model, from your Private Registry, under which you want to create a new ModelVersion with the current Experiment. Then give a name and description to the new ModelVersion. It is created in your Private Registry once you click Export.

2. Exported items

Exporting an Experiment as a ModelVersion creates a new ModelVersion in your Private Registry, and this new object inherits several items from its source Experiment.

For more details about the Model and ModelVersion structure on Picsellia, see Model Versionning system.

Note that if some items are missing from the Experiment, they simply won't be exported, and the related item in the exported ModelVersion will be empty — all these items are optional for both Experiment and ModelVersion, and their usage depends on your needs and how you use the Picsellia platform.

A. Model Files

The most important item, which defines the ModelVersion itself, is the model file(s).

When exporting an Experiment as a new ModelVersion, Picsellia duplicates all the Artifacts related to the Experiment and copies them as model files attached to the exported ModelVersion.

B. LabelMap

The LabelMap is also inherited by the exported ModelVersion.

To be precise, the LabelMap is initialized when creating an Experiment, based on the Label defined during DatasetVersion creation. When exporting the Experiment as a new ModelVersion, the LabelMap is inherited from the source Experiment and attached to the created ModelVersion.

The LabelMap of the ModelVersion remains editable, but editing it might compromise the consistency of your ModelVersion.

C. Parameters

The parameters are also inherited from the Experiment by the exported ModelVersion, for two main purposes:

Keep track, at the ModelVersion level, of the training parameters used to train the current ModelVersion in the frame of its source Experiment.
Have the list of training parameters and their default values pre-filled in case the exported ModelVersion is used as Base architecture for a new Experiment.

D. Docker Image

As with the training parameters, the Docker image is also inherited from the Experiment by the exported ModelVersion, again for two main purposes:

Keep track, at the ModelVersion level, of the Docker image used to train the current ModelVersion in the frame of its source Experiment.
Have the Docker image ready to use for the training script execution, in case the exported ModelVersion is used as Base architecture for a new Experiment.

E. Source Experiment

To ensure the structure and consistency of your Computer Vision work — which aims at creating and operating the best-performing ModelVersion — it is essential to have a clear view of where your ModelVersion comes from.

This is why, when exporting an Experiment as a ModelVersion of your Private Registry, Picsellia ensures that you can easily retrieve, from your ModelVersion, the Experiment that gave birth to it.

F. Dataset Versions

For the exact same reasons as for the source Experiment, from your ModelVersion you can easily retrieve the DatasetVersion used to train it.