Create a ModelVersion manually

There are two ways to create a new ModelVersion on Picsellia:

• Automatically by exporting an Experiment as explained here
• Manually in the Private Model Registry

This page will focus on the manual creation of a ModelVersion directly from the Private Model Registry of your Organization.

1. Access a Model

First of all, you need to access the Model under which you want to create a ModelVersion.

From your Private Model Registry, click on this Model.

The list of existing ModelVersion will then be displayed. From this ModelVersion list, the creation of a new ModelVersion can be initialized by clicking on + New Version.

2. ModelVersion creation form

The ModelVersion creation form will then open. This form allows you to define the main characteristics of your ModelVersion.

You will be requested to fill in the following fields:

• Name
• Description (optional)
• Picture (optional)
• Type
• Framework (optional)
• LabelMap (optional)
• Parameters (optional)
• Docker information: Image, Tag, Flags, and Environment variables (optional)

Once the form is filled, click on Create and you will land on the ModelVersion overview of the freshly created ModelVersion.

More details about the elements that composed a ModelVersion are detailed in this page.

3. How a ModelVersion is used

A ModelVersion is more than a set of files. It can be used in two ways, which are not exclusive: as a model in its own right, ready to be deployed, and as the starting point of a future training. Understanding these mechanics tells you what to fill in the ModelVersion creation form, and what must be consistent for the ModelVersion to work as expected.

A. As a model, ready to be deployed

A ModelVersion is, first of all, a model: a performant model that you can deploy and use to make predictions, thanks to its Model files (weights, configuration, and so on). Depending on where you want it to run, you can deploy it:

• On the Picsellia Serving engine, where Picsellia handles the whole deployment workflow for you.
• On your own infrastructure, whether it runs at the edge or in the cloud, by downloading the Model files and serving the model yourself.

In both cases, you can connect the deployed model to a Deployment to monitor it and visualize its predictions. The two options are detailed in Deploy a ModelVersion.

B. As Base architecture of a next training

A ModelVersion can also be selected as the Base architecture of a new Experiment, to train a new ModelVersion from it. In that case, Picsellia relies on several of its elements to prepare and run the training:

a. The Docker image: where the training script lives

The Docker image associated with a ModelVersion contains the training script. When the ModelVersion is used as Base architecture and the training is launched from the Experiment (on the Picsellia training engine, through a Compute Connector, or on your own resources, as detailed here), this is the image that Picsellia runs, with the Tag, Flags, and Environment variables you defined.

The Docker image is therefore what makes a ModelVersion trainable from Picsellia. It is optional if you only want to store or deploy the ModelVersion, but it must be filled in, and accessible from the computing resource that runs the training, for the ModelVersion to be trained through an Experiment.

b. Parameters: the inputs your training script expects

The Parameters of a ModelVersion are the list of training parameters, each with its value, that Picsellia exposes to the user when the ModelVersion is used as Base architecture: they are pre-filled as editable Hyperparameters in the Experiment creation form, as detailed here.

The values you enter on the ModelVersion are the default values. The user of the ModelVersion can modify them when creating the Experiment.

Whatever the values are, Picsellia provides these parameters to the training script when the image is executed. The script must therefore expect them and handle them properly: the list of Parameters you define should match exactly the parameters your script reads, with default values that make it work. A missing or unexpected parameter can make the training fail or behave unexpectedly.

c. The LabelMap: a starting point, updated with the DatasetVersion

The LabelMap of a ModelVersion is the Label (and their indexes) the model was built around. When the ModelVersion is used as Base architecture, it is inherited by the Experiment, then updated with the LabelMap of the DatasetVersion attached to the Experiment.

Your training script must therefore be able to work with a LabelMap that differs from the one of the ModelVersion, for instance with a different number of classes. The ModelVersion later exported from the Experiment carries the updated LabelMap, which is why it matches the DatasetVersion used for its training.

d. Dataset Constraints: securing the DatasetVersion given to the script

A training script usually expects a precise set of DatasetVersion: for instance one aliased train and one aliased eval. Dataset Constraints let you declare this expectation on the ModelVersion: each constraint defines the number of DatasetVersion and the alias of each one.

When the ModelVersion is used as Base architecture, these constraints ensure that the number and the aliases of the DatasetVersion provided to the Experiment conform to what the script expects, which reduces errors at execution time. In the Experiment creation form, selecting a constraint pre-fills the aliases to attach, as detailed here.

Dataset Constraints are not part of the ModelVersion creation form: you define them once the ModelVersion is created, from the Constraints tab of its Settings, as detailed here.

4. Add Model files

Attaching Model files such as weight, configuration file, or checkpoint file to your ModelVersion can be done once the ModelVersion has been initialized as explained just before.

To do so, click on the pen icon next to Files in the ModelVersion overview.

A modal will then open allowing you to select the Model file and give it a Name.

Once uploaded, the Model file will be related to the ModelVersion and physically stored through the Storage Connector defined as default for the current Organization.

Further details on Model Files are available here.

5. Modify a ModelVersion

Once created you can still modify some elements of a ModelVersion.

• The Name and the Description can be modified in the Settings tab.
• The LabelMap, Parameters, Docker image and Model files can be modified directly from the ModelVersion overview using the pen icon next to each section.
• The Framework and Type can be modified using the SDK

Other elements can not be modified as they are automatically generated by Picsellia, so modifying them could endanger the consistency of the ModelVersion.

However more globally, if the ModelVersion is already used in a Deployment or somewhere else, modifying it could put at risk the consistency of your workspace or create issues later in your project. This is why you should be really careful when performing such modifications.

A good practice that would avoid such risks would be to create a new ModelVersion that will carry those modifications. This way, you keep track of your work and you avoid endangering the consistency of your workspace.