Creation

It is now time to create your first Experiment that will track the training of your ModelVersion!

Once the DatasetVersion you're planning to use for ModelVersion training has been attached to your Project, you can click on the "+ New Experiment" button located in the the Experiment tab to reach the Experiment creation form.

The Experiment creation form is a two-step flow: you first pick a Base architecture, then fill in the Experiment's details on a single screen.

1. Base architecture

The first step asks you to pick the Base architecture of your Experiment, among three options: From Model, From Experiment, or From Scratch.

A. From Model

From Model lets you use any ModelVersion stored in the Model Registry (Public or Private) as Base architecture.

Selecting it opens, inline, a Select Model browser with Public and Private tabs, letting you browse the SOTA models packaged by the Picsellia team, or the ModelVersion you or another member of your Organization have already developed or imported. Once you pick a Model, a Select Version list appears below it, letting you pick the ModelVersion to actually use as Base architecture.

B. From Experiment

From Experiment lets you use an Experiment already performed in the frame of the current Project as Base architecture.

Selecting it opens, inline, the list of existing Experiment of the current Project — pick the one to use as Base architecture.

C. From Scratch

From Scratch starts your Experiment with no pre-existing configuration, letting you define your own architecture and parameters from the ground up. Unlike the two other options, it doesn't inherit any LabelMap, parameter, or Docker image, and the next step's Hyperparameters table starts empty.

In summary, the selection of the Base architecture defines the ML base (neural network architecture, weights, training script, and hyperparameters...) that will be used to train a ModelVersion in your new Experiment, and, for From Model and From Experiment, the LabelMap and parameters your Experiment inherits.

Click on Next to move to the second step.

2. Experiment details

The second step gathers everything else needed to create the Experiment: a summary of the Base architecture you picked, the general information, the Hyperparameters, and the DatasetVersion to attach.

Let's picture it with a very basic example, the finetuning of a simple YoloV8-Segmentation Model, it means that we selected From Model at the first step, selected YoloV8-Segmentation Model and YoloV8-m-segmentation as Version from the public Registry at the previous step.

A. General information

Here again, and as is the case for all Picsellia objects for traceability purposes, you'll be asked to give a name and description to your Experiment.

B. Hyperparameters

If you picked From Model or From Experiment as Base architecture, Picsellia automatically inherits its Hyperparameters and their associated default values into an editable Name / Value table; with From Scratch, this table starts empty.

You can modify the value of the inherited Hyperparameters to configure the training as expected, and add or remove Hyperparameters. Be careful though, as those Hyperparameters and their associated values will be passed as training parameters to the training script when launching the ModelVersion training in the frame of the Experiment — if some training parameters are missing or unexpected, it might impact the training script execution and cause errors.

The creation, edition or removal of parameters is made as shown below though the text filed for adding some, the pen icon for modification or the thrash icon for deletion:

C. Dataset Versions

Last but not least, you need to attach DatasetVersion to your Experiment — those DatasetVersion will be used during the training to train your ModelVersion.

The Dataset versions section lists the DatasetVersion attached to the Project, with their asset count — click on one to move it to the Selected dataset versions list below. From there, you can pick the DatasetVersion that will be used during the model training, depending on the ability of the training script to handle one or several DatasetVersion as input.

For each DatasetVersion you select, you'll be asked to provide an alias in the Selected dataset versions list. This alias is passed along with the DatasetVersion to the training script, allowing the training script (if written to receive such an alias) to understand the purpose of each DatasetVersion — for instance, differentiating the training DatasetVersion from the validation or test ones.

a. Dataset Constraints

If your Base architecture defines Dataset Constraints, a Constraint Group dropdown appears above the dataset version list.

A Dataset Constraint is a set of constraints, each of which predefines the number of DatasetVersion that can be attached to an Experiment and the alias that must be given to each attached DatasetVersion. A Dataset Constraint is always related to a ModelVersion, meaning the Dataset Constraint you'll have to fulfill depends on the selected Base architecture.

Dataset Constraints reflect the fact that the training script is designed to receive a given number of DatasetVersion with particular aliases as input. Their main purpose is to prevent the training script of the Base architecture from failing because the DatasetVersion attached and their associated aliases aren't handled.

Selecting a Constraint Group from the dropdown pre-fills the Selected dataset versions list with empty placeholders and their associated aliases, so you only need to pick the matching DatasetVersion for each, without having to write the alias yourself. This ensures that you attach the right number of DatasetVersion with the proper aliases, depending on the Base architecture selected.

3. Created Experiment

Once every required field is filled in, click on Create to instantiate your Experiment.

Once created, you'll land on the Experiment overview, where you'll retrieve all the elements defined during the Experiment creation.