Experiment - Export as a model
At this stage, you should have already created several Experiment
, executed the related training, and analyzed their results. You should now have selected the Experiment
that you consider a success and that suits your needs. The next step is basically to make from this Experiment
a ModelVersion
that will be stored on your Private registry and ready for deployment.
1. Export as a Model
After having accessed the Experiment
want to create a ModelVersion
from, you simply need to click on the Export as a Model button:
A modal will then open, from this one you have two options:
- Create a brand new
Model
and its very firstModelVersion
- Create a new
ModelVersion
of an existingModel
Model Versionning System
First of all it is important to have the Picsellia Model Versioning system in mind. Indeed, as is the case for
Dataset
, aModel
on Picsellia is composed of one or severalModelVersion
.The
ModelVersion
is actually what can be considered as a model in the data science ecosystem as the Model files and all model properties are contained under aModelVersion
.This Model Versioning system mainly aims at ensuring the model's traceability during its whole lifecycle and potential evolution or improvements.
A. New Model
Model
In case you want to create from the current Experiment
a brand new Model
and its first ModelVersion
, you need to jump into the Create Model tab.
In this modal, you will be requested to first initialize your new Model
with a name and a description, then initialize its first ModelVersion
with a name and description. The Model
and ModelVersion
will be created in your Private Registry once the Export
button clicked.
A good practice would be to name & describe your Model
with the actual purpose of detection and the ModelVersion
with more specific details that might be changed in a further version. But name and description conventions are under your responsibility.
B. New ModelVersion
ModelVersion
In case the current Experiment
needs to be exported as a new ModelVersion
of a Model
already existing in your Private Registry, you can switch to the Create Model Version tab.
In this modal, you will be requested to select an existing Model
among your Private Registry under which you want to create a new ModelVersion
with the current Experiment
. Once done, you just need to give a name and description to the new ModelVersion
of the selected Model
. The ModelVersion
will be created in your Private Registry once the Export
button clicked.
2.Exported items
It is very important to know that exporting an Experiment
as a ModelVersion
will obviously create a new ModelVersion
in your Private Registry but this new object will have many items that are inherited from its source Experiment
.
For more specific details about the Model
& ModelVersion
structure on Picsellia, you can reach this page.
However please note that if some items are missing in the Experiment
, they will simply not be exported and the related item in the exported ModelVersion
will be empty, as a summary all items are optional for Experiment
as for ModelVersion
. Their usage and presence depend on your need and your use of the Picsellia platform.
A. Model Files
The most important item, that defines the ModelVersion
itself is the Model file(s).
When exporting an Experiment
as a new ModelVersion
, Picsellia will simply duplicate all the Artifact related to the Experiment
and copy them as Model files attached to the exported ModelVersion
.
B.LabelMap
LabelMap
The LabelMap
is also inherited to the exported ModelVersion
.
To be very precise, the LabelMap
is initialized when creating an Experiment
thanks to the Label
defined during DatasetVersion
creation. Then 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 it might compromise the consistency of your ModelVersion
.
C.Parameters
The parameters are also inherited from the Experiment
to the exported ModelVersion
. Basically, there are two main purposes for that.
- Keep the trace at
ModelVersion
level of the training parameters used to train the currentModelVersion
in the frame of its sourceExperiment
. - Have the list of training parameters and their default value prefilled in case the exported
ModelVersion
is used as Base Architecture of a newExperiment
D.Docker Image
As for the training parameters the Docker Image is also inherited from the Experiment
to the exported ModelVersion
. Basically, there are again the two main purposes for that.
- Keep the trace at
ModelVersion
level of the Docker Image used to train the currentModelVersion
in the frame of its sourceExperiment
. - Have the Docker Image ready to be used for the training script execution in case the exported
ModelVersion
is used as Base Architecture of a newExperiment
E.Source Experiment
In order to ensure the structure and consistency of your Computer Vision which aims at creating and operating the best-performing ModelVersion
, this is absolutely mandatory to have a very clear view of where your ModelVersion
comes from.
This is why, while exporting an Experiment
as a ModelVersion
of your Private Registry, Picsellia will ensure that from your ModelVersion
you are able to retrieve easily the Experiment
that gave birth to this ModelVersion
.
F.Dataset Versions
For the exact same reasons as for the Source Experiment, from your ModelVersion
, you will be able to retrieve easily the DatasetVersion
used to train the ModelVersion
.
Updated about 1 year ago