wandb_mv_logo

Description

This library provides an abstraction to perform Model Versioning using Weight & Biases.

Features

  • Version a new trained model
  • Promote a model to another stage (e.g production)

Example of usage

The following code snippet shows how to promote a newly trained model to the best_model on the validation set. First, it creates the model providing the checkpoint path, the artifact name and type, a description, and the metadata, where the metrics or any desired information is stored.

The comparision_type indicates how to compare the metrics between the two models. E.g if smaller, the smaller the metric from the new model the better

run = wandb.init(...)

# Create the desired artifact
artifact = versioner.create_artifact(
                            checkpoint='model.ckpt',
                            artifact_name='prueba',
                            artifact_type='model',
                            description='Prueba Wandb-MV',
                            metadata={
                                'val_metric': 78.0,
                                'test_metric': 0.0
                            })

# Promote the desired artifact to the 'best_model' tag
versioner.promote_model(new_model=artifact,
                        artifact_name='prueba',
                        artifact_type='model',
                        comparision_metric='val_metric',
                        promotion_alias='best_model',
                        comparision_type='smaller'
                       )

This code snippet shows how to promote a trained model to production after being validated on the test set. The already_deployed parameter indicates that the model is already logged, so it only needs to be updated.

versioner = Versioner(run)
versioner.promote_model(new_model=artifact,
                        artifact_name='prueba',
                        artifact_type='model',
                        comparision_metric='test_metric',
                        promotion_alias='production',
                        comparision_type='smaller',
                        already_deployed=True
                        )

Next features

  • Allow providing a custom comparision function to promote a model

GitHub

View Github