IntroductionΒΆ
Mantik allows to make use of MLflow to track parameters, metrics and/or articfacts when training ML models or using a model for inference.
MLflow organizes its tracking in two entities:
Experiments
Runs
A single execution of your training code usually makes up a single run. Multiple runs are then grouped in experiments. Each run might, for example, represent an individual training process of a model. Any kind of input parameter and output metrics of the training process might be tracked to a run by using MLflow commands inside your training script. Multiple runs (i.e. training processes) may then be compared with one another to find the optimal set of hyperparameters for that model.
A single execution of your training code usually makes up a single run. Multiple runs are then grouped in experiments. For example, while a run might represent an individual training process of a model, an experiment would contain all runs that train a specific model (e.g. U-Net). Any kind of input parameter and metrics related to model training may be tracked to a run by using MLflow commands inside your training script. Multiple runs (i.e. training processes) may then be compared on the Mantik tracking server to find the optimal hyperparameters for that model.
We recommend to get familiar with the basic MLflow functionalities in their Quickstart.
The following page will give you a detailed overview on how you can make use of MLflow to track your ML experiments with Mantik and MLflow.