Logging

Motivation

Reliable experiment tracking is essential for debugging, reproducibility, and collaboration. We recommend using MLflow to centralize solver runs, compare configurations, and archive artifacts produced by DuaLip.

How to use MLflow with DuaLip

DuaLip can be used alongside MLflow to record parameters, per-iteration metrics, and final results. At a high level:

Log run parameters (what you passed to the solver): By default, the solver logs max_iter, initial_step_size, gamma etc.
Log per-iteration metrics (how optimization is progressing): custom metrics such as feasibility, step size, timing, etc.
Log objective result: the result of the solver including dual objective, primal objective, etc.

See tables below for the variables that are logged by default.

What gets logged

Solver Variables	Description
`max_iter`	Maximum number of iterations.
`gamma`	Ridge regularization parameter.
`gamma_decay_type`	Type of gamma decay.
`initial_step_size`	Initial step size.
`max_step_size`	Maximum step size.

Objective Result	Description
`dual_objective`	Dual objective value.
`primal_objective`	Primal objective value.
`regularization_penalty`	Regularization penalty.
`max_positive_slack`	Maximum positive slack.
`sum_positive_slack`	Sum of positive slack.

Quick start

Install MLflow:

pip install mlflow

(Optional) Configure MLflow: If mlflow is not enabled or not configured, MLflow will use the default local configuration.

class MLflowConfig:
        """Configuration for MLflow logging."""

        enabled: bool
        tracking_uri: str = ""
        experiment_name: str = ""
        run_name: str = ""
        log_hyperparameters: bool = True
        log_metrics: bool = True
        synchronous: bool = False

Example configuration to runninng locally:

mlflow_config = MLflowConfig(
        enabled=True,
        tracking_uri="http://localhost:5000",
        experiment_name="DuaLip",
        run_name="DuaLip",
)

Tips

Use mlflow.log_metrics(..., step=iter) so MLflow plots metrics over iterations.