Tracking Module

The researchlab.tracking module provides tools to ensure that every experiment you run can be exactly reproduced, even if you didn't commit your code.

Experiment Tracker

The core of the tracking module is the ExperimentTracker context manager. It wraps your training code and automatically: 1. Starts an MLflow run. 2. Captures the current Git commit hash. 3. Generates a patch file containing all uncommitted changes (staged, unstaged, and untracked). 4. Logs these as tags and artifacts to MLflow.

Usage

from researchlab import ExperimentTracker

def train():
    with ExperimentTracker(experiment_name="my_experiment") as tracker:
        # Code execution here is now tracked.
        # If you change a file and run this, the change is captured in the patch.
        print(f"Run ID: {tracker.run_name}")

        # Log hyperparameters (automatically flattened)
        tracker.log_config("config.yaml")

        # ... training loop ...

CLI Reference

rlab provides a command-line interface to interact with these tracked experiments.

rlab list

Lists all local branches that have been restored from experiments.

rlab restore <run_id>

Restores the code state of a specific run into a new git branch. * Creates a branch experiment/<run_id> from the base commit. * Applies the captured patch.

rlab diff <run_id_1> <run_id_2>

Performs a three-way diff between the code states of two runs. This is incredibly useful for debugging regressions ("It worked yesterday, what changed?").