Prerequisites

Before using Malet, you need two things: an experiment folder and a training function.

Experiment folder

Create a folder for each experiment containing a YAML config file. Malet generates log files and figures in the same folder automatically:

experiments/
└── my_experiment/
    ├── exp_config.yaml   (you create this)
    ├── log.tsv            (generated by malet.experiment)
    ├── log_splits/        (generated when using parallel splits)
    └── figure/            (generated by malet-plot)

Organizing all experiments under a shared experiments/ directory keeps configs, logs, and plots together and makes CLI plotting straightforward.

Training function

Your training function accepts a configuration object and returns a dictionary of metrics. Each metric value must be a list — one entry per epoch or step. This is what lets Malet plot training curves, select best-performing steps, and compute statistics across seeds.

def train(config, ...):
    # ... training logic ...

    return {
        'train_accuracies': train_accuracies,   # list[float], one per epoch
        'val_accuracies': val_accuracies,
        'train_losses': train_losses,
        'val_losses': val_losses,
    }