HyperGBM：Job management with Hyperctl

Hyperctl is a general multi-job management tool, which includes but not limit to training, testing and comparison. This section will introduce how to use hyperctl to manage the HyperGBM training tasks.

Firstly, use the python script hypergbm/job.py or hypergbm/gpu_job.py(for using gpu to train) provided by HyperGBM to read all parameters of the job of hyperctl. Then configure these parameters and transfer them to the function hypergbm.make_experiment to create an experiment. Lastly, the experiment is executed.

Example: Use Hyperctl to train a HyperGBM classification model

• Create an directory and all operation will be executed within this directory

mkdir /tmp/hyperctl-example
cd /tmp/hyperctl-example

• Download the training dataset heart-disease-uci：

# curl -O heart-disease-uci.csv https://raw.githubusercontent.com/DataCanvasIO/Hypernets/master/hypernets/tabular/datasets/heart-disease-uci.csv
python -c "from hypernets.tabular.datasets.dsutils import load_heart_disease_uci;load_heart_disease_uci().to_csv('heart-disease-uci.csv', index=False)"

• Create the hyperctl job configuration file batch.json. We could see that it includes the parameters: train_data,target,log_level and run_kwargs:

{
    "jobs": [
        {
            "params": {
                "train_data": "/tmp/hyperctl-example/heart-disease-uci.csv",
                "target": "target",
                "log_level": "info",
                "run_kwargs": {
                  "max_trials": 10
                }
            },
            "execution": {
                "command": "python -m hypergbm.job"
            }
        }
    ]
}

Please note:

1. In the configuration file, parameters like ‘train_data’,’eval_data’ and ‘test_data’ should be replaced by the corresponding file path. And the file format coule be ‘csv’ or ‘parquet’.

2. In the configuration file, sub parameters within ‘run_kwargs’ are used to configure the experiment. They will be transfered to the function hypernets.experiment.compete.CompeteExperiment.run

• Execute the configured job：

hyperctl run --config ./batch.json