Improving Model Training with Multi-fidelity Hyperparameter Evaluation