An Approach to Tuning Hyperparameters in Parallel: A Performance Study Using Climate Data CyberTraining: Big Data + High-Performance Computing + Atmospheric Sciences

Abstract

The ability to predict violent storms and bad weather conditions with current models can be difficult due to the immense complexity associated with weather simulation. For example when predicting a tornado caution must be used when attempting to quickly classify a weather pattern as tornadic or not tornadic. Thus one can use machine learning to quickly classify these weather patterns but great care must be taken to obtain the maximal amount of accuracy while maintaining prediction wall time. We then create a general framework for determining hyperparameters with tensorflow and keras and use it for training a convolutional neural network that specializes in classifying storms as tornadic or not tornadic based on important factors like vorticity. We demonstrate our framework’s ability to determine optimal hyperparameter values for batch size, epochs, and learning rate by examining accuracy and training time with regards to a small amount of application data. In the context of training time we leverage both CPUs and GPUs and found the performance of GPUs to be vastly superior in time taken to train the various networks than CPUs.