Performance Optimization for Feature Extraction Section of DeepChem

Abstract

Based on the popular deep learning technique, the authors at Stanford implement DeepChem as an open source methods for the research in the fields of drug discovery, biology and so on. For the performance problem of training process of DeepChem neural network, this paper rebuilds the original serial feature extraction algorithm of DeepChem and optimizes the rebuilt serial algorithm based on the multiple processes algorithm. The experiment results show that the parallel algorithm achieves 15.38× speedup at the best compared with the serial algorithm. For the future work, first, in addition to the multiprocessing package, the other packages such as concurrent, subprocess and so on could be considered to optimize the feature extraction algorithm; Second, the serial and parallel algorithms run slower when the data block size is 150 compared with the other block sizes, optimization of training process performance for the smaller data block size is the second direction in future work.