On the issue of increasing machine learning performance at the data sampling stage when solving classification problems

Abstract

The purpose of this study is to determine a data storage method for machine learning tasks of neural networks and semantic segmentation of point clouds. The existing methods of working with large files are considered, experimental studies are carried out to determine the speed of the data reading operation. The experiment consisted in reproducing the process of accessing information from files for which the volume and structure of stored information with time measurement were described. For the research, the most common file formats used for storing information were taken *.csv, *.npy and *.h5. The result of the experiment was statistical information about the file reading time depending on the selected structure and the amount of information stored in it, as well as recommendations for choosing a storage method.