Deep learning data handling: exploring file formats and access strategies

Abstract Accessing large volumes of data presents a significant challenge when finding the best strategies to manage the data efficiently. Deep learning applications require the processing of massive amounts of data, which implies a considerable access Input/Output (I/O) load on computer systems. During training, interaction with the I/O system intensifies as files are continuously accessed to read data sets. This persistent access could overload the file system, which, in turn, adversely impacts application performance and efficient storage system utilization. Several factors influence the I/O of these applications, and one of the most relevant is the variety of file formats in which datasets can be stored. The choice of file format depends on the use case, as each format defines how information is stored. Some file formats have features that promote efficient access to datasets during the training phase, which can improve the performance of deep learning applications. Likewise, it is also important that the format adapts to the context, in this case, to an HPC system with a parallel file system. We will propose an image preprocessing method for cases where performance improves with parallel file access. This method will transform image data sets from their original JPEG format to the more efficient HDF5 format. Thus, our research will focus on the importance of understanding the mode of data access, spatial and temporal patterns, and the level of parallelism in file access to determine whether it is advisable to change the storage format.