Comparative Analysis of Machine Learning Performance : Conventional Computer Versus Super-computer in Convolutional Neural Network Training
Tauriainen, Joni-Heikki (2024)
2024
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202405079890
https://urn.fi/URN:NBN:fi:amk-202405079890
Tiivistelmä
This study examined the difference in running time when a complete Convolutional Neural Network (CNN) learning process was performed on a regular computer versus a supercomputer with data collected by the Oulu University of Applied Sciences (OAMK) 6G studies. The data collection was conducted by an application provided by the Oulu University of Applied Sciences for 6G simulation. The simulation tool is intended to enhance the positioning of the simulated 6G signal transmitter and receiver. This positioning is expected to be accomplished with the assistance of a machine learning script in the future.
In this research, a significant number of teaching image with label pairs for the Convolutional Neural Network (CNN) were collected. This was achieved by running a simulation program and capturing images from the computer screen with a webcam into a Pickle file, which contains the necessary metadata. To ensure the reliability of the collected data, an automated validation system was developed in Python.
The CNN training code was developed in Python using the TensorFlow library. The code was designed to facilitate the modification of the teaching parameters employed. The primary parameters that could be altered were the number of images utilized for training and the pixel size of the image. Additionally, the program prints out indicating the total time required to complete the process, which included the time spent loading the teaching images into memory and the time dedicated to the actual teaching phase. The times were tabulated and subjected to statistical analysis in order to determine the performance of the machines in the given task.
The findings of the research indicate that it is not possible to establish a definitive threshold at which it would be advantageous to transition from the use of a conventional machine to that of a supercomputer in the context of machine learning. This decision is largely contingent upon the availability of resources. However, the study revealed that as the quantity of training data increases, the performance of a standard machine becomes increasingly constrained in the absence of optimization. Furthermore, it was observed that an increasing the pixel size of the training image resulted in an increase in the time required for training. The execution of machine learning scripts places a significant load on the machine, potentially preventing machine from performing other tasks during that time. Therefore, if the execution time is several tens of minutes, it would be beneficial to transfer that work to a supercomputer.
In this research, a significant number of teaching image with label pairs for the Convolutional Neural Network (CNN) were collected. This was achieved by running a simulation program and capturing images from the computer screen with a webcam into a Pickle file, which contains the necessary metadata. To ensure the reliability of the collected data, an automated validation system was developed in Python.
The CNN training code was developed in Python using the TensorFlow library. The code was designed to facilitate the modification of the teaching parameters employed. The primary parameters that could be altered were the number of images utilized for training and the pixel size of the image. Additionally, the program prints out indicating the total time required to complete the process, which included the time spent loading the teaching images into memory and the time dedicated to the actual teaching phase. The times were tabulated and subjected to statistical analysis in order to determine the performance of the machines in the given task.
The findings of the research indicate that it is not possible to establish a definitive threshold at which it would be advantageous to transition from the use of a conventional machine to that of a supercomputer in the context of machine learning. This decision is largely contingent upon the availability of resources. However, the study revealed that as the quantity of training data increases, the performance of a standard machine becomes increasingly constrained in the absence of optimization. Furthermore, it was observed that an increasing the pixel size of the training image resulted in an increase in the time required for training. The execution of machine learning scripts places a significant load on the machine, potentially preventing machine from performing other tasks during that time. Therefore, if the execution time is several tens of minutes, it would be beneficial to transfer that work to a supercomputer.