10 Mar 2021
14:00 Master's Defense Fully distance
Theme
Selecting Efficient Virtual Machines for Training Deep Learning Models on the Cloud
Student
Eva Maia Malta
Advisor / Teacher
Edson Borin (Advisor) / Sandra Eliza Fontes de Avila (Co-supervisor)
Brief summary
Deep Learning Models have been increasingly used to solve complex problems. Its characteristic of hierarchical analysis of information allows the extraction of complex relationships existing in a set of data. However, with the increase in the complexity of the models and the amount of data, the training of these models has required the use of increasingly powerful computer systems with high acquisition costs. The Computational Cloud is a business model that allows access to several types of computer systems, including high performance systems, upon payment for use, without the user having to bear the cost of purchasing the equipment. However, correctly choosing the most appropriate computer system for training a Deep Learning model in the cloud is a challenge, as the choice must take into account factors such as execution time and cost, for example. With this in mind, this work presents a study on the behavior of training Deep Learning models in virtual machines with GPU in the computational cloud. In this study, we observed that the batch size configuration affects the model's training time and the number of times required for the model's accuracy to stabilize. In addition, we observed that the execution times of the iterations and the validation processes of each training period are stable, with the exception of the first iteration and the validation of the first period. From these observations, we proposed two methodologies to identify the ideal virtual machine type to train a given model of Deep Learning in the computational cloud. Finally, we validate the accuracy of the proposed methodologies with two different Deep Learning applications and show that, in both cases, the methodologies were able to identify the type of virtual machine with the lowest cost and / or the fastest to carry out the training.
Examination Board
Headlines:
| Edson Borin | IC / UNICAMP |
| João Paulo Papa | DCo / UNESP |
| Lúcia Maria de Assumpção Drummond | IC / UFF |
Substitutes:
| Edmundo Roberto Mauro Madeira | IC / UNICAMP |
| Moacir Antonelli Ponti | ICMC / USP |