Distributed Training Based on Horizontal Scaling for Food Image Classification

Deep learning models for food image classification rely on vast amounts of data to effectively recognize and differentiate between various food items. However, training these models on such extensive datasets presents significant computational challenges. Big data and distributed training are inherently complementary in addressing these challenges. Big data involves large volumes of information that traditional single-machine processing cannot handle efficiently. Distributed training offers a solution by enabling machine learning models to scale across multiple machines or nodes, thereby more effectively processing and learning from extensive datasets. By partitioning tasks into smaller chunks processed in parallel, distributed training optimizes computational resources, ensuring efficient analysis of big data while maintaining fault tolerance and performance scalability. This paper explores various strategies for distributed training and communication between GPUs. We compare the performance and utilization of a single P100 GPU versus two T4 GPUs on the Food-11 dataset, focusing on accuracy and training time. Our results indicate that using two GPUs yields slightly better performance than a single GPU, reducing the training time by 125 seconds, decreasing the validation dataset loss by 0.02, and increasing the accuracy by 1%. The overall accuracy achieved is 92%, with a corresponding loss of 0.28.