Efficient Large Scale Image Classification: A Streamlit Framework Integrating Mixed Precision Training and Real Time Performance Visualization (Case Study)

Abstract

Training massive deep learning classifiers for images introduces major hurdles in computational efficiency, resource usage, and convergence stability. This work introduces an interactive Streamlit‑based simulator for effective image classification on the Tiny ImageNet dataset (200 classes, 100,000 training samples). The simulator combines transfer learning with a pretrained ResNet‑18 backbone, mixed‑precision (AMP) training for faster computation, and GPU‑accelerated data augmentation using parallel data loaders. Experiments were run on an NVIDIA RTX GPU employing batch sizes ranging from 32 to 256 and learning rates between 1e‑4 and 1e‑3. The tuned setup reached a top‑1 validation accuracy of 64.2 % and a top‑5 accuracy of 86.7 % after five training epochs, cutting training time by 40‑45 % relative to conventional full‑precision training. Real‑time visualizations in Streamlit offered dynamic views of loss convergence, accuracy trends, and hardware usage. The findings show that merging mixed‑precision optimization with pretrained models markedly boosts training throughput while preserving model performance. Consequently, this interactive simulator acts as both a teaching aid for large‑scale deep learning optimization and a prototype framework for building scalable, resource‑efficient image classification systems.