Optimizing Learning Rate, Epoch, and Batch Size in Deep Learning Models for Skin Disease Classification

Abstract

This study explores the best combination of learning rate, number of epochs, and batch size for training deep learning models to classify skin diseases. The experiments involved analyzing how loss changes with learning rates on a logarithmic scale. The findings reveal that a learning rate of approximately 10^-2  is most effective, with 5×10⁻³  offering additional stability during training.  Various combinations of epochs and batch sizes were tested, ranging from 20 to 100 epochs and batch sizes between 32 and 128. The results show that using a batch size of 32 yielded the best outcomes, achieving a validation accuracy of 97.35% and the lowest validation loss of 0.1074. While a batch size of 128 was more efficient in terms of time, it resulted in slightly lower accuracy. The model performed optimally with 25 epochs and a batch size of 32, avoiding any signs of overfitting.  Data preparation also played a crucial role, involving steps like image resizing, pixel normalization, and data augmentation to align with the requirements of models such as VGG-19, Inception-V4, and ResNet-152. Visualizing the dataset distribution ensured data quality and class balance, allowing the model to better recognize patterns. This study offers practical insights for effectively and efficiently training deep learning models, particularly for tasks related to skin disease classification.