Akademik Veri Yönetim Sistemi
4th Global Conference on Engineering Research (GLOBCER’24), Balıkesir, Türkiye, 16 - 19 Ekim 2024, ss.14-20
Many hyperparameters are used for classification in convolutional neural networks (CNNs). Optimum tuning of hyperparameters plays an important role in the classification success of CNN. One of the important hyperparameters used in CNNs is batch size (BS). Changing the BS, just like changing other hyperparameters during training, is a research topic. However, changing the BS during training has been less investigated than other hyperparameters. In this study, the effect on the performance of CNNs of increasing the BS during training was examined. For this purpose, the test loss value was checked with early stopping during training, and BS was increased if the loss value increased by the determined number of steps. To examine the effect of increasing the BS, training was performed on the MNIST dataset using pre-trained MobileNet and MobileNetV2 models. The results showed that increasing the BS during training provided slightly higher accuracy and lower loss compared to using a constant BS.