Investigation of Transfer Learning Freeze Model Perfomance in Different Epoch

Abstract

This research aimed to investigate the performance of transfer learning in the Oxford Flower 102 class data. The study used a Resnet 34 model with three different epochs, 10, 15, and 20. The training data was augmented by applying random rotation, cropping, mirroring, and normalization, while the test data was augmented by resizing and normalizing. The model was trained using cross-entropy as a measure of accuracy. The results showed that the model with 20 epochs achieved the highest accuracy of 97.96%. The training loss was 0.0884, while the validation loss was 0.1008. The model with 15 epochs showed an accuracy of 97.48% with a training loss of 0.126 and a validation loss of 0.00965. The model with 10 epochs showed an accuracy of 97.84% with a training loss of 0.1505 and a validation loss of 0.1052. These results indicate that increasing the number of epochs can improve the accuracy of transfer learning in image classification, but there is also a risk of overfitting. This study provides insights into the potential of transfer learning in image classification and the impact of different epochs on the model's performance.