A simple CNN model in Keras without transfer learning achieves 67% accuracy on test set of CIFAR-100
In this project, I develope some CNN models with various experiment showing benefits of following techniques:
- Increasing number of filters in convolution layers.
- Increasing number of convolution layers.
- Taking advantages of Dropout layers.
- Taking advantages of Data augmentation.
- Taking advantages of Batch norm layers.
Although I create a very simple model, performance of my model is similar to ResNet50 (#15) in the below link:
https://paperswithcode.com/sota/image-classification-on-cifar-100?tag_filter=3
Training process of model is approximately 3 hours on Google colab with 1 GPU.
My project is implemented as a small project when studying in University of Sydney with email: tung6100@uni.sydney.edu.au