Standard Convolutional Neural Networks (ConvNets) rely on data augmentation (particularly, rotating training images) to deal with rotation invariant pictures. We design a new convolution layer that is rotation invariant by nature. As a result, we don't need to rotate the training images for preprocessing.
Tested under Ubuntu with Python 3.10.
pip install -r requirements.txt
Execute main.py in each folder. The required datasets are automatically downloaded.
Here are several commands for you to use :
--dataset 1 #choose dataset MNIST
--dataset 2 #choose dataset FashionMNIST
--train #retrain a model, default : false, load trained model
--augment #do data augmentation, default : false