An VGG net (with batchnorm and dropout) trained on CIFAR-100. You can easily modify this code to train on CIFAR-10 by changing a line in the data loader class. Achieves around 64% accuracy without data augmentation. Record on this dataset is 75%. I plan to add data agumentation to get performance up to state of the art.
Important - Please download the saves folder into the project directory. It contains the weights
https://www.dropbox.com/sh/gxwsc9u9m6f3pga/AADZdSpkOyLMk2ofeSJXrlEqa?dl=0
Here's the architecture:
Useful Links https://www.cs.toronto.edu/~kriz/cifar.html