Utilize various data augmentation techniques and explore accuracy prediction on CIFAR-10.
First train a conditional gan:
cd cgan/
python main.py
Then generate samples from the trained conditional gan:
python sample.py
Finally train the classifier:
cd ../vggnet/
python main.py
You can additionally train a discriminator to differentiate real from generated images:
cd ../cgan/
python threshold.py
Then use it to generate synthetic datasets that have a "realness" above a certain threshold:
python sample.py --threshold=0.9
Then train the classifier with this thresholded dataset:
python main.py --p_cifar=1. --p_thresholded=1. --threshold=0.9