deciphering-autoencoders

arXiv

CIFAR-10 random generation

CelebA random generation

Dataset

Download CIFAR-10 and store it in folder ./dataset.
Generate a number of random patterns corresponding to each data point as npz file by running gen_masks.py.
Store the npz file in folder ./dataset.

Training (CIFAR-10)

Run train.py with default settings.
Finetune the model by changing some training parameters as follows:

args.run_name = 'nf128_na1-4-16_nc32_finetune' \
args.epochs = 2000 \
args.warmup = False \
args.shift_range = 0 \
args.weight_path = 'PATH to initially trained model' \

Training (CelebA)

Please refer to the paper.

Testing

Run sample.py.