Learning-By-Teaching

Learning Implicit Generative Models by Teaching Explicit Ones

Important Dependencies

- Python 2.7.14
- tensorflow-gpu 1.12.0
- numpy 1.15.4
- scikit-learn 0.20.2
- scipy 1.1.0

Datasets

- The default path for data is '/home/Data/[dataset_name]'. Please change it in dataset.py
- The MNIST dataset will be downloaded automatically.
- Cifar10 can be downloaded from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
- CelebA dataset is cropped into 64x64

Toy Data:

- ring (LBT): python LBT_toy.py -dataset ring -mog_scale 1.0 -mog_std 0.1 -gpu [GPU_ID]
- ring (LBT-GAN): python LBT-GAN_toy.py -dataset ring -mog_scale 1.0 -mog_std 0.1 -gpu [GPU_ID]

- grid (LBT): python LBT_toy.py -dataset grid -n_mixture 100 -mog_scale 0.2 -mog_std 0.01 -batch_size 2048 -batch_size_est 2048 -max_iter 2000000 -n_est 3 -n_viz 51200 -gpu [GPU_ID]
- grid (LBT-GAN): python LBT-GAN_toy.py -dataset grid -n_mixture 100 -mog_scale 0.2 -mog_std 0.01 -batch_size 2048 -batch_size_est 2048 -max_iter 2000000 -n_est 3 -n_viz 51200 -gpu [GPU_ID]

Stacked-MNIST:

- Baseline: python LBT-GAN_smnist.py -gpu [GPU_ID]
- LBT-GAN: python LBT-GAN_smnist.py -lbt -gpu [GPU_ID]

Cifar10 & MNIST:

- Cifar10: python LBT-GAN_cifar10.py -lbt -gpu [GPU_ID]
- CelebA: python LBT-GAN_celeba.py -lbt -gpu [GPU_ID]