Code accompanying the papers:
- Sobolev GAN (arXiv) - Appeared at ICLR 2018.
- Semi-Supervised Learning with IPM-based GANs: an Empirical Study - Appeared at NIPS 2017 Workshop: Deep Learning: Bridging Theory and Practice
Tested for python 2.7, PyTorch 0.3.0.
To reproduce the CIFAR-10 result for 4000 labeled samples using the K+1
critic, imposing Fisher constraint on full critic f
and Sobolev constraint on f-
only:
python main.py --dataset cifar10 --dataroot ${DATAROOT} --cuda --outputdir ${OUTPUTDIR} \
--labeledSamples 4000 \
--SSL_critic_type Kp1 --f_component_Fisher f --f_component_Sobolev fneg