Skip to content

PyTorch implementation of Adversarially Learned Inference (BiGAN).

Notifications You must be signed in to change notification settings

9310gaurav/ali-pytorch

Repository files navigation

Adversarially Learned Inference

PyTorch implementation of Aversarially Learned Inference

SVHN samples

After 1 epoch :

After 10 epochs :

After 40 epochs :

CIFAR10 samples :

After 1 epoch :

After 100 epochs :

After 500 epochs :

To test SVHN pretrained embeddings for semi-supervised learning using L2-SVMs :

python3.5 test_semisup.py --dataset=svhn --dataroot=<dataroot> --model_path=<saved_model_path>

Note : The provided model was trained for 100 epochs and gives an error rate of 23% as opposed to 19.5% reported in the paper. The original model's training was not stable in this implementation and I've used some GAN hacks like adding instance noise and selective training. Refer to this for more details : https://github.com/soumith/ganhacks.

Cite

@article{DBLP:journals/corr/DumoulinBPLAMC16,
  author    = {Vincent Dumoulin and
               Ishmael Belghazi and
               Ben Poole and
               Alex Lamb and
               Mart{\'{\i}}n Arjovsky and
               Olivier Mastropietro and
               Aaron C. Courville},
  title     = {Adversarially Learned Inference},
  journal   = {CoRR},
  volume    = {abs/1606.00704},
  year      = {2016},
  url       = {https://arxiv.org/abs/1606.00704},
}

Releases

No releases published

Packages

No packages published

Languages