GibbsNet

PyTorch implementation of GibbsNet: Iterative Adversarial Inference for Deep Graphical Models.

Requirements

• Python 3
• Pytorch
• visdom

Usage

SVHN, CIFAR10 dataset are currently supported.

Train

$ python train.py --model=GibbsNet --batch_size=100 --lr=1e-5 --dataset=SVHN --sampling_count=20
$ python train.py --model=GibbsNet --batch_size=100 --lr=1e-5 --dataset=SVHN --gpu_ids=0,1 --sampling_count=20

Visualize

• To visualize intermediate results and loss plots, run python -m visdom.server and go to the URL http://localhost:8097

Test

$ python test.py --test_count=20 --model=GibbsNet --repeat_generation=10

• Test result will generate in ./[opt.test_dir]/[opt.model/, of which default value is ./test/GibbsNet/
• Test result consists of real_[i].png files and fake_[i]_[j].png files. real_[i].png files are sampled from real dataset, and fake_[i]_[j].png files are generated from sampled latent variable of real_[i].png

Results

1. SVHN

• epoch 100, lr 1e-5, sampling_count 20
• generated results

2. CIFAR10

• Working in Progress

Implementation detail

• Original implementation of discriminator network for CIFAR10 dataset uses maxout activation layer, but this implementation uses leaky ReLU rather than maxout layer because of lack of GPU memory.
• all hyper parameters references to paper Adversarially Learned Inference.
• To train GibbsNet, appropriate learning rate is 1e-5 for sampling count 20. You can increase learning rate when you sample less than 20 times.

TODOs

• Custom dataset support
• Visualize test results

Code reference

Visualization code(visualizer.py, utils.py) references to pytorch-CycleGAN-and-pix2pix(https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) by Jun-Yan Zhu

Author

Tony Kim