octoberchang / mmd-gan

MMD-GAN: Towards Deeper Understanding of Moment Matching Network

MMD-GAN

Code accompanying the paper MMD-GAN: Towards Deeper Understanding of Moment Matching Network.

Prerequisites

- Python, NumPy, Scipy
- PyTorch (v0.1.12)
- A recent Nvidia GPU

Usage

./mmd_gan [OPTIONS]
OPTIONS:
    --dataset DATASET: type of dataset (mnist/cifar10/celeba/lsun)
    --dataroot DATAROOT: path to dataset
    --workers WORKERS: number of threads to load data
    --batch_size BATCH_SIZE: batch size for training
    --image_size IMAGE_SIZE: image size of dataset
    --nc NC: number of channels in images
    --nz NZ: hidden dimension in z and codespace
    --max_iter MAX_ITER: max iteration for training
    --lr LR: learning rate (default 5e-5)
    --gpu_device GPU_DEVICE: gpu id (default 0)
    --netG NETG: path to generator model
    --netD NETD: path to discriminator model
    --Diters DITERS: number of updates for discriminator per one generator update
    --experiment EXPERIMENT: output directory of sampled images

For a quick start, please set the DATA_PATH variable in run_exp.sh to

    ./data

and run

	$ ./run_exp.sh [mnist/cifar10/celeba/lsun]

Dataset

For mnist and cifar10, the dataset will be automatically download if not exist in the designated DATAROOT directory.

For CelebA and LSUN dataset, please run the download script in ./data directory.

More Info

This repository is by Chun-Liang Li, Wei-Cheng Chang, Yu Cheng, Yiming Yang, Barnabás Póczos, and contains the source code to reproduce the experiments in our paper MMD GAN: Towards Deeper Understanding of Moment Matching Network. If you find this repository helpful in your publications, please consider citing our paper.

@article{li2017mmd,
    title={MMD GAN: Towards Deeper Understanding of Moment Matching Network},
    author={Li, Chun-Liang and Chang, Wei-Cheng and Cheng, Yu and Yang, Yiming and P{\'o}czos, Barnab{\'a}s},
    journal={arXiv preprint arXiv:1705.08584},
    year={2017}
}

For any questions and comments, please send your email to [email protected]

mmd-gan's Issues

Incorrect depth for MNIST dataset

When running the code on MNIST dataset, the depth of the "transform" object is 3 instead of 1, raising a run-time error.

weight clipping

Thanks for posting the code.
I have one question relating to one sentence in the paper.
' the gradient DELTA_θ (maxφ fφ ◦ gθ) has to be bounded, which can be done by clipping φ.'
Shouldn't it be DELTA_x? if not, what will make the gradient explode？ Thanks

How can I match higher-order moments with kernel tricks if I do not have the complete distribution but only the higher-order moments?

As you said in Appendix B.3 of ref, "It is difficult to match high-order moments, because we have to deal with high order tensors directly."
However, if I only know little information about the true distribution such as up to 4-order moment or cumulant, how can I match these information with kernel tricks?

Out-of-date Pytorch run-time errors

General out-of-date use of Pytorch (e.g. data[0] instead of data.item() for 0d tensors).

Invalid module name, containing "."

From Pytorch 0.4 on, modules are not allowed to contained "." pytorch/pytorch#6941
Raising errors in base_module.py where modules do have "."