Skip to content

Official PyTorch implementation of Harmonizing Maximum Likelihood with GANs for Multimodal Conditional Generation (ICLR 2019)

License

Notifications You must be signed in to change notification settings

soochan-lee/MR-GAN

Repository files navigation

MR-GAN

Project | Paper

Official PyTorch implementation of ICLR 2019 paper: Harmonizing Maximum Likelihood with GANs for Multimodal Conditional Generation.

Conditional GAN models are often optimized by the joint use of the GAN loss and reconstruction loss. We show that this training recipe shared by almost all existing methods is problematic and has one critical side effect: lack of diversity in output samples. In order to accomplish both training stability and multimodal output generation, we propose novel training schemes with a new set of losses named moment reconstruction losses that simply replace the reconstruction loss.

mismatch

model

multiple tasks

Reference

If you are willing to use this code or cite the paper, please refer the following:

@inproceedings{
    lee2019harmonizing,
    title={Harmonizing Maximum Likelihood with {GAN}s for Multimodal Conditional Generation},
    author={Soochan Lee and Junsoo Ha and Gunhee Kim},
    booktitle={International Conference on Learning Representations},
    year={2019},
    url={https://openreview.net/forum?id=HJxyAjRcFX},
}

Requirements

  • Python >= 3.6
  • CUDA >= 9.0 supported GPU with at least 10GB memory

Installation

$ pip install -r requirements.txt

Preprocessing

Cityscapes

We expect the original Cityscapes dataset to be located at data/cityscapes/original. Please refer to Cityscapes Dataset and mcordts/cityscapesScripts for details.

$ python ./scripts/preprocess_pix2pix_data.py \
    --src data/cityscapes/original/leftImg8bit \
    --dst data/cityscapes/256x256 \
    --size 256 \
    --random-flip

Maps

We expect the original Maps dataset to be located at data/maps/original. We recommend you to use the dataset downloading script of junyanz/CycleGAN.

$ python ./scripts/preprocess_pix2pix_data.py \
    --src data/maps/original \
    --dst data/maps/512x512 \
    --size 512 \
    --random-flip \
    --random-rotate

CelebA

We expect the original CelebA dataset to be located at data/celeba/original with the directory structure of data/celeba/original/train and data/celeba/original/val.

# For Super-Resolution
$ python ./scripts/preprocess_celeba.py \
    --src data/celeba/original \
    --dst data/celeba/64x64 \
    --size 64

# For Inpainting
$ python ./scripts/preprocess_celeba.py \
    --src data/celeba/original \
    --dst data/celeba/128x128 \
    --size 128

Training

MR-GAN

$ python main.py --mode mr --config ./configs/{model}-{dataset}-{distribution}-{method}.yaml --log-dir ./logs/mr

Proxy MR-GAN

Train a predictor first and determine the checkpoint where the validation loss is minimized.

$ python main.py --mode pred --config configs/{model}-{dataset}-{distribution}-{method}.yaml --log-dir ./logs/predictor

Use the checkpoint as --pred-ckpt to train the generator.

$ python main.py --mode mr --config configs/{model}-{dataset}-{distribution}-{method}.yaml --log-dir ./logs/pmr --pred-ckpt ./logs/predictor/ckpt/{step}-p.pt

About

Official PyTorch implementation of Harmonizing Maximum Likelihood with GANs for Multimodal Conditional Generation (ICLR 2019)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages