Git Product home page Git Product logo

lecam-gan's Introduction

Regularizing Generative Adversarial Networks under Limited Data

[Project Page][Paper]

Implementation for our GAN regularization method. The proposed regularization 1) improves the performance of GANs under limited training data, and 2) complements the exisiting data augmentation approches.

Please note that this is not an officially supported Google product.

Paper

Please cite our paper if you find the code or dataset useful for your research.

Regularizing Generative Adversarial Networks under Limited Data
Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, Weilong Yang
Computer Vision and Pattern Recognition (CVPR), 2021

@inproceedings{lecamgan,
  author = {Tseng, Hung-Yu and Jiang, Lu and Liu, Ce and Yang, Ming-Hsuan and Yang, Weilong},
  title = {Regularing Generative Adversarial Networks under Limited Data},
  booktitle = {CVPR},
  year = {2021}
}

Installation and Usage

We provide three implementations: biggan_cifar, biggan_imagenet, and stylegan2. Plesase refer to the README.md file under each sub-folder for the installation and usage guides.

lecam-gan's People

Contributors

hytseng0509 avatar roadjiang avatar

Stargazers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar

Watchers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar

lecam-gan's Issues

StyleGAN2

What does 'Third_party' refer to in the stylegan2 implementation?

The formula in the paper

The formula (4) in the paper is
R_{LC}= E_{x∼T}[||D(x)−\alpha_{F}||^2]+ E_{z∼p_z}[||D(G(z))−\alpha_R||^2]
but according to my understanding, this will not make the result converge to a certain point stably, but will cause the result to oscillate, and this does not conform to the legend (2) in the paper, is this a clerical error?

Low-shot training

How to train the Lecam-gan on the low-shot image generation datasets,THX.

Implementation for StyleGAN2

Hi
I noticed there is only the code for Lecam regularizer in the StyleGAN2 folder. Is the full implementation not available in the repo? Are we supposed to manually add the regularizer to the official StyleGAN2 code ourselves?

Unable to run demo file in imagenet folder

The model file is not accessible and call to model file raises the following exception.

model_path=gs://robust-gan/lc-biggan-imagenet128-100/240000
batch_size=2
rm: cannot remove '/tmp/models/*': No such file or directory
ServiceException: 401 Anonymous caller does not have storage.objects.list access to the Google Cloud Storage bucket.
CommandException: 1 file/object could not be transferred.
download model file to :

Can you please provide the file or share where can I get it?

ReLU on Discriminator activations?

Thank you for making your implementation available, I'm quite excited to apply the regularizer to ongoing projects.

There was some discussion online recently about a possible mismatch between Eqn 4 in a public preprint of the paper (available here) and the code implementation at these lines: {tf, pytorch}.

Could you please clarify if there's supposed to be a ReLU (present in the code but not the preprint) in Eqn. 4 as well? Thanks again!

Results in Cifar10

Hi,

It seems like the baseline results in your paper are different from the results in ADA/DA. For example, FID reported for Cifar10 is 5.33 in ADA and 2.68 in your paper. Please advise.

Training stylegan 2

Hello!

Thanks for the great paper and code!

Unfortunately I am having some issues with integrating the lecam_loss.py with the original stylegan2-ada repo.

I tried hardcoding the functionality of the lecam_loss.py function into the stylegan2 loss in here. It trains but it starts producing aliens and in it quite quickly turns into green images.

Could you please clarify how it is done in more detail?

Decay factor is 0.99 in the paper but 0.9 in the code

Hello, thanks for sharing this helpful repo! I found that in your code the decay factor is 0.9 which is reported as 0.99 in the paper. Although this is not a critical problem, which value should I exactly use to better reproduce the performance in the paper?

Paper:

We fix the decay factor γ to 0.99 in all experiments.

Code:

def __init__(self, init=1000., decay=0.9, start_itr=0):

Best.

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.