zcemycl / matlab-gan Goto Github PK

MATLAB implementations of Generative Adversarial Networks -- from GAN to Pixel2Pixel, CycleGAN

License: MIT License

MATLAB 94.42% Jupyter Notebook 5.58%

matlab-gan infogan cyclegan pix2pix gans image-generation deep-learning deep-neural-networks computer-vision lsgan

matlab-gan's Introduction

Matlab-GAN

Collection of MATLAB implementations of Generative Adversarial Networks (GANs) suggested in research papers. This repository is greatly inspired by eriklindernoren's repositories Keras-GAN and PyTorch-GAN, and contains codes to investigate different architectures of GAN models.

Configuration

To run the following codes, users should have the following packages,

MATLAB 2019b
Deep Learning Toolbox
Parallel Computing Toolbox (optional for GPU usage)

Datasets

Generative Adversarial Network (GAN) [code] [paper]
Least Squares Generative Adversarial Network (LSGAN) [code] [paper]
Deep Convolutional Generative Adversarial Network (DCGAN) [code] [paper]
Conditional Generative Adversarial Network (CGAN) [code] [paper]
Auxiliary Classifier Generative Adversarial Network (ACGAN) [code] [paper]
InfoGAN [code] [paper]
Adversarial AutoEncoder (AAE) [code] [paper]
Pix2Pix [code] [paper]
Wasserstein Generative Adversarial Network (WGAN) [code] [paper]
Semi-Supervised Generative Adversarial Network (SGAN) [code] [paper]
CycleGAN [code] [paper]
DiscoGAN [paper]

Outputs

GAN -Generator, Discriminator	LSGAN -Least Squares Loss	DCGAN -Deep Convolutional Layer	CGAN -Condition Embedding

ACGAN -Classification	InfoGAN mnist -Continuous, Discrete Codes	AAE -Encoder, Decoder, Discriminator	Pix2Pix -Pair and Segments checking -Decovolution and Skip Connections

WGAN	SGAN	CycleGAN -Instance Normalization -Mutli-agent Learning	InfoGAN CelebA

References

Y. LeCun and C. Cortes, “MNIST handwritten digitdatabase,” 2010. [MNIST]
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, andL. Fei-Fei, “ImageNet: A Large-Scale Hierarchical Image Database,” inCVPR09, 2009. [Apple2Orange (ImageNet)]
R. Tyleček and R. Šára, “Spatial pattern templates forrecognition of objects with regular structure,” inProc.GCPR, (Saarbrucken, Germany), 2013. [Facade]
Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learn-ing face attributes in the wild,” inProceedings of In-ternational Conference on Computer Vision (ICCV),December 2015. [CelebA]
Goodfellow, Ian J. et al. “Generative Adversarial Networks.” ArXiv abs/1406.2661 (2014): n. pag. (GAN)
Radford, Alec et al. “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.” CoRR abs/1511.06434 (2015): n. pag. (DCGAN)
Denton, Emily L. et al. “Semi-Supervised Learning with Context-Conditional Generative Adversarial Networks.” ArXiv abs/1611.06430 (2017): n. pag. (CGAN)
Odena, Augustus et al. “Conditional Image Synthesis with Auxiliary Classifier GANs.” ICML (2016). (ACGAN)
Chen, Xi et al. “InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets.” NIPS (2016). (InfoGAN)
Makhzani, Alireza et al. “Adversarial Autoencoders.” ArXiv abs/1511.05644 (2015): n. pag. (AAE)
Isola, Phillip et al. “Image-to-Image Translation with Conditional Adversarial Networks.” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016): 5967-5976. (Pix2Pix)
J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpairedimage-to-image translation using cycle-consistent ad-versarial networks,” 2017. (CycleGAN)
Arjovsky, Martín et al. “Wasserstein GAN.” ArXiv abs/1701.07875 (2017): n. pag. (WGAN)
Odena, Augustus. “Semi-Supervised Learning with Generative Adversarial Networks.” ArXiv abs/1606.01583 (2016): n. pag. (SGAN)

matlab-gan's People

Contributors

Stargazers

Watchers

Forkers

binzhou-com hustcxl robot520 minoue-xx aung2phyowai yzq1979 tang16 sawon1234 subhankar21 ashevel1 abdurad fanleishifanlei judithcodes cakchit chiuyuen-hsui zhanqizhang66 israeljsf95 zachwelborn1 kaiifu misbsear iqiukp freesciencecommunity ballic helalaashehri zggl deroes yueseu guojinglei kamkare artdatai augmingyue micklexqg wyh5201314 samira-ai-2020 fenggaoluyuan liangwenkai azhangyulin v-italy nabilaisham weihong1021 h1sat0r1 guevara111 eagle034 lighting-zhu gracebaoxp cqhxping dliofindia celineserb xiaokangdemo abcxq seokbeomroh zzxfaith samsam6414 liyuqi01 happybsln alqushaibi cawenxuan mir107 nakhmani hanpuz yangyanglcy misswayguy csyifangli limx59 xizhang-yao 1213v linz-l luchi2333 609959998 pobbyleesh zht-1994 anki6h shaoyuhlq bm137128 joy-zhang-0303 cailab-tamu

matlab-gan's Issues

Gan1

Testing code for cycleGAN

I have trained cycleGAN for dehazing the fog but I don't know how to test and predict the images. I am trying to test the images which is given in the below link:
https://www.mathworks.com/help/deeplearning/ug/unsupervised-medical-image-denoising-using-cyclegan.html
It would be great if you provide me the code for testing the GAN.
If I am trying to predict I am getting the following error
Error using predict (line 85)
No valid system or dataset was specified.

Error in testing (line 59)
generatedImageHD = predict(GradGBA,imageTestLD);

GradGBA is the weights obtained after training cycleGAN

kindly help me !!!!!!

negate the GPU Usage?

hey @zcemycl could you tell me what I should do to use only my laptop I don't have GPU so what should I do . is there a way to run the code just for testing purposes on my laptop . Like some code to negate the GPU Usage?

Originally posted by @hurmaeht in #11 (comment)

The error of code of LoadFacadeDatabase.mlx

I run the code（LoadFacadeDatabase.mlx）on the Matlab, and prompted for an error that he index exceeds the number of elements in the array .And LatCode.mat and ImgLib.mat were not found.

Hello, author, I like this app very much, but I am a beginner and have some questions I would like to ask you

Can A CGAN be used to generate images for data? What is the specific name of the generated neural network? How do you save a trained network?

With cycleGAN, the center is black even after huge epoch.

The center is black even after huge epoch with cycleGAN.

Original code without any modification.

Hope someone can point out my running errors.

Pix2Pix implementation: Bug in the Generator?

First of all, thanks for the codes.

I think that the upsampling part of the Generator of the Pix2Pix code has a couple of issues. The order of the batchnorm and relu are permuted, and the u1 layer has no relu layer defined.
I have tested it with the "right" order and it works much better. Here is the code (with the changes in bold):

u1=dltranspconv(d7,params.TCW1,params.TCb1,'Stride',2,'Cropping','same');
u1=batchnormwrap(u1,params,st,7);
u1=relu(u1);
u1=cat(3,u1,d6);

exstattran = 'u%d = dltranspconv(u%d,params.TCW%d,params.TCb%d,"Stride",2,"Cropping","same");';
exstatrelu = 'u%d = relu(u%d);';
exstatbatch2='[u%d,st] = batchnormwrap(u%d,params,st,%d);';
exstatcat = 'u%d=cat(3,u%d,d%d);';
for i = 2:7
eval(sprintf(exstattran,i,i-1,i,i));
if i ~= 7
eval(sprintf(exstatbatch2,i,i,6+i));
eval(sprintf(exstatrelu,i,i));
eval(sprintf(exstatcat,i,i,7-i));
end
end
dly = tanh(u7);
end

CGAN Minst Dataset

Hello, I wanted to ask how I can format my own dataset to be like the MINST dataset for MATLAB. I want to use the CGAN for my own dataset of coin images but I don't know how to format those images to work with this code. could you please help?

Would you consider adding data augmentation and random grouping？

How to save a model.

Hello.
Thanks for these nice codes.
Normally we do have checkpoints and model to save to a file.

I want to know how to save the model.
Is it good to save the paramGen and paramDis? and if it is then how to use as a model.
2 .And also I want to know how to test pix2pix.m to a specific image file.
Thanks in advance.

Although the function gpdl store the array on GPU, the code doesn’t run on the GPU.

Hello,thanks for your code. Although the function gpdl store the array on GPU, the code doesn’t run on the GPU.

Retraining Encoder/Generator in AAE

Hello,

Iam currently implementing an AAE and found your great work!
However, is it possible that you forgot to add the part where the Encoder/Generator is retrained? Currently the Encoder/Generator weights are only updated once with the reconstruction error from the Encoder-Decoder Part.

Dont have GPU

Hello, I wanted to use the CycleGAN but it keeps saying that the CUDA Driver is unable to load and that the GPU Driver needs to be updated or reinstalled. I checked my laptop's GPU and it is Intel(R) UHD Graphics 620. I am slightly confused as to how to make it work if I don't have nvcuda.dll installed. I even tried installing nvcuda.dll but that also didn't seem to work. Please help me out here. How do I make this work if I don't have a GPU. or how to install a GPU on my computer to make this work.

batching indexing missing is some gans

First, thank you. This is an impressive collection of work you are sharing.

Some of the GANS seem to be missing batch indexing.

Should they have

idx = (i-1)*settings.batch_size+1:i*settings.batch_size;
ABatch=gpdl(trainAshuffle(:,:,:,idx),'SSCB');
BBatch=gpdl(trainBshuffle(:,:,:,idx),'SSCB');

rather than just

ABatch=gpdl(trainAshuffle(:,:,:,i),'SSCB');
BBatch=gpdl(trainBshuffle(:,:,:,i),'SSCB');

Latent_fake AAE

Hello again,

in line 108-110 of your AAE.m file the latent space is build by:

latent_fake = dly(1:settings.latent_dim,:)+...
    dly(settings.latent_dim+1:2*settings.latent_dim)*...
    randn(settings.latent_dim,settings.batch_size);

You are only taking the first pair of standard deviations (109) and multiplicate it with all samples from the standard gaussian (110). This results into a 1 x batch_size vector instead of a latent_dim x batch_size matrix. This vector is then added to the matrix of means (line 108). I think the correct way is:

latent_fake = dly(1:settings.latent_dim,:)+...
    dly(settings.latent_dim+1:2*settings.latent_dim,:) .*...
    randn(settings.latent_dim,settings.batch_size);

Greetings
Nico