only class 2 with GAN
class 0-9 with infoGAN
I use feature matching to train Generative model. (I define this Loss in the /matlab/+dagnn/Feature_Match_Loss.m)
1.Compile matconvnet by run gpu_compile.m which you should remove comment in it.
2.You can test this code by run test_gan_3.m or test_gan_info.m
3.If you wanna train this code, you can run train_gan_3.m or train_gan_info.m
You can find the network structure in GDnet_3.m and GDnet_info.m
1.I may miss some thing or not select a good initial parameter. So any advice is welcome.
GDnet_1 is using 32*32 random map as input
GDnet_2 is using 100 random vector and using deconv
GDnet_3 is using 100 random vector and using conv (like fc layer)
In my experiment, deconv show that the output adjacent pixel is likely. So in the minist using conv(fc layer) is better. (deconv may suit for real images such as CIFAR)
I am sorry for that. I think my GAN training code on github is not good enough to rehearsal the result in the original paper. In fact, I give up my code and turn to use the dcgan wrote in the tensrflow. The code url is https://github.com/carpedm20/DCGAN-tensorflow. You may try it. Recently I also test the code for wgan. https://github.com/martinarjovsky/WassersteinGAN It’s also awesome. I hope it can help you.