Trouble getting the mnist example to converge about dcgan-tensorflow HOT 7 OPEN

I meet the same problem,it just can't converge.
But the picture in the samples folder are getting more and more clear.

from dcgan-tensorflow.

for GAN, however, convergence is likely an indication of failure...

from dcgan-tensorflow.

I suggest you not to use my code as experiments but just a reference. This code is one of my earliest practice (a year ago) with TensorFlow and now there are "too" many good papers and its implementation in generative model fields which published and open-sourced after DCGAN.

• https://github.com/openai/improved-gan/
• https://github.com/openai/pixel-cnn
• https://github.com/openai/iaf
• https://github.com/openai/InfoGAN
• and literally a lot...

But @kaihuchen showed that this code still can make some promising results like #56.

from dcgan-tensorflow.

What's the algorithms used to generate MNIST digitals? I found a lot of source programs using label information, that original paper not contains.

            h0 = tf.nn.relu(self.g_bn0(linear(z, self.gfc_dim, 'g_h0_lin')))
            h0 = tf.concat(1, [h0, y])

            h1 = tf.nn.relu(self.g_bn1(linear(h0, self.gf_dim*2*s4*s4, 'g_h1_lin')))
            h1 = tf.reshape(h1, [self.batch_size, s4, s4, self.gf_dim * 2])

from dcgan-tensorflow.

@auroua Yes , just like "conditional adversarial networks"
Adding label to training data will makes result realistic in general.

from dcgan-tensorflow.

the loss just bounced around 1.3/0.7 means "converge" in my experiments.

from dcgan-tensorflow.

Recently I did a whole bunch of experiments using cGAN, such a getting it to convert from black-and-white portraits to color ones, learning monocular depth perception, or performing image operations such as erasing the background. But eventually I ran into problem when trying to use cGAN's z vector as a neural image model, which I intend to apply to videos. The problem I had with cGAN is that its U-Net architecture is great for image-to-image mapping, but it also weakens the z vector as a latent representation. So for this purpose I think I need to go back to use DCGAN instead. Just want to point out here that DCGAN is by no means obsolete.

from dcgan-tensorflow.