tntrung / gan Goto Github PK

Hello! I have a question about this piece of code. Why are you training the whole autoencoder only on the MNIST dataset?

In 1D demo , after model = GAN(...).fit(data)
how can I use this model to generate new data?
I have try to feed z to model._create_generator(z) but I got
AttributeError: 'numpy.ndarray' object has no attribute 'get_shape'

hello guys,
correct me if i'm wrong.
in the paper, you have f(x,G(z)) and lambda_w = sqrt(dim_z/dim_x)
but i glanced at the code, you guys used features of x and G(z). should lambda_w be fixed to
lambda_w = sqrt(dim_z/dim_ft_of_x)?
thank you!

It seems unreasonable. And it didn't work when I applied it to my own GAN model.

Hi, Trung,

In gaan.py,
self.md_x = tf.reduce_mean(self.f_recon - self.f_fake)
According to Eq. (7) in your paper, maybe
self.md_x = tf.reduce_mean(self.f_real - self.f_fake)

Is it correct?
Thanks,
Sungwoong.

Hi, Ngoc-Trung,

Thanks for your code sharing.
I have a question regarding a computation of gradient penalty in your code.

In gaan.py,
epsilon = tf.random_uniform(shape=tf.shape(self.X), minval=0., maxval=1.)

I think for the convex combination for each sample (same epsilon should be applied to all dims in each sample),
epsilon = tf.random_uniform(shape=[tf.shape(self.X)[0],1], minval=0., maxval=1.)

Is it correct?

Thanks,
Sungwoong.