aylien/gan-intro

gan-intro's Issues

Request for rewriting gan-intro codes in keras?

I came across this article http://blog.aylien.com/introduction-generative-adversarial-networks-code-tensorflow/ and I am impressed by how you used a toy example of 1-D gaussian distribution to explain GAN's. I myself is not well versed in TensorFlow and because of that I understood only a small parts in your tutorial. Can I request an article about implementation of this GAN code in keras with theano or tensorflow backend.

Thank you

confuse about the define of placeholder

when I change the position of placeholder of x, it goes wrong
the original is:

self.z = tf.placeholder(tf.float32, shape=(params.batch_size, 1), name="z")

with tf.variable_scope('G'):
 ……

self.x = tf.placeholder(tf.float32, shape=(params.batch_size, 1), name="x")

with tf.variable_scope('D'):
 ……

I change to:

self.z = tf.placeholder(tf.float32, shape=(params.batch_size, 1), name="z")
self.x = tf.placeholder(tf.float32, shape=(params.batch_size, 1), name="x")

with tf.variable_scope('G'):
 ……

with tf.variable_scope('D'):
 ……

the training result is completely wrong. I am so confused, can anyone tell me why this change will effect the result.

Loss nan when "--minibatch True"

hi @johnglover
It's well to run python gan.py
But i got all nan for loss when running with minibatch discrimination python gan.py --minibatch True:

0: inf  nan
10: nan nan
20: nan nan
30: nan nan
40: nan nan
50: nan nan
60: nan nan
70: nan nan
80: nan nan
90: nan nan
100: nan    nan
110: nan    nan
120: nan    nan
130: nan    nan
140: nan    nan
150: nan    nan
160: nan    nan
170: nan    nan
180: nan    nan
190: nan    nan
200: nan    nan
......

My numpy, scipy, seaborn and tensorflow are all in the newest version.

Any tips about this nan?

Question on minibatch discrimination

In Tim Salimans' paper, minibatch discrimination is applied separately to generated and real samples. Did you try that as well ? If yes, did it have a different effect on the results ?

reordering matplotlib in requirement.txt

upon running gan.py, I was getting below error.
RuntimeError: module compiled against API version 0xa but this version of numpy is 0x9

Searching for the problem in stackoverflow gave me this solution.

placing matplotlib after numpy solved the issue:

like this:

numpy==1.11.3
matplotlib==1.5.3

Can we fix this in upstream?

Cost of generative model not reducing,

The cost of discriminate model converges fast to low value and then no perceivable change to generative model. Can you please suggest what the issue might be.
this.mp4.tar.gz