tensorflow version error about dcgan-tensorflow HOT 5 CLOSED

You can add the line

            with tf.variable_scope(tf.get_variable_scope(), reuse=False):

before ema.apply in ops.py:35.
Note that this is a quick hack and the real fix should come from tensorflow.
Also a duplicate of #57.

from dcgan-tensorflow.

I've seen the same problem. This is due to changes in TF0.11 where exponential moving average cannot be under a reuse=True scope. See tensorflow/tensorflow#2740.
You can modify the code to force reuse=False around exponential moving average, like what's been done here.

from dcgan-tensorflow.

@ppwwyyxx Could you explain how you would apply such a modification in this code? I don't know TF well enough to figure out where exactly you would force the EMA to be done outside of a scope (or... whatever is going on...).

from dcgan-tensorflow.

We're trying to update the code to use tf.contrib.layers.batch_norm to avoid this issue, but the results we see on mnist are noticeably worse. Are there any important differences between the batch_norm code used here in ops.py and the methods used in the contrib.layers version that might explain the discrepancy?

from dcgan-tensorflow.

On tensorflow 1.12.0, I had the same problem and fixed it by adding the line:

        with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):

before ema.apply

from dcgan-tensorflow.