Error for running with TensorFlow: ValueError: Variable d_w1/Adam/ does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope? about generative-adversarial-networks HOT 5 CLOSED

Its running perfectly :), thank you so much

from generative-adversarial-networks.

First, try restarting the kernel, and running the notebook again. A lot of times creating Tensorflow graphs in iPython notebooks can get a little messy.

If that doesn't fix it, this is most likely an issue with scope. I think that the new Tensorflow version deals with optimizers (other than GradientDescentOptimizer) a bit differently. I think adding this with tf.variable_scope(tf.get_variable_scope(), reuse=False): before the declaration of the Adam optimizer should fix the problem, but not completely sure if this is a long term solution. Have you come across this error @jonbruner

from generative-adversarial-networks.

Adit is correct; this is usually a scope issue. When the Adam optimizer is instantiated, it creates a bunch of additional TensorFlow operations, and if TF thinks it can only reuse existing operations it throws an error.

I did come across this error and added the with tf.variable_scope line, which seemed to fix it. I'll try running again shortly to see if I can reproduce the error.

from generative-adversarial-networks.

OK, I think I've fixed the problem in gan-script.py; TF v1.0 and v1.1 might handle variable scope in slightly different ways. The code previously worked in v1.0.1 but broke in v1.1.

In the revised script, variable scope is handled at a slightly higher level; reuse_variables() is false until after the optimizers are defined, at which point it's set to true for the rest of the script. The exception is the discriminator network, which is set up once with variable reuse off and a second time with variable reuse on.

@adeshpande3, would you mind running gan-script.py to make sure I didn't miss anything? I'll make the same changes to gan-script-fast.py now.

from generative-adversarial-networks.

I've now updated gan-script-fast.py along the same lines, and I've tested it and gan-script.py on both TF v1.0.1 and v1.1.0.

@heathervica, thanks for bringing this to our attention!

I'm going to go ahead and close this issue for now; @adeshpande3, let me know if you run into any problems when you run the files.

from generative-adversarial-networks.