Open catherio opened 6 years ago
Thanks for the PR :) Re. the Inception Score changes, are those bugs also present in the canonical implementation (https://github.com/openai/improved-gan/blob/master/inception_score/model.py)? If so then I'm hesitant to change because enough people have reported results with them that they're de-facto part of the eval metric now.
Ah, I hadn't realized this code was almost verbatim copied from the openai version. Yes, these two bugs are also there. They're not the type of bugs that cause the score to be different, though, they're stacktrace errors that mean the program just doesn't run (anymore?). I'll go file this same PR against their implementation and see what they say.
PR on the canonical implementation is here: https://github.com/openai/improved-gan/pull/31
FYI, the changes proposed are now also in the canonical model. No numerical or semantic changes, merely compatibility. https://github.com/openai/improved-gan/pull/31
The o.set_shape(tf.TensorShape(new_shape))
part actually doesn't work (Tensorflow won't allow you to change the first dimension to None), and that's why you need bs=1
instead of 100. I suggest using o.shape.dims[0] = tf.Dimension(None)
instead. See https://github.com/wchen342/SketchyGAN/commit/f8fe961044da245c4f3fb812772c7a6f0a564dc4.
@catherio Thank you for your changes!
Two bugfixes:
Improved error message explaining how to download CIFAR-10
Whitespace removal (sorry, my editor does this automatically, and I assume you don't mind)