Hi Eric,

I really enjoyed this tutorial! I do have one question about the manifold alignment part, though.

This concept of aligning the two manifolds seems to be quite important indeed, and several authors are investigating it more in depth.

In this example, I'm a little unsure how the sorting works. Namely, I thought the goal in minimizing the transformation distance from Z to X would require changing the initial weights of the generator such that the 'ordering' of z inputs to x output is (loosely) as monotonic as possible, assuming one has a definition of ordering.

So my understanding was that each specific input z_i corresponds to its specific generated pair x_i = G(z_i), and the goal was to make this relationship as monotonic as possible. But you don't keep this correspondence, right? You generate a set of x_i's from a sampled batch of z_i's and force the ordering, but isn't this artificial? I don't see why this would actually align the latent z-space and the X space?

Dear eric,
I want to understand GAN but myself I am not good in TensorFlow and I didnt get much from your tensorflow tutorial. Can you please write another tutorial code in keras (with Tf or Th backend) instead of pure TensorFlow.

Thank you for considering this
Potholiday

Is setting trainable = False not necessary in the variable batch in the function momentum_optimzer?
ie. batch = tf.Variable(0,trainable=False,name=global_step) instead of just
batch = tf.Variable(0).

Inside plot_fig function,
the x-range of histc may fall roughly between -3.0<x<1.0,
because np.histogram range is gs.min() < x < gs.max(), if no range is specified

This may result in a horizontal shift of the following ax.plot which specified -5<x<5