Add method for reversing GAN to get latent representation for images

[avlaskin]

Add method for reversing GAN to get latent representation for images. This can help with future utilisation of the generator network. Also this pr removes some trailing space.

[veqtor]

Nice!

[leweohlsen]

This is really useful. However, the latents being returned are all nan-values. I am working with half-precision floats. Is anyone else encountering the same problem?

[Wuvist]

@avlaskin I tried to use the reverse_gan_for_etalons method with:

latents = np.random.RandomState(1).randn(1000, *Gs.input_shapes[0][1:]) # 1000 random latents
latents = latents[[0]] # hand-picked top-1
labels = np.zeros([latents.shape[0]] + Gs.input_shapes[1][1:])
img = load_image("test.png")
Gs.reverse_gan_for_etalons(latents, labels, img)

However, keep getting the error:

InvalidArgumentError (see above for traceback): Incompatible shapes: [2] vs. [0]

Appearently, it happens to at the line

gradient = tf.gradients(loss, input_latents)

The tensor input_latents seems wrong. Is it because I shouldn't construct latents from random state?

Thank you.

[yjs0704]

This is really useful. However, the latents being returned are all nan-values. I am working with half-precision floats. Is anyone else encountering the same problem?

I got the same problem. It turned out that all my g values are greater than the initial c_min (1e9). I have changed it to 1e12 and obtained non-nan outputs but the actually generated images from the recovered latent space representations do not quite match my original inputs.

[dmenig]

Thanks for this work. I was also getting nan. I was trying to reconstitue an image with an fp16 trained model on a custom dataset. i jsut put a loss = tf.reduce_sum(tf.div(tf.pow(out_expr[0] - psy, 2), 1000.)) instead of the loss you wrote, changed c_min to 1e12 and it works.