search

timsainb / GAIA

Generative Adversarial Interpolative Autoencoder (GAIA) is a Generative Adversarial Network (GAN) made up of Autoencoders (AE) trained explicitly on interpolations to promote convexity and better latent interpolations.
77 stars 19 forks source link

Questions in the paper #3

Closed Abel1802 closed 3 years ago

Abel1802 commented 3 years ago

Excuse me, I have a question to ask. In your paper: "The linear interpolations are Euclidean interpolations between pairs of points in Z, sampled from a 1-dimensional Gaussian distribution5 centered around the midpoint between zgeni and zgenj ." In other words, it requires two latent representations.

But in your code: The function of "get_midpoints(z, midpoint_div_1, midpoint_div_2)" in the network.py (line 502) only one latent representation is passed in.

There is a screenshot below that describes my problem. Thank you! code

timsainb commented 3 years ago

Hi @Abel1802 the tf.slice function is cutting the batch z into two batches first_half_batch and second_half_batch, then interpolating between them.

Abel1802 commented 3 years ago

I thought that picture 1 and picture 2 are a pair, which turned out to be half of batch. thank you