some questions

[taki0112]

1. why do not use mean in kl_loss why sum ? in other kl_loss.. you use mean !

2. why 2 discriminator(disA, disA2) link what do you think just update it once with batch size 1?

3. why did you calculate latent regression loss using only mu? link link2 According to the paper, should we calculate the random_normal distribution using mu and stddev and calculate the loss?

4. According to the paper, KL_loss seems to fit only for the attribute, why does the code calculate the KL_loss for the content? link

[hytseng0509]

1. For kl_loss, please refer to the official Pytorch example and Appendix B from VAE paper.
2. For other kl_loss, it is just an L2 regularization loss, not a KL divergence loss. We'll change the confusing variable name.
3. From the observation stated in Section 4 of the BicycleGAN paper, using separate discriminators yields better visual results.
4. For latent regression, since the goal is to recover back to a single randomly sampled point in attribute latent space, we constrain the predicted mu with an L1 loss.
5. Sorry for the confusing loss name. It is just an L2 regularization loss, not a KL divergence loss. We will change the loss name.

[taki0112]

OK. Here is my tensorflow implementation

Let me know if there's an error in my code. Thank you for sharing your good research.