Bin Dai

Hi @mmderakhshani 1. KL loss. What do you mean by a regular implementation? The KL loss is exactly the same as equation 6 in "Tutorial on variational autoencoders". Note that...

@mmderakhshani 1. About the negative loss. We just leave it negative. There is no need to force the loss to be positive. 2. About the adam optimizer and the learning...

@chanshing yes you are correct. I made a typo in the response. Thanks for pointing this out.

@mago876 yeah in the code we use N(x | \hat{x}, \gamma^2 I). In the paper and the discussion, we use N(x | \hat{x}, \gamma I). Sorry for the confusion. In...

Hi @chanshing 1. About mode collapse. \gamma goes to zeros does not correspond to the mode collapse issue. Let the latent space be R^\kappa, the data manifold be \chi and...

@chanshing 1. About point 2 in my last response. Yes I mean q(z) will occupy the whole R^\kappa. 2. About your experiment results. These results are interesting and I agree...

Hi @XavierXiao, thanks for your careful check about the experiment details. We originally used what we believed at the time was one of the most common crop for CelebA, but...

Hi @mmderakhshani, sorry for the late reply. For celeba, we just use the default setting in the repository. You can run `python demo.py --dataset celeba` using this code. For cifar10,...

@mmderakhshani Table 2 is applied on WAE network which is defined in https://github.com/daib13/TwoStageVAE/blob/871862382746d619e7dc7527f3d02f924e1b0af5/network/two_stage_vae_model.py#L194 We exactly follow the training protocol of the WAE paper. You can reproduce the result using the...