The implementation is not variational

[amir-abdi]

Hi Shengija,

Thank you for your work.
The implementation does not seem to be a variational approach. The mapping between the x and z is deterministic with no Gaussian density estimation and sampling.
I may be misinterpretting here, but I can't match this to the InfoVAE paper: https://arxiv.org/pdf/1706.02262.pdf

Thanks,
Amir

[rojinsafavi]

Any updates on this?

[amir-abdi]

@rojinsafavi
I implemented another library for disentanglement which also includes the infoVAE.
https://github.com/amir-abdi/disentanglement-pytorch
Check it out and let me know if it helped.

[jwparks]

any updates?

[Sean0719]

Who can give me some suggestions that how to get the Figure 1 in paper. Thanks.

[ecfm]

any updates?

[ShengjiaZhao]

Hi sorry for this late reply. I haven't been up to date on looking at this repo. In the tutorial, we use a simplified implementation without adding random noise. For the version with random noise please see this file
https://github.com/ShengjiaZhao/MMD-Variational-Autoencoder/blob/master/mmd_vae_eval.py

[ecfm]

Thank you Shengjia!

…

On Wed, Jul 21, 2021 at 1:33 AM Shengjia Zhao ***@***.***> wrote: Hi sorry for this late reply. I haven't been up to date on looking at this repo. In the tutorial, we use a simplified implementation without adding random noise. For the version with random noise please see this file https://github.com/ShengjiaZhao/MMD-Variational-Autoencoder/blob/master/mmd_vae_eval.py — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABMKI5AYNUAWAYRTEDJOU53TYZLZJANCNFSM4IGUSQ7Q> .