# Why train VAE twice

**URL:** <https://discourse.scverse.org/t/why-train-vae-twice/48>\
**Category:** scvi-tools\
**Tags:** scvi\
**Created:** [March 17, 2021, 5:39am UTC](https://discourse.scverse.org/t/why-train-vae-twice/48 "2021-03-17T05:39:05Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![lu7777](https://yyz1.discourse-cdn.com/flex035/user_avatar/discourse.scverse.org/lu7777/32/26_2.png) [@lu7777](https://discourse.scverse.org/u/lu7777)\
**Post date:** [March 17, 2021, 5:39am UTC](https://discourse.scverse.org/t/why-train-vae-twice/48/1 "2021-03-17T05:39:05Z")

</div>

Hi，  
Thank you for your excellent tutorial.

I read the article and code. It seems you split the raw data into source data and target data. I have a question why train VAE separately with respect to different data?

Best regards,  
Lu

---

<div class="post-metadata">

**Author:** ![adamgayoso](https://yyz1.discourse-cdn.com/flex035/user_avatar/discourse.scverse.org/adamgayoso/32/100_2.png) [@adamgayoso](https://discourse.scverse.org/u/adamgayoso)\
**Post date:** [August 8, 2022, 7:08am UTC](https://discourse.scverse.org/t/why-train-vae-twice/48/2 "2022-08-08T07:08:24Z")

</div>

Can you point to the tutorial? I believe there we are showing off a feature of how you can fine-tune models for new data that is obtained.