Fork me on GitHub
#uncomplicate
<
2017-03-22
>
jholkeboer17:03:30

Overall it is great. If anything is holding me back it is not Bayadera but rather the fact that I still have much to learn about Bayesian statistics.

jholkeboer17:03:20

And maybe even dynamically adding parameters to the OpenCL kernel. I was able to add an additional parameter manually and I think this could be done by treating the whole file as a string

blueberry18:03:53

@jholkeboer The whole OpenCL code is treated as a string already. Most Bayadera methods related to building cl models just help in creating the model without writing (and optimizing!) the whole file manually. One thing that may not be obvious to users is that kernels that execute on the GPU are very fragile things: introduce too much conditionals, and the performance tanks. Dynamicity is something to be avoided on the GPU. The host code is open game 🙂 Another thing that is obvious is that dynamic code may not be a good match for the MCMC algorithm itself...

jholkeboer18:03:23

Yes i definitely will not be adding ifs and loops in the kernel. Just b3, b4... in the multiple linear regression

jholkeboer18:03:12

and only ahead of time. as if I had typed it in. This is more about building a workflow for me iterating my models quickly

jholkeboer18:03:54

to have more than two MLR predictors

blueberry18:03:06

@jholkeboer Cool. Please write about your experience and share the code. I would love to add such functionality if it works well.

blueberry18:03:10

Bayadera exactly is a low-level engine that should enable you to build such convenient high-level tools on top of it!