Improve the training time

[OmkarEllicium]

Hi @gufengzhou ,
I am trying to train the model with data spread across ~17 states. Approx record count is around 40k.
For each state, robyn takes around 21 minutes to run with 2000 iterations and 5 trials.

Is there any way we can reduce the time taken to run by keeping the same parameters mentioned above?
I have a requirement to run it across a few hundred regions hence the question.

Any help would be appreciated.

Thank you

[gufengzhou]

You could speed things up by using more cores in robyn_run. Otherwise it'd be a standalone effort to focus on speed improvement that's unfortunately not on the current plan due to resource.

[OmkarEllicium]

I tried increasing cores from 8 cores to 32 cores. But the speed is more or less the same

[benshush]

Hey @OmkarEllicium , @gufengzhou
I just submitted PR #1217, which optimizes _geometric_adstock reducing overall model training time by ~40% (tested on tutorial1.ipynb with Trial=1, iterations=2000).

Since _geometric_adstock was a major bottleneck, this should help speed up training, especially for large datasets. Would love your thoughts!

@gufengzhou , since you’re assigned to this issue, I wanted to bring this to your attention in case it's relevant.