Chains are not converging

[devin30000]

Hi all,

First i ran my model on a cpu, which took 7 hours (10 chains), but everything worked out well. Now I am using a T4 GPU and when I run "model_diagnostics.plot_rhat_boxplot()" I always get "MCMCSamplingError: MCMC sampling failed with a maximum R-hat value of 1273814037504.0" or something like this. So my chains do not converge anymore.
I have no NaN values and i have 6 channels and 25 geos and 2 control variables for now. some channels have many weeks without spending. I have als tried to set the chains with [1,1,1,1,1]. Then the R-hat values are around 1-1.05, but it takes too long.
Possible reasons?
Thank you for helping me

[dattatreyam23]

Hi @devin30000,

Thank you for contacting us!

The behaviour you mentioned is unexpected. Please also check our documentation on Debugging MCMC Convergence issues which lists some probable causes for the same and the respective debugging steps that may help resolve this problem.

Since the issue is not replicable using the demo dataset, to help me understand the issue better, would it be possible for you to share a bit more about your dataset and whether you're using holdout sets? If you could try to reproduce the error with a set seed (sample_prior and sample_posterior both accept this), that would be incredibly helpful for debugging. Any code snippet you're able to share would also be greatly appreciated.

Hope the debugging documentation would resolve your issue. If the issue persists, feel free to reach out for further assistance, along with the requested information if possible.

Thank you,

Google Meridian Support Team

[devin30000]

It turned out there was an error in our data, so everyhting works out fine now, thank you for the help anyways :)