Error with training of model for Jura dataset

[mc19952004]

Hi,

I have been working through the demo code for the Jura dataset as it is a similar structure to my own data. I am able to run the code and create the initial model, but discover an error when training the model using 'multigpOptimise'. The error appears to relate to the computation of the model's gradients w.r.t log likelihood using 'modelLogLikeGradients'. The issue seems to be that the dimension of the gradients, 'g', and the dimension of the 'model.paramGroups' are not compatible for matrix multiplication as 'g' only contains 17 columns in comparison to the 20 rows withing 'model.paramGroups'.

The same issue occurs when using my own data. Are you aware of any way that this issue can be fixed?

Thanks

[lawrennd]

HI Michael ... haven't dug into this particular problem. But I suspect it's a problem/incompatibility to how paramGroups is being set.

One thing that can help is to provide a minimal example that recreates the problem, there's a nice description on this here:

https://stackoverflow.com/help/minimal-reproducible-example

My recollection would be that paramGroups is being used for parameter tieing ... so that one parameter is being shared multiple places in the covariance function. With luck that might give some clues on to what's gone wrong. Are you doing parameter tieing in your own code?

The other trick might be to look for changes in code over the last 10 or 15 years ... that might have introduced the problem. As I think the demo should have been working when we published. The alternative is that something we were relying on in matlab may have changed.