Alex Robson

Old thread, but I believe that I'm also hitting a variant of this when deploying in the wild with some heavy tailed data: MWE: ``` rx = [randn(95,); 10^1; 10^2;...

> Is the suggestion that we implement the same hack in construct? (We should probably consider a hack in CRC more carefully than in FD). Agree with this. I wasn't...

Thanks for the quick response - that was helpful iiuc, the code you've linked creates general rrules for a set of matrix operations, ones that would under the hood call...

This should be ready I believe. I could also include a deprecation to remove `mvtdist` because, as you point out, we should be using the constructor. Or possibly we can...

For context, in my case, it was in exploring some time-series data of length O(1000). Regarding the comment on using different geometries, I don't doubt that that in my case...

I havne't looked at this for a while. That said, last I left it, I thought it was ready. I've moved well out of the context of what I was...

FWIW i had a look at this. IIUC, when generating the [training data](https://github.com/SciML/NeuralPDE.jl/blob/104ccc07aec0241be659a7662f15c6ccd8f3b5e2/src/discretize.jl#L248) product is used which doesn't promote, making CUDA unhappy: ``` span = [0, (0.0:0.1:1.0)]; map(points -> collect(points),...

I'm pikcing this back up. Just working through it now. I should have an update shortly.

Looking at this again, I recall the reason I hit a block on it, and why it shouldn't go in as-is. There are two reasons for this, one major and...

> This may be helpful: if you set `primary = false`, a new series will take the color of the previous series. And if `label = ""`, it's excluded from...