Dan Foreman-Mackey
Dan Foreman-Mackey
Maybe we need a celerite 2.0 paper?
Thanks! This was an incompatibility introduced in the most recent PyMC/PyTensor release. Once https://github.com/exoplanet-dev/celerite2/pull/94 is merged, it should work again.
I expect that this is happening because 3.10 wheels are available on PyPI, but they aren't for 3.11. (Although I'm not sure why the source build is failing!) We'll have...
There is no GPU support planned for celerite2. It's possible to parallelize some of the algorithms but it's slower than the CPU version for all the tests I've done and...
I don't know of any good JAX libraries for GPs, but it's not too hard to implement the math yourself to try it out. If the GP is your bottleneck,...
@tagordon — @bmorris3 is right, but there are some options. In the (released) version of celerite2 that you're using, the suggested approach is to directly use a `OverdampedSHOTerm` or `UnderdampedSHOTerm`,...
@ericagol: The catch here is actually that K1 will be dense (in the celerite sense), but then K2 will be block diagonal (with celerite kernels in each block). The idea...
So - I worked out how to do this. Basically, if you want a kernel with the form: ``` k(tn, tm) = ( a(tn) * a_prime(tm) * cos(d * (tn...
I expect this has more to do with how you're setting up the model when you restart - rather than anything to do with loading the chain. How are you...
I can look more closely later, but what if you use the globally defined `gp` instead of passing it as an argument? I can see a problem here where, even...