Extend tests for GPEmulator.jl and Observations.jl #62
Conversation
| @@ -117,12 +117,12 @@ function GPObj(inputs, data, package::GPJL; GPkernel::Union{K, KPy, Nothing}=not | |||
| GPkernel_i = deepcopy(GPkernel) | |||
| # inputs: N_samples x N_parameters | |||
| # data: N_samples x N_data | |||
| logstd_obs_noise = log(sqrt(0.5)) # log standard dev of obs noise | |||
| logstd_obs_noise = log(sqrt(0.01)) # log standard dev of obs noise | |||
There was a problem hiding this comment.
Should we make this dependent on the ensemble covariance? I also think that Ollie wanted to base this on the SVD stuff.
There was a problem hiding this comment.
Yes, we should! This will come in another (probably the next) PR though - the purpose of this one is merely to extend the test coverage. The change of the hard-coded regularization noise from log(sqrt(0.5)) to log(sqrt(0.01)) is just there because I experimented with it after our discussion and thought I could as well leave it at a smaller value until we have implemented the SVD (since the idea is that the regularization noise should be smaller than the learned white noise, a criterion that is more likely to be met the smaller the regularization noise is - speaking of which, one of the next PRs should also address the issue of naming all these noises (white noise, observational noise, regularization noise etc.) in a consistent way that minimizes confusion as much as possible).
There was a problem hiding this comment.
Sounds good. And yes, I think with all the covariances involved, being precise in the notation would be helpful.
Update src/Observations.jl Co-authored-by: Charles Kawczynski <kawczynski.charles@gmail.com>
bielim
force-pushed
the
extend_unit_tests
branch
from
4237680
to
a163957
Compare
| @@ -117,12 +117,12 @@ function GPObj(inputs, data, package::GPJL; GPkernel::Union{K, KPy, Nothing}=not | |||
| GPkernel_i = deepcopy(GPkernel) | |||
| # inputs: N_samples x N_parameters | |||
| # data: N_samples x N_data | |||
| logstd_obs_noise = log(sqrt(0.5)) # log standard dev of obs noise | |||
| logstd_obs_noise = log(sqrt(0.01)) # log standard dev of obs noise | |||
There was a problem hiding this comment.
Sounds good. And yes, I think with all the covariances involved, being precise in the notation would be helpful.
| # Zero mean function | ||
| kmean = MeanZero() | ||
| m = GPE(inputs', dropdims(data[:, i]', dims=1), kmean, GPkernel_i, | ||
| logstd_obs_noise) | ||
| optimize!(m, noise=false) | ||
| optimize!(m, noise=true) |
There was a problem hiding this comment.
FYI this is the default of optimize!. I will try to read the documentation of GaussianProcesses.jl to see what would be the best setting for our purposes.
|
bors r+ |
|
Build succeeded: |
This PR extends the coverage of the tests for the modules
GPEmulator.jlandObservations.jl.