Hessian PDF Covariance Matrix for theory predictions#1743
Conversation
scarlehoff
left a comment
scarlehoff
left a comment
There was a problem hiding this comment.
Could you also add a test, an equivalent of
?Also, could you run black on your code? (now it should be completely safe to do so :D)
| return mat | ||
|
|
||
|
|
||
| @check_pdf_is_montecarlo |
There was a problem hiding this comment.
There should still be some check, for the error member types we don't support.
| X = hessian_eigenvectors - central_predictions.reshape((central_predictions.shape[0],1)) | ||
| # need to rescale the Hessian eigenvectors in case the eigenvector confidence interval is not 68% | ||
| X = X / rescale_fac | ||
| pdf_cov = np.einsum("ij,kj->ik",X, X) |
There was a problem hiding this comment.
This could be X@X.T which is clearer, if less efficient.
|
|
||
| # need to subtract the central set which is not the same as the average of the | ||
| # Hessian eigenvectors. | ||
| X = hessian_eigenvectors - central_predictions.reshape((central_predictions.shape[0],1)) |
There was a problem hiding this comment.
This could be something like central_prediction[:, np.newaxis] (or central_prediction[:, None])
|
Hi @Zaharid , @scarlehoff not sure of why |
|
You need to rebase / merge master. I think it is due to numpy 1.24.X, you probably have a different (older) version in your computer. |
comane
force-pushed
the
hessian_pdf_covmat_for_theory_predictions
branch
from
0dd394b to
7efa7fd
Compare
| raise CheckError(f"Error type of PDF {pdf} must be 'replicas' and not {etype}") | ||
|
|
||
|
|
||
| @make_check |
There was a problem hiding this comment.
This should be using make_argcheck afaict.
| def check_pdf_is_montecarlo_or_symmhessian(ns, **kwargs): | ||
| pdf = ns['pdf'] | ||
| etype = pdf.error_type | ||
| if (etype != 'replicas') and (etype != 'symmhessian'): |
There was a problem hiding this comment.
this could be check(etype in {'replicas', 'symhessian'}, f"Error type of PDF {pdf} must be either 'replicas' or 'symmhessian' and not {etype}")
|
I wonder if it makes sense to move this kind of functionality to the stats classes themselves (clue being that we are using plenty of internal methods). Would be a different pr anyway. |
3 participants
The idea of this PR is to have the
covmats.pdferr_plus_covmatfunction to take into account for the case in which the PDF haserror_type = 'symmhessian'