[PIPE2D-1012] Quality of FluxModelInterpolator's output is bad

Wed, 8 Jun 2022 00:27:40 +0000

The quality of FluxModelInterpolator's output is evidently bad.

[truth vs interpolation.png] FluxModelInterpolator cannot even reproduce one of the input points. Since this is interpolation, such a phenomenon should not be observed if calculations are exact.

[fullset vs modelInterpolator.png]FluxModelInterpolator cannot reproduce the interpolation "fullset") that Yamashita-san made.

According to Yamashita-san, the current FluxModelInterpolator works differently from what his interpolator did because he had forgot to upload his latest code to drp_fstar repository before FluxModelInterpolator was made.

However, I made some experiments and concluded that it was the oldness of scipy that was to blame. In addition, the latest scipy introduced RBFInterpolator that is very useful to FluxModelInterpolator. Currently, FluxModelInterpolator uses Rbf class, whose codomain must be one-dimensional in our old scipy. It uses 100k Rbf instances to output a spectrum with 100k sampling points. On the other hand, RBFInterpolator's codomain can be multidimensional. If it is used, computation will be 100k times faster as I have already confirmed personally.

Because I don't think the latest scipy (>= 1.8) will be available very soon, I am going to write a small implementation of RBFInterpolator compatible with scipy. After scipy >= 1.8 is available, the compatible class must be abandoned.

PFS-JIRA

[PIPE2D-1012] Quality of FluxModelInterpolator's output is bad