Closed jeffjennings closed 9 months ago
Example of a split with the new implementation -- none of the splits have red points at the smallest (u,v) locations (but they do still have points at >530 klambda, i.e., including where the weighted visibility distribution peaks):
This figure omits the data (missing in the circle around the origin) that are being given to every train set:
Weighted visibility distribution for reference. The train sets are all getting the cells in the smallest dartboard bin (0 - 530 klambda in this case):
Example of the loss evolution, now without an outlier k-fold:
Alters
crossval.DartboardSplitGridded
to give every train split the cells belonging to the smallest baseline bin in the dartboard. Withholds these cells from all test sets.Sets the default
Dartboard
inCrossValidate
to setq
cells based on the max baseline in the data, not the max of the Fourier grid (which can be quite a bit larger). This ensures the points that every train cell gets aren't too large a fraction of the total vis distribution. I've set theq
bins to be the same as invis_histogram_fig
so that the figure made by that corresponds to the splits in a cross-val loop by default (also added args to pass inq
andphi
bins toCrossValidate
).Updates the figure produced by
plot.split_diagnostics_fig
to look a bit better