Diagnostics for Fit Results

[annayqho]

Current Diagnostics:

• Warning if the answer is more than 2-sig away from the mean & variance of training labels
• Plot of the leading coefficients & scatter as a function of wavelength, 2 panels, one showing 1 +/- 0.05
• triangle plot for the leading coefficients
• triangle plot for the N test labels
• 1-to-1 plots of the input and output labels + histogram of the difference
• Histogram of the chi-sq of the fits, w/ a dashed line corresponding to DOF
• Overplot each Cannon best-fit spectrum with the original spectrum.
• Stack best-fit spectral residuals and sort by each of the labels. (colorbar goes from -3 to +3 sig) + histogram of RMS at each pixel and each star
• Autocorrelation of residual, incl. blown-up version of central peak

  Modifications:

• Make zoomed versions of the best-fit vs. original flux plot
• Error bars on the output labels. @HWRix has homework to tell us whether the diagonal elements are true error bars :)
• Error bars on the output flux values are fishy... (?)
• triangle.py smoothing (I don't think that triangle.py has a built-in smoothing function. Not sure what a good, robust way to do smoothing is. @HWRix ?)

[davidwhogg]

Choose some places (wavelengths) where quadratic model looks good and where it looks bad and plot the residuals away from the quadratic model as a function of the label values. That would be useful as a starting point.