cballand / desibest Goto Github PK

A fit with zero error leads to the following warning:

desibest/bin/parsezbest.py:281: RuntimeWarning: divide by zero encountered in divide
  pull =(bz-tz)/ez

For running on NERSC or running over all target classes, it would be useful to have a command-line option that suppresses the display of all plots (which then have to be individually closed). Instead, they could be automatically saved to disk (or not produced at all).

parsezbest will refuse to calculate statistics if the zbest file is missing the BRICKNAME and SUBTYPE columns, even though it doesn't use them. It would be better to just print a warning telling the user to fix that, but proceed with the calculations anyway.

Similarly, if EXTNAME='ZBEST' doesn't exist, print a warning and proceed with trying to use whatever is in HDU 1.

Precision and accuracy are estimated using the standard deviation and mean:

            # precision
            ok = np.where(zw==0)[0]
            zerr = np.std(dz[ok])
            verr = np.std(dv[ok])

            #accuracy
            zacc = np.abs(np.mean(dz[ok]))
            vacc = np.abs(np.mean(dv[ok]))

These are not robust estimators when the distribution of residuals is non-Gaussian, especially in the presence of long tails.

I recommend using quantiles instead (or as an alternative), e.g.

    dv_quant = np.percentile(dz[ok], (2.5, 16, 50, 84, 97.5))
    dv_err68 = dv_quant[3] - dv_quant[1]
    dv_err95 = dv_quant[4] - dv_quant[0]
    dv_acc50 = dv_quant[2]