plots.tgz
  Each subdir is a "label", which generally corresponds to a plot or two in the paper.
  Within each label\'s subdir, there are single-metric subdirs (e.g. lbi, cons-dist-aa), which are like the scan plots in the paper, but since they only have one metric, they can show all the different parameter values at once, i.e. several different lines for each metric).
  So e.g. Fig 1 shows performance vs obs time for one carry cap for lots of metrics, but in this dir (in vary-obs-times-v5/) we have the same but showing carry cap of 350, 1000, and 3000 for each metric by itself.
  There are also subdirs of each label with combined-XXX, where combined-with-aa-lb/ is usually the one in the figures in the paper, and combined-with-dtr-and-aa-lb/ is the same, but also with the dtr.
  The one exception is cons-dist-accuracy/, since these plots are showing a totally different thing; these show the plots in S7 Figure, but for additional seeds, obs times, carry caps, and min-target-distance values.
simulation-samples.tgz (not in github cause theyre huge, were copied with cp-simulation-samples.sh)
  This contains all simulated samples that were used in the paper.
  To figure out which samples go with which plots, look for the plot you want in plots.tgz (the subdir name should help you guess); they have the same names in here.
  All samples are in partis yaml format https://github.com/psathyrella/partis/blob/master/docs/output-formats.md.
  For details on how these were processed to get to the plots, or for any of the processed outputs (theyre way too much to include here), please contact the authors.
dtr-scan-logs.tgz
  Text logs of dtr performance for different dtr training parameter values (made with test/dtr-scan.py)
  Several different versions of parameter distributions were made, corresponding to different dtr training samples: v0-v3.
  Samples are generated in sets of 2-5 with different random seeds; training was always on seed 0, while the other seeds are for performance evaluation.
  dtr-scan-all-seeds/
    performance for all samples: this compares performance on the training sample (seed 0) to other samples that both differ only by random seed (same vN, different seeds), and by different parameter distributions (different vN).
  dtr-scan/
    same, but shows mean +/- std err only for non-training samples, i.e. omits seed 0 for the training version
data.tgz
  Partis parameter and output files (which include all input sequences and affinity values) for the real data samples from Landais 2017 and Wu 2011.