Track down pipeline problem.

[bitsofbits]

Despite it's name, this branch is mainly about tracking down the pipeline problem.

Look near the end of compare single model for how the models did on the new transits I downloaded. The main takeaway is that they worked fine on newly trained models, and on models that I dumped and reloaded, but failed on the models I loaded from github. It appears that the parameter sets on github are broken somehow.

I put the new npz files, both for transists and for fixed measures at: https://drive.google.com/drive/u/0/folders/0B5lEDe8f8fbcU1pzRHIzUzh6UEk

We should move them eventually, but I didn't want to clobber anything till things settle down.

Take a look at Compare_single_models, particularly the end where I check on the new tracks with both models I trained locally and those I downloaded from github.

My suggestion is to:

* rename your existing model parameters (in the models dir), then download the new branch.

* Download my measure files from GoogleDrive

• Rerun Compare_single_models.

  - The main thing to verify is that newly trained models work and that the old models are failing.
  
  - If that checks out then try the new parameters in the pipeline.
  
  - If not we may have some weird portability problem and I"m not sure what the next step is.

  Here's the commit log for the file that git hub includes (for completeness):

• Improve way Kristina's data is fixed
• Add method to turn CSV transit data
• test how model is performing on new transit data and on previously dumped models.

[redhog]

I can verify that

• Loading the old models from master scores entirely wrong
• Rerunning add_measures from master, then re-generating the models, scores correctly.