Poor Pipeline Performance

[bitsofbits]

Completed

• Downloaded 5 new tracks from BigQuery. One turned out to be a duplicate, so discarded it.
  • Uploaded remaining tracks to GoogleDrive
• Ran these tracks against model trained in compare single model.
  • Tracks still had low False positive rate (~4%)
  • False positives were clumped, not clustered, so shouldn't have appearance of dotted tracks.
• Ran against model parameters in models directory
  • Parameters that I generated locally using train_models
  • Worked identically
  • Parameters that I downloaded from github
  • * 90% false positive rate**

Implication

Model parameters on github are broken

To Do

• Download more tracks for testing
• Train again after fixing Kristina's data

[bitsofbits]

@redhog can we close this now?

[redhog]

Yes