OOD: Document the workflow that will be used #748

Activity

1. f-PLT

   added this to the OOD Integration milestone on Feb 27, 2025
2. f-PLT

   assigned

   Yuyan-C

   and

   mihow

   on Feb 27, 2025
3. f-PLT

   added a parent issue on Feb 27, 2025

   • Get antenna to work locally on laptops for Panama trip #747
4. f-PLT

   removed a parent issue on Feb 27, 2025

   • Get antenna to work locally on laptops for Panama trip #747
5. mihow

   changed the title [-]Document the workflow that will be used[/-] [+]OOD: Document the workflow that will be used[/+] on Mar 10, 2025
6. mihow commented on Mar 14, 2025

   

   mihow

   on Mar 14, 2025

   Collaborator

   More actions

   @Yuyan-C can you make a short summary of the workflow you/we proposed before the dev meeting next tuesday? I know you have a more detailed one in the works.
7. mihow commented on Mar 24, 2025

   

   mihow

   on Mar 24, 2025

   Last edited by mihow

   Collaborator

   More actions

   Here are the dev steps & requirments for implementing the software tools we need

   1. Decide on classification model for Panama species - 2 new options with better OOD results and expanded checklist. Likely the new model with ~2,360 species (Resnet50). Called "Panama Plus Species Classifier - Mar 2025"

   • TODO:
   • • Get familiar with the Data Companion, and test PG Vector query methods
   • • • Software to install DBeaver, Docker for Mac
   • • Add new Panama Plus classifier to AMI Data Companion (@Yuyan-C)
   • • Add new output field to API response for additional score(s)

   2. New column / field for "OOD Score" (either calculated from logits or existing softmax score). (Current score in Antenna is Softmax Score). Can score names in Antenna be dynamic? We can hardcode for now, but each model will likely want to add a definition of it's calibrated confidence score.
   3. If occurrences OOD Score below threshold - Unknown Species (catch-all)
   4. Decide on / Test 2 approaches. Start with vector database locally, start with pgvector query of current ADC database
   5. Cluster all "Unknown Species" to Unknown Species n01, n02, etc. (based on feature vector of image classifier)
8. mihow

   closed this as completedon Mar 25, 2025