Train a model on some intermediate rank

[sylvain-morin]

iNat problem: Sometimes the common ancestor isn't super useful. Can we make it better?

Experiment Idea: Train a model on some intermediate rank like family.

Evaluation notes: Compare the predictions from this intermediate rank model to the family prediction suggested by rolling up from leaf predictions in the current production model.

[alexshepard]

When I described this in the AWS experiment ideas doc in late 2021, I was thinking of this as a computer vision issue. Should we try to map this into the current geo model or is it still useful to think of this as an improvement to the computer vision model?

[sylvain-morin]

• [x] Write a script to prepare the datasets by adding the Family ID / Family Leaf Class ID
• [x] Run this script on the 2.12 datasets to get Family 2.12 datasets (remove observations that were identified ad higher rank than Family)
• [x] Run download_photo scripts for these train and val datasets, on bluey
• [x] Run make_clean scripts for these datasets
• [x] Run the training...
• [ ] Adapt the test model to support this Family Leaf Class ID
• [x] Convert the trained model (.pb files) into vision.h5
• [ ] Run the Family test model script on the 2.12 model and the Family trained model

[sylvain-morin]

EXPERIMENT 1 I'm using the 2.12 training datasets, and adding the Family information to each selected observations. This observation selection was done at the species level, using the species threshold.

EXPERIMENT 2 Generate a new training dataset, including observations from any taxon of a family, without limiting to the "species threshold".