model metadata

[dbuscombe-usgs]

we need better ways to keep track of individual trained models ... what data they were used on, what hyperparameters, etc

[option 1] a bespoke option
--> config
--> history
--> metrics
json files that can be appended, then parsed

[option 2] conform to a standard e.g. W&B or MLFlow

what it needs to contain

• store info on what training data was used to train a model
• use unique IDs for trained model weights
• store all values from config file with model weights?
• store metrics?
• during implementation, keep track of model implementation choices

what type of database?

[ebgoldstein]

is this like MLFlow or Weights & Biases?

[dbuscombe-usgs]

could be

[dbuscombe-usgs]

First implementation in bf1962f

[dbuscombe-usgs]

Implement Hugginface-style model cards see #41