Organizational metadata for experiments

[bfirsh]

Why

You can add params to experiments when starting them, but you might also want to add metadata after the fact to annotate them with. For example:

• "bad" for experiments that did not perform well, but you want to keep the information around so you know you tried it
• "interesting" if you want to include them in the paper
• "converged" "failed to converge" depending on if they converged or not
• "bad-labels" if you messed up somehow
• "production" to mark which experiment/checkpoint is deployed to production. (Though maybe this is more like a singleton branch-like pointer? It doesn't make much sense for 10 experiments to be marked as "production")

This feels like a tag in Flickr, or perhaps a label in Docker (although they are key/value).

How are these different to params?

• Param is something that changes how the model behaves — fixed inputs to start of model
• A tag is metadata that I add for my own organization — can be modified

Strawman

Related

• #297