Experiments resource-based interface

[acroz]

This PR prototypes adding a resource-based interface for experiment runs, with the following features:

• New primary models using attrs with class methods for querying / getting resources
• Convert client models to 'resource interface' models
• A query return type that behaves as an iterable but can provide extra methods like as_dataframe for other view types

Notes:

• I discussed with @zblz about using dataclasses but we agreed that attrs was a reasonable substitute allowing us to maintain Python 2.7 support.
• I've not done anything clever with any of:
  • Caching the session / client
  • Looking up the project ID when not set
• Only part of the run objects are currently mapped to df columns
• What are the columns in the output df called when metrics / params / tags have conflicting names?

[acroz]

Example usage:

from faculty.experiments import ExperimentRun

print(
    list(
        ExperimentRun.query(project_id="f20f5eaa-9cff-4216-8ebd-b88fd294bb3e")
    )
)

print(
    ExperimentRun.query(
        project_id="f20f5eaa-9cff-4216-8ebd-b88fd294bb3e"
    ).as_dataframe()
)

[eliasbenussi]

• Caching the session / client

I talked to @zblz about this. In faculty.client we call faculty.session.get_session which already handles caching of the session. I think it's not worth the effort to implement caching just for the client, so I would keep the implementation as is?

• Looking up the project ID when not set

Sorry my confusion; by this you mean using the PROJECT_ID env variable by default if an argument is not passed?

• What are the columns in the output df called when metrics / params / tags have conflicting names?

I am not very experienced with pandas, my first thought was to namespace keys by metrics. / params. / tags., but I don't know if this would be a nuisance in practice?

[eliasbenussi]

The checks are not passing because apparently yield from in faculty/experiments.py (line 62) is not valid Python 2 syntax. I am not really sure how to address this?