Jarvis Reproduce

[rlnsanz]

Reproduce an experiment or trial by materializing it on a user-specified directory. The execution of a python script enables the data scientist to re-run some experiment, or some trial within the experiment.

[rlnsanz]

• Experiment artifacts and version information, data context exist in a hidden location. User can “materialize” these contents to facilitate exploration, but changes to these contents do not update any artifacts.
• Trials replaced with constants. Trials don’t look like general experiment. Iteration semantics
  • Pick one aggregates (min/max)
  • Sweep vs. for each.
    • Alpha can have one value
    • For each value of alpha, I may want to export a trial that iterates over every value of beta. This needs to be expressed somehow.

[malharpatel]

[dcrankshaw]

Glancing through those notes on the board, I noticed it says that if we re-run an experiment it does not get versioned. What's the reasoning behind that? I think we should be very careful about deciding to not version things.

[rlnsanz]

What we meant is that if you reproduce an experiment it does not get versioned. Re-running an experiment always gets versioned.

The argument was that if you checkout some past experiment-run to some target directory, move to that directory, and call python reproduce.py -- if the run-reproduction is a "true" one -- then the results would already have been versioned, and don't need to be versioned again.

That said, it's unlikely that the experiment-run reproduction will be a "true" one, sources of randomness could alter some result, and we may have good reasons for tracking how many times an experiment has been reproduced.

Still... what it means to version something well when you re-run an experiment vs. when you reproduce a past one are something that we wanted to call to the attention of the group, and something we could discuss in meeting.