experiment management

[EdAyers]

Right,

@experiment
def foo(x : int, y : float):

   ... long running computation
   checkpoint = snapshot('./out/checkpoint1')
   ... long running
   return {checkpoint, result}

This is treated the same as @memo except an 'experiment' flag is set.

Experiments are treated differently in that

• they appear in their own visualisation on the website
• they are not deleted like memo results
• (to design) logs are streamed and captured
• (to design) other artefacts are captured and streamed up
• (to design) you can annotate them after the fact (eg "this evaluation is the best one")
• (eventually) you can do bayesian optimisation on parameters