Train with
anaconda-project run train
Demo with
anaconda-project run demo
IMPORTANT:
- Now regex.Alt has a continuous parameter. Need to pay for it in the score (either AIC or something variational)
Training/Proposals:
- Evaluate multiple proposals in parallel
- Keep training NN while evaluating proposals
- Don't add the same proposal multiple times (after counterexamples)
- Repropose on existing concept
- Given a valid inferred regex, parse examples with that regex, choose a matching subgroup, and run inference on that parse
Library:
- When removing library item, delete other redundant items that are only referenced by it
Trace:
- Pitman-Yor Process
- Move state (other than observations) into concept itself?
- Clean up CRPRegexConcept. Currently CRPconcept has no prior on whether it has regex or CRP inside (and concepts not in baseConcepts don't get scored in addConcept)
- Rich concepts get richer (DP?)
- Move all model logic out of trace.py
- Make it clearer which functions mutate, and also when a returned trace is clean/dirty
- concept.py?
- Refactor CRPConcept (shared functionality between observe and observe_partial)
- CRPConcept: If there are existing tables with some value, sample rather than choosing the first?
- ErrorConcept to allow errors?
- Replace TempDict and TempList with https://pythonhosted.org/pysistence/