Optimizer¶

Optimizer()

The base class of optimizers: subclass it to write your own.

An optimizer has one method, compile(program, *, trainset, valset=None): given an AI function or a @module and rows with known answers (a list of dicts), it returns the new state of each AI function it improves, as {fn: ProgramState(instructions=..., demos=...)}. It never changes the functions itself: fn.opt(rows, optimizer=MyOptimizer()) builds the improved copy from what it returns. metric (metric(row, prediction)) is set from fn.opt(metric=...) when the optimizer has none.

class FirstRows(functai.Optimizer):          # the first three rows become worked examples
    def compile(self, program, *, trainset, valset=None):
        demos = tuple({"inputs": {"message": r["message"]}, "outputs": {"result": r["team"]}}
                      for r in trainset[:3])
        return {program: functai.ProgramState(demos=demos)}

better = team.opt(rows, optimizer=FirstRows())