def _select_optimizer_cls()

in nevergrad/optimization/optimizerlib.py [0:0]

• 71 lines of code
• 39 McCabe index (conditional complexity)

    def _select_optimizer_cls(self) -> base.OptCls:
        if self.fully_continuous and self.has_noise:  # In particular for neuro-DPS.
            DeterministicMix = ConfPortfolio(optimizers=[DiagonalCMA, PSO, GeneticDE])
            return Chaining([DeterministicMix, OptimisticNoisyOnePlusOne], ["half"])

        cma_vars = max(1, 4 + int(3 * np.log(self.dimension)))
        num36 = (
            1 + int(np.sqrt(4.0 * (4 * self.budget) // (self.dimension * 1000)))
            if self.budget is not None
            else 1
        )
        num21 = 1 + (4 * self.budget) // (self.dimension * 1000) if self.budget is not None else 1
        num_dim10 = (
            1 + int(np.sqrt(8.0 * (8 * self.budget) // (self.dimension * 1000)))
            if self.budget is not None
            else 1
        )
        num_dim20 = self.budget // (500 * self.dimension) if self.budget is not None else 1
        para = 1
        if self.budget is not None and self.budget > 5000 * self.dimension:
            para = num36 * cma_vars
        elif self.dimension < 5:
            para = num21 * cma_vars
        elif self.dimension < 10:
            para = num_dim10 * cma_vars
        elif self.dimension < 20:
            para = num_dim20 * cma_vars

        # Special cases in the bounded case
        if (
            self.budget is not None
            and self.budget > 500 * self.dimension
            and self.fully_continuous
            and not self.has_noise
            and self.num_objectives < 2
            and self.num_workers <= para
            and p.helpers.Normalizer(self.parametrization).fully_bounded
        ):
            if (
                self.budget > 5000 * self.dimension
            ):  # Asymptotically let us trust NGOpt36 and its subtle restart.
                return NGOpt36
            if self.dimension < 5:  # Low dimension: let us hit the bounds.
                return NGOpt21
            if self.dimension < 10:  # Moderate dimension: reasonable restart + bet and run.
                num = 1 + int(np.sqrt(8.0 * (8 * self.budget) // (self.dimension * 1000)))
                return ConfPortfolio(optimizers=[NGOpt14] * num, warmup_ratio=0.7)
            if self.dimension < 20:  # Nobody knows why this seems to be so good.
                num = self.budget // (500 * self.dimension)
                return ConfPortfolio(
                    optimizers=[Rescaled(base_optimizer=NGOpt14, scale=1.3 ** i) for i in range(num)],
                    warmup_ratio=0.5,
                )
            if self.num_workers == 1:
                return CmaFmin2
            # We need a special case for dim < 30 ---> let's see later.
            # Otherwise, let us go back to normal life: NGOpt16 which rocks in many cases, possibly Cobyla.
            return NGOpt16
        elif (  # This might be specific of high-precision cases.
            self.budget is not None
            and self.fully_continuous
            and not self.has_noise
            and self.num_objectives < 2
            and self.num_workers <= cma_vars
            and self.budget > 50 * self.dimension
            and p.helpers.Normalizer(self.parametrization).fully_bounded
        ):
            if self.dimension < 3:
                return NGOpt8
            if self.dimension <= 20 and self.num_workers == 1:
                MetaModelFmin2 = ParametrizedMetaModel(multivariate_optimizer=CmaFmin2)
                MetaModelFmin2.no_parallelization = True
                return MetaModelFmin2
            return NGOpt15
        else:
            return super()._select_optimizer_cls()