def from_config(cls, config: Config) -> MonotonicRejectionGP:
classname = cls.__name__
num_induc = config.gettensor(classname, "num_induc", fallback=25)
num_samples = config.gettensor(classname, "num_samples", fallback=250)
num_rejection_samples = config.getint(classname,
"num_rejection_samples",
fallback=5000)
lb = config.gettensor(classname, "lb")
ub = config.gettensor(classname, "ub")
dim = config.getint(classname, "dim", fallback=None)
mean_covar_factory = config.getobj(
classname,
"mean_covar_factory",
fallback=monotonic_mean_covar_factory)
mean, covar = mean_covar_factory(config)
monotonic_idxs: List[int] = config.getlist(classname,
"monotonic_idxs",
fallback=[-1])
return cls(
monotonic_idxs=monotonic_idxs,
lb=lb,
ub=ub,
dim=dim,
num_induc=num_induc,
num_samples=num_samples,
num_rejection_samples=num_rejection_samples,
mean_module=mean,
covar_module=covar,
)
def from_config(cls, config: Config):
classname = cls.__name__
lb = config.gettensor(classname, "lb")
ub = config.gettensor(classname, "ub")
dim = config.getint(classname, "dim", fallback=None)
seed = config.getint(classname, "seed", fallback=None)
return cls(lb=lb, ub=ub, dim=dim, seed=seed)
def from_config(cls, config: Config) -> GPClassificationModel:
"""Alternate constructor for GPClassification model.
This is used when we recursively build a full sampling strategy
from a configuration. TODO: document how this works in some tutorial.
Args:
config (Config): A configuration containing keys/values matching this class
Returns:
GPClassificationModel: Configured class instance.
"""
classname = cls.__name__
inducing_size = config.getint(classname, "inducing_size", fallback=10)
lb = config.gettensor(classname, "lb")
ub = config.gettensor(classname, "ub")
dim = config.getint(classname, "dim", fallback=None)
mean_covar_factory = config.getobj(classname,
"mean_covar_factory",
fallback=default_mean_covar_factory)
mean, covar = mean_covar_factory(config)
max_fit_time = config.getfloat(classname,
"max_fit_time",
fallback=None)
inducing_point_method = config.get(classname,
"inducing_point_method",
fallback="auto")
likelihood_cls = config.getobj(classname, "likelihood", fallback=None)
if likelihood_cls is not None:
if hasattr(likelihood_cls, "from_config"):
likelihood = likelihood_cls.from_config(config)
else:
likelihood = likelihood_cls()
else:
likelihood = None # fall back to __init__ default
return cls(
lb=lb,
ub=ub,
dim=dim,
inducing_size=inducing_size,
mean_module=mean,
covar_module=covar,
max_fit_time=max_fit_time,
inducing_point_method=inducing_point_method,
likelihood=likelihood,
)
def monotonic_mean_covar_factory(
config: Config,
) -> Tuple[ConstantMeanPartialObsGrad, gpytorch.kernels.ScaleKernel]:
"""Default factory for monotonic GP models based on derivative observations.
Args:
config (Config): Config containing (at least) bounds, and optionally LSE target.
Returns:
Tuple[ConstantMeanPartialObsGrad, gpytorch.kernels.ScaleKernel]: Instantiated mean and
scaled RBF kernels with partial derivative observations.
"""
lb = config.gettensor("monotonic_mean_covar_factory", "lb")
ub = config.gettensor("monotonic_mean_covar_factory", "ub")
assert lb.shape[0] == ub.shape[0], "bounds shape mismatch!"
dim = lb.shape[0]
fixed_mean = config.getboolean("monotonic_mean_covar_factory",
"fixed_mean",
fallback=False)
mean = ConstantMeanPartialObsGrad()
if fixed_mean:
try:
target = config.getfloat("monotonic_mean_covar_factory", "target")
mean.constant.requires_grad_(False)
mean.constant.copy_(torch.tensor([norm.ppf(target)]))
except NoOptionError:
raise RuntimeError(
"Config got fixed_mean=True but no target included!")
ls_prior = gpytorch.priors.GammaPrior(
concentration=__default_invgamma_concentration,
rate=__default_invgamma_rate,
transform=lambda x: 1 / x,
)
ls_prior_mode = ls_prior.rate / (ls_prior.concentration + 1)
ls_constraint = gpytorch.constraints.Positive(transform=None,
initial_value=ls_prior_mode)
covar = gpytorch.kernels.ScaleKernel(
RBFKernelPartialObsGrad(
lengthscale_prior=ls_prior,
lengthscale_constraint=ls_constraint,
ard_num_dims=dim,
),
outputscale_prior=gpytorch.priors.SmoothedBoxPrior(a=1, b=4),
)
return mean, covar
def from_config(cls, config: Config, name: str):
gen_cls = config.getobj(name, "generator", fallback=SobolGenerator)
generator = gen_cls.from_config(config)
model_cls = config.getobj(name, "model", fallback=None)
if model_cls is not None:
model = model_cls.from_config(config)
else:
model = None
acqf_cls = config.getobj(name, "acqf", fallback=None)
if acqf_cls is not None and hasattr(generator, "acqf"):
if generator.acqf is None:
generator.acqf = acqf_cls
generator.acqf_kwargs = generator._get_acqf_options(
acqf_cls, config)
n_trials = config.getint(name, "n_trials")
refit_every = config.getint(name, "refit_every", fallback=1)
lb = config.gettensor(name, "lb")
ub = config.gettensor(name, "ub")
dim = config.getint(name, "dim", fallback=None)
outcome_type = config.get(name,
"outcome_type",
fallback="single_probit")
if model is not None and not generator._requires_model:
if refit_every < n_trials:
warnings.warn(
f"Strategy '{name}' has refit_every < n_trials even though its generator does not require a model. Consider making refit_every = n_trials to speed up point generation.",
RuntimeWarning,
)
return cls(
lb=lb,
ub=ub,
dim=dim,
model=model,
generator=generator,
n_trials=n_trials,
refit_every=refit_every,
outcome_type=outcome_type,
)
def default_mean_covar_factory(
config: Config,
) -> Tuple[gpytorch.means.ConstantMean, gpytorch.kernels.ScaleKernel]:
"""Default factory for generic GP models
Args:
config (Config): Object containing bounds (and potentially other
config details).
Returns:
Tuple[gpytorch.means.Mean, gpytorch.kernels.Kernel]: Instantiated
ConstantMean and ScaleKernel with priors based on bounds.
"""
lb = config.gettensor("default_mean_covar_factory", "lb")
ub = config.gettensor("default_mean_covar_factory", "ub")
fixed_mean = config.getboolean("default_mean_covar_factory",
"fixed_mean",
fallback=False)
lengthscale_prior = config.get("default_mean_covar_factory",
"lengthscale_prior",
fallback="gamma")
outputscale_prior = config.get("default_mean_covar_factory",
"outputscale_prior",
fallback="box")
kernel = config.getobj("default_mean_covar_factory",
"kernel",
fallback=gpytorch.kernels.RBFKernel)
assert lb.shape[0] == ub.shape[0], "bounds shape mismatch!"
dim = lb.shape[0]
mean = gpytorch.means.ConstantMean()
if fixed_mean:
try:
target = config.getfloat("default_mean_covar_factory", "target")
mean.constant.requires_grad_(False)
mean.constant.copy_(torch.tensor([norm.ppf(target)]))
except NoOptionError:
raise RuntimeError(
"Config got fixed_mean=True but no target included!")
if lengthscale_prior == "invgamma":
ls_prior = gpytorch.priors.GammaPrior(
concentration=__default_invgamma_concentration,
rate=__default_invgamma_rate,
transform=lambda x: 1 / x,
)
ls_prior_mode = ls_prior.rate / (ls_prior.concentration + 1)
elif lengthscale_prior == "gamma":
ls_prior = gpytorch.priors.GammaPrior(concentration=3.0, rate=6.0)
ls_prior_mode = (ls_prior.concentration - 1) / ls_prior.rate
else:
raise RuntimeError(
f"Lengthscale_prior should be invgamma or gamma, got {lengthscale_prior}"
)
if outputscale_prior == "gamma":
os_prior = gpytorch.priors.GammaPrior(concentration=2.0, rate=0.15)
elif outputscale_prior == "box":
os_prior = gpytorch.priors.SmoothedBoxPrior(a=1, b=4)
else:
raise RuntimeError(
f"Outputscale_prior should be gamma or box, got {outputscale_prior}"
)
ls_constraint = gpytorch.constraints.Positive(transform=None,
initial_value=ls_prior_mode)
covar = gpytorch.kernels.ScaleKernel(
kernel(
lengthscale_prior=ls_prior,
lengthscale_constraint=ls_constraint,
ard_num_dims=dim,
),
outputscale_prior=os_prior,
)
return mean, covar
def song_mean_covar_factory(
config: Config,
) -> Tuple[gpytorch.means.ConstantMean, gpytorch.kernels.AdditiveKernel]:
"""
Factory that makes kernels like Song et al. 2018:
Linear in intensity dimension (assumed to be the last
dimension), RBF in context dimensions, summed.
Args:
config (Config): Config object containing (at least) bounds and optionally
LSE target.
Returns:
Tuple[gpytorch.means.ConstantMean, gpytorch.kernels.AdditiveKernel]: Instantiated
constant mean object and additive kernel object.
"""
lb = config.gettensor("song_mean_covar_factory", "lb")
ub = config.gettensor("song_mean_covar_factory", "ub")
assert lb.shape[0] == ub.shape[0], "bounds shape mismatch!"
dim = lb.shape[0]
mean = gpytorch.means.ConstantMean()
try:
target = config.getfloat("song_mean_covar_factory", "target")
except NoOptionError:
target = 0.75
mean.constant.requires_grad_(False)
mean.constant.copy_(torch.tensor([norm.ppf(target)]))
ls_prior = gpytorch.priors.GammaPrior(
concentration=__default_invgamma_concentration,
rate=__default_invgamma_rate,
transform=lambda x: 1 / x,
)
ls_prior_mode = ls_prior.rate / (ls_prior.concentration + 1)
ls_constraint = gpytorch.constraints.Positive(transform=None,
initial_value=ls_prior_mode)
stim_dim = config.getint("song_mean_covar_factory",
"stim_dim",
fallback=-1)
context_dims = list(range(dim))
stim_dim = context_dims.pop(stim_dim) # support relative stim dims
if dim == 1:
# this can just be LinearKernel but for consistency of interface
# we make it additive with one module
return (
mean,
gpytorch.kernels.AdditiveKernel(
gpytorch.kernels.ScaleKernel(
gpytorch.kernels.LinearKernel(ard_num_dims=1),
outputscale_prior=gpytorch.priors.SmoothedBoxPrior(a=1,
b=4),
)),
)
else:
context_covar = gpytorch.kernels.ScaleKernel(
gpytorch.kernels.RBFKernel(
lengthscale_prior=ls_prior,
lengthscale_constraint=ls_constraint,
ard_num_dims=dim - 1,
active_dims=context_dims,
),
outputscale_prior=gpytorch.priors.SmoothedBoxPrior(a=1, b=4),
)
intensity_covar = gpytorch.kernels.ScaleKernel(
gpytorch.kernels.LinearKernel(active_dims=stim_dim,
ard_num_dims=1),
outputscale_prior=gpytorch.priors.SmoothedBoxPrior(a=1, b=4),
)
return mean, context_covar + intensity_covar