# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# pyre-strict
import itertools
from collections.abc import Callable
from copy import deepcopy
from typing import Any
import numpy.typing as npt
import torch
from ax.adapter.torch import TorchAdapter
from ax.core.data import MAP_KEY
from ax.core.search_space import SearchSpaceDigest
from ax.generators.torch.botorch_modular.generator import BoTorchGenerator
from ax.utils.sensitivity.derivative_measures import (
compute_derivatives_from_model_list,
sample_discrete_parameters,
)
from ax.utils.sensitivity.fixed_feature_model import (
FixedFeatureModel,
prepare_fixed_feature_inputs,
)
from botorch.models.gpytorch import GPyTorchModel
from botorch.models.model import Model
from botorch.sampling.normal import SobolQMCNormalSampler
from botorch.utils.sampling import draw_sobol_samples
from botorch.utils.transforms import unnormalize
from pyre_extensions import none_throws
from torch import Tensor
[docs]
class SobolSensitivity:
def __init__(
self,
bounds: torch.Tensor,
input_function: Callable[[torch.Tensor], torch.Tensor] | None = None,
num_mc_samples: int = 10**4,
input_qmc: bool = False,
second_order: bool = False,
first_order_idcs: torch.Tensor | None = None,
num_bootstrap_samples: int = 1,
bootstrap_array: bool = False,
discrete_features: list[int] | None = None,
) -> None:
r"""Computes three types of Sobol indices:
first order indices, total indices and second order indices (if specified ).
Args:
bounds: Parameter bounds over which to evaluate model sensitivity.
input_function: The objective function.
num_mc_samples: The number of montecarlo grid samples
input_qmc: If True, a qmc Sobol grid is use instead of uniformly random.
second_order: If True, the second order indices are computed.
bootstrap: If true, the MC error is returned.
first_order_idcs: Tensor of previously computed first order indices, where
first_order_idcs.shape = torch.Size([dim]).
num_bootstrap_samples: If bootstrap is true, the number of bootstraps has
to be specified.
bootstrap_array: If true, all the num_bootstrap_samples extimated indices
are returned instead of their mean and Var.
discrete_features: If specified, the inputs associated with the indices in
this list are generated using an integer-valued uniform distribution,
rather than the default (pseudo-)random continuous uniform distribution.
"""
self.input_function = input_function
self.dim: int = bounds.shape[-1]
self.num_mc_samples = num_mc_samples
self.second_order = second_order
self.bootstrap: bool = num_bootstrap_samples > 1
self.num_bootstrap_samples: int = (
num_bootstrap_samples - 1
) # deduct 1 because the first is meant to be the full grid
self.bootstrap_array = bootstrap_array
self.device: torch.device = bounds.device
self.A: torch.Tensor
self.B: torch.Tensor
if input_qmc:
seed_A, seed_B = 1234, 5678 # to make it reproducible
self.A = draw_sobol_samples(
bounds=bounds, n=num_mc_samples, q=1, seed=seed_A
).squeeze(1)
self.B = draw_sobol_samples(
bounds=bounds, n=num_mc_samples, q=1, seed=seed_B
).squeeze(1)
else:
self.A = unnormalize(
torch.rand(num_mc_samples, self.dim, device=self.device),
bounds=bounds,
)
self.B = unnormalize(
torch.rand(num_mc_samples, self.dim, device=self.device),
bounds=bounds,
)
# uniform integral distribution for discrete features
self.A = sample_discrete_parameters(
input_mc_samples=self.A,
discrete_features=discrete_features,
bounds=bounds,
num_mc_samples=num_mc_samples,
)
self.B = sample_discrete_parameters(
input_mc_samples=self.B,
discrete_features=discrete_features,
bounds=bounds,
num_mc_samples=num_mc_samples,
)
self.A_B_ABi: torch.Tensor = self.generate_all_input_matrix().to(torch.double)
if self.bootstrap:
subset_size = 4
self.bootstrap_indices: torch.Tensor = torch.randint(
0, num_mc_samples, (self.num_bootstrap_samples, subset_size)
)
self.f_A: torch.Tensor | None = None
self.f_B: torch.Tensor | None = None
self.f_ABis: list[torch.Tensor] | None = None
self.f_total_var: torch.Tensor | None = None
self.f_A_btsp: list[torch.Tensor] | None = None
self.f_B_btsp: list[torch.Tensor] | None = None
self.f_ABis_btsp: list[list[torch.Tensor]] | None = None
self.f_total_var_btsp: list[torch.Tensor] | None = None
self.f_BAis: list[torch.Tensor] | None = None
self.f_BAis_btsp: list[list[torch.Tensor]] | None = None
self.first_order_idcs: torch.Tensor | None = first_order_idcs
self.first_order_idcs_btsp: torch.Tensor | None = None
[docs]
def evalute_function(self, f_A_B_ABi: torch.Tensor | None = None) -> None:
r"""evaluates the objective function and devides the evaluation into
torch.Tensors needed for the indices computation.
Args:
f_A_B_ABi: Function evaluations on the entire grid of size M(d+2).
"""
if f_A_B_ABi is None:
f_A_B_ABi = none_throws(self.input_function)(self.A_B_ABi)
# for multiple output models, average the outcomes
if len(f_A_B_ABi.shape) == 3:
f_A_B_ABi = f_A_B_ABi.mean(dim=0)
self.f_A = f_A_B_ABi[: self.num_mc_samples]
self.f_B = f_A_B_ABi[self.num_mc_samples : self.num_mc_samples * 2]
# first, there is simply A and B, and then there are all the ablated variants
# of each (which are retrieved here by slicing the input matrix)
# but this only retrieves the ablations of A (and not B)
self.f_ABis = [
f_A_B_ABi[self.num_mc_samples * (i + 2) : self.num_mc_samples * (i + 3)]
for i in range(self.dim)
]
# Get the variances of A and B (so simply the variance of 2 num_mc samples)
self.f_total_var = torch.var(f_A_B_ABi[: self.num_mc_samples * 2])
if self.bootstrap:
f_A = none_throws(self.f_A)
f_B = none_throws(self.f_B)
f_ABis = none_throws(self.f_ABis)
self.f_A_btsp = [
torch.index_select(f_A, 0, indices)
for indices in self.bootstrap_indices
]
self.f_B_btsp = [
torch.index_select(f_B, 0, indices)
for indices in self.bootstrap_indices
]
self.f_ABis_btsp = [
[torch.index_select(f_ABi, 0, indices) for f_ABi in f_ABis]
for indices in self.bootstrap_indices
]
f_A_btsp = self.f_A_btsp
f_B_btsp = self.f_B_btsp
self.f_total_var_btsp = [
torch.var(
torch.cat(
(f_A_btsp[i], f_B_btsp[i]),
dim=0,
)
)
for i in range(self.num_bootstrap_samples)
]
if self.second_order:
# If second order, we also need to retrieve the ablations of B
# In total, we have f_ABis and f_BAis which are both of size M(d), and are
# the ablations of each other
self.f_BAis = [
f_A_B_ABi[
self.num_mc_samples * (i + 2 + self.dim) : self.num_mc_samples
* (i + 3 + self.dim)
]
for i in range(self.dim)
]
if self.bootstrap:
f_BAis = self.f_BAis
self.f_BAis_btsp = [
[torch.index_select(f_BAi, 0, indices) for f_BAi in f_BAis]
for indices in self.bootstrap_indices
]
[docs]
def first_order_indices(self) -> Tensor:
r"""Computes the first order Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values,var_mc,stderr_mc)x dim
else
Tensor: (values)x dim
"""
f_A = none_throws(self.f_A)
f_B = none_throws(self.f_B)
f_ABis = none_throws(self.f_ABis)
f_total_var = none_throws(self.f_total_var)
first_order_idcs = []
for i in range(self.dim):
# The only difference between self.f_ABis[i] and self.f_A is in dimension i
# So we get the variance in the component that corresponds to dimension i
vi = torch.mean(f_B * (f_ABis[i] - f_A)) / f_total_var
first_order_idcs.append(vi.unsqueeze(0))
self.first_order_idcs = torch.cat(first_order_idcs, dim=0).detach()
if not self.bootstrap:
return self.first_order_idcs
else:
f_B_btsp = none_throws(self.f_B_btsp)
f_A_btsp = none_throws(self.f_A_btsp)
f_ABis_btsp = none_throws(self.f_ABis_btsp)
f_total_var_btsp = none_throws(self.f_total_var_btsp)
first_order_idcs_btsp = [torch.cat(first_order_idcs, dim=0).unsqueeze(0)]
for b in range(self.num_bootstrap_samples):
first_order_idcs = []
for i in range(self.dim):
vi = (
torch.mean(f_B_btsp[b] * (f_ABis_btsp[b][i] - f_A_btsp[b]))
/ f_total_var_btsp[b]
)
first_order_idcs.append(vi.unsqueeze(0))
first_order_idcs_btsp.append(
torch.cat(first_order_idcs, dim=0).unsqueeze(0)
)
self.first_order_idcs_btsp = torch.cat(first_order_idcs_btsp, dim=0)
if self.bootstrap_array:
return self.first_order_idcs_btsp.detach()
else:
btsp = none_throws(self.first_order_idcs_btsp)
return (
torch.cat(
[
btsp.mean(dim=0).unsqueeze(0),
btsp.var(dim=0).unsqueeze(0),
torch.sqrt(
btsp.var(dim=0) / (self.num_bootstrap_samples + 1)
).unsqueeze(0),
],
dim=0,
)
.t()
.detach()
)
[docs]
def total_order_indices(self) -> Tensor:
r"""Computes the total Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values,var_mc,stderr_mc)x dim
else
Tensor: (values)x dim
"""
f_A = none_throws(self.f_A)
f_ABis = none_throws(self.f_ABis)
f_total_var = none_throws(self.f_total_var)
total_order_idcs = []
for i in range(self.dim):
vti = 0.5 * torch.mean(torch.pow(f_A - f_ABis[i], 2)) / f_total_var
total_order_idcs.append(vti.unsqueeze(0))
if not (self.bootstrap):
total_order_idcs = torch.cat(total_order_idcs, dim=0).detach()
return total_order_idcs
else:
f_A_btsp = none_throws(self.f_A_btsp)
f_ABis_btsp = none_throws(self.f_ABis_btsp)
f_total_var_btsp = none_throws(self.f_total_var_btsp)
total_order_idcs_btsp = [torch.cat(total_order_idcs, dim=0).unsqueeze(0)]
for b in range(self.num_bootstrap_samples):
total_order_idcs = []
for i in range(self.dim):
vti = (
0.5
* torch.mean(torch.pow(f_A_btsp[b] - f_ABis_btsp[b][i], 2))
/ f_total_var_btsp[b]
)
total_order_idcs.append(vti.unsqueeze(0))
total_order_idcs_btsp.append(
torch.cat(total_order_idcs, dim=0).unsqueeze(0)
)
total_order_idcs_btsp = torch.cat(total_order_idcs_btsp, dim=0)
if self.bootstrap_array:
return total_order_idcs_btsp.detach()
else:
return (
torch.cat(
[
total_order_idcs_btsp.mean(dim=0).unsqueeze(0),
total_order_idcs_btsp.var(dim=0).unsqueeze(0),
torch.sqrt(
total_order_idcs_btsp.var(dim=0)
/ (self.num_bootstrap_samples + 1)
).unsqueeze(0),
],
dim=0,
)
.t()
.detach()
)
[docs]
def second_order_indices(
self,
first_order_idcs: torch.Tensor | None = None,
first_order_idcs_btsp: torch.Tensor | None = None,
) -> Tensor:
r"""Computes the Second order Sobol indices:
Args:
first_order_idcs: Tensor of previously computed first order indices, where
first_order_idcs.shape = torch.Size([dim]).
first_order_idcs_btsp: Tensor of all first order indices given by bootstrap.
Returns:
if num_bootstrap_samples>1
Tensor: (values,var_mc,stderr_mc)x dim
else
Tensor: (values)x dim
"""
if not self.second_order:
raise ValueError(
"Second order indices has to be specified in the sensitivity definition"
)
# TODO Improve this part of the code by vectorization T204291129
if first_order_idcs is None:
if self.first_order_idcs is None:
self.first_order_indices()
first_order_idcs = self.first_order_idcs
f_A = none_throws(self.f_A)
f_B = none_throws(self.f_B)
f_ABis = none_throws(self.f_ABis)
f_BAis = none_throws(self.f_BAis)
f_total_var = none_throws(self.f_total_var)
first_order_idcs_val = none_throws(first_order_idcs)
second_order_idcs = []
for i in range(self.dim):
for j in range(i + 1, self.dim):
vij = torch.mean(f_BAis[i] * f_ABis[j] - f_A * f_B)
vij = (
(vij / f_total_var)
- first_order_idcs_val[i]
- first_order_idcs_val[j]
)
second_order_idcs.append(vij.unsqueeze(0))
if not self.bootstrap:
second_order_idcs = torch.cat(second_order_idcs, dim=0).detach()
return second_order_idcs
else:
f_A_btsp = none_throws(self.f_A_btsp)
f_B_btsp = none_throws(self.f_B_btsp)
f_ABis_btsp = none_throws(self.f_ABis_btsp)
f_BAis_btsp = none_throws(self.f_BAis_btsp)
f_total_var_btsp = none_throws(self.f_total_var_btsp)
second_order_idcs_btsp = [torch.cat(second_order_idcs, dim=0).unsqueeze(0)]
if first_order_idcs_btsp is None:
first_order_idcs_btsp = none_throws(self.first_order_idcs_btsp)
for b in range(self.num_bootstrap_samples):
second_order_idcs = []
for i in range(self.dim):
for j in range(i + 1, self.dim):
vij = torch.mean(
f_BAis_btsp[b][i] * f_ABis_btsp[b][j]
- f_A_btsp[b] * f_B_btsp[b]
)
vij = (
(vij / f_total_var_btsp[b])
- first_order_idcs_btsp[b][i]
- first_order_idcs_btsp[b][j]
)
second_order_idcs.append(vij.unsqueeze(0))
second_order_idcs_btsp.append(
torch.cat(second_order_idcs, dim=0).unsqueeze(0)
)
second_order_idcs_btsp = torch.cat(second_order_idcs_btsp, dim=0)
if self.bootstrap_array:
return second_order_idcs_btsp.detach()
else:
return (
torch.cat(
[
second_order_idcs_btsp.mean(dim=0).unsqueeze(0),
second_order_idcs_btsp.var(dim=0).unsqueeze(0),
torch.sqrt(
second_order_idcs_btsp.var(dim=0)
/ (self.num_bootstrap_samples + 1)
).unsqueeze(0),
],
dim=0,
)
.t()
.detach()
)
[docs]
def GaussianLinkMean(mean: torch.Tensor, var: torch.Tensor) -> torch.Tensor:
return mean
[docs]
def ProbitLinkMean(mean: torch.Tensor, var: torch.Tensor) -> torch.Tensor:
a = mean / torch.sqrt(1 + var)
return torch.distributions.Normal(0, 1).cdf(a)
[docs]
class SobolSensitivityGPMean:
def __init__(
self,
model: GPyTorchModel,
bounds: torch.Tensor,
num_mc_samples: int = 10**4,
second_order: bool = False,
input_qmc: bool = False,
num_bootstrap_samples: int = 1,
link_function: Callable[
[torch.Tensor, torch.Tensor], torch.Tensor
] = GaussianLinkMean,
discrete_features: list[int] | None = None,
) -> None:
r"""Computes three types of Sobol indices:
first order indices, total indices and second order indices (if specified ).
Args:
model: BoTorch model whose posterior is a `GPyTorchPosterior`.
bounds: `2 x d` parameter bounds over which to evaluate model sensitivity.
method: if "predictive mean", the predictive mean is used for indices
computation. If "GP samples", posterior sampling is used instead.
num_mc_samples: The number of montecarlo grid samples
second_order: If True, the second order indices are computed.
input_qmc: If True, a qmc Sobol grid is use instead of uniformly random.
num_bootstrap_samples: If bootstrap is true, the number of bootstraps has
to be specified.
link_function: The link function to be used when computing the indices.
Indices can be computed for the mean or on samples of the posterior,
predictive, but defaults to computing on the mean (GaussianLinkMean).
discrete_features: If specified, the inputs associated with the indices in
this list are generated using an integer-valued uniform distribution,
rather than the default (pseudo-)random continuous uniform distribution.
"""
self.model = model
self.second_order = second_order
self.input_qmc = input_qmc
self.bootstrap: bool = num_bootstrap_samples > 1
self.num_bootstrap_samples = num_bootstrap_samples
self.num_mc_samples = num_mc_samples
def input_function(x: Tensor) -> Tensor:
assert x.ndim == 2, "Input must be a 2D tensor."
with torch.no_grad():
# NOTE: Do not unsqueeze x_ here. Likely due to a matmul issue
# this ends up using a lot of memory with batched models.
# See https://github.com/pytorch/botorch/issues/2310.
means, variances = [], []
for x_ in x.split(4096):
p = self.model.posterior(x_)
means.append(p.mean)
variances.append(p.variance)
mean = torch.cat(means, dim=-2)
variance = torch.cat(variances, dim=-2)
return link_function(mean, variance)
self.sensitivity = SobolSensitivity(
bounds=bounds,
num_mc_samples=self.num_mc_samples,
input_function=input_function,
second_order=self.second_order,
input_qmc=self.input_qmc,
num_bootstrap_samples=self.num_bootstrap_samples,
discrete_features=discrete_features,
)
self.sensitivity.evalute_function()
[docs]
def first_order_indices(self) -> Tensor:
r"""Computes the first order Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values,var_mc,stderr_mc)x dim
else
Tensor: (values)x dim
"""
return self.sensitivity.first_order_indices()
[docs]
def total_order_indices(self) -> Tensor:
r"""Computes the total Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values,var_mc,stderr_mc)x dim
else
Tensor: (values)x dim
"""
return self.sensitivity.total_order_indices()
[docs]
def second_order_indices(self) -> Tensor:
r"""Computes the Second order Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values,var_mc,stderr_mc)x dim(dim-1)/2
else
Tensor: (values)x dim(dim-1)/2
"""
return self.sensitivity.second_order_indices()
[docs]
class SobolSensitivityGPSampling:
def __init__(
self,
model: Model,
bounds: torch.Tensor,
num_gp_samples: int = 10**3,
num_mc_samples: int = 10**4,
second_order: bool = False,
input_qmc: bool = False,
gp_sample_qmc: bool = False,
num_bootstrap_samples: int = 1,
discrete_features: list[int] | None = None,
) -> None:
r"""Computes three types of Sobol indices:
first order indices, total indices and second order indices (if specified ).
Args:
model: Botorch model.
bounds: `2 x d` parameter bounds over which to evaluate model sensitivity.
num_gp_samples: If method is "GP samples", the number of GP samples
has to be set.
num_mc_samples: The number of montecarlo grid samples
second_order: If True, the second order indices are computed.
input_qmc: If True, a qmc Sobol grid is use instead of uniformly random.
gp_sample_qmc: If True, the posterior sampling is done using
SobolQMCNormalSampler.
num_bootstrap_samples: If bootstrap is true, the number of bootstraps has
to be specified.
discrete_features: If specified, the inputs associated with the indices in
this list are generated using an integer-valued uniform distribution,
rather than the default (pseudo-)random continuous uniform distribution.
"""
self.model = model
self.second_order = second_order
self.input_qmc = input_qmc
self.gp_sample_qmc = gp_sample_qmc
self.bootstrap: bool = num_bootstrap_samples > 1
self.num_bootstrap_samples = num_bootstrap_samples
self.num_mc_samples = num_mc_samples
self.num_gp_samples = num_gp_samples
self.first_order_idcs_list: torch.Tensor = torch.empty(0)
self.sensitivity = SobolSensitivity(
bounds=bounds,
num_mc_samples=self.num_mc_samples,
second_order=self.second_order,
input_qmc=self.input_qmc,
num_bootstrap_samples=self.num_bootstrap_samples,
bootstrap_array=True,
discrete_features=discrete_features,
)
# TODO: Ideally, we would reduce the memory consumption here as well
# but this is a tricky since it uses joint posterior sampling.
posterior = self.model.posterior(self.sensitivity.A_B_ABi)
if self.gp_sample_qmc:
sampler = SobolQMCNormalSampler(
sample_shape=torch.Size([self.num_gp_samples]), seed=0
)
self.samples: torch.Tensor = sampler(posterior)
else:
with torch.no_grad():
self.samples = posterior.rsample(torch.Size([self.num_gp_samples]))
@property
def dim(self) -> int:
"""Returns the input dimensionality of `self.model`."""
return self.sensitivity.dim
[docs]
def first_order_indices(self) -> Tensor:
r"""Computes the first order Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values, var_gp, stderr_gp, var_mc, stderr_mc) x dim
else
Tensor: (values, var, stderr) x dim
"""
first_order_idcs_list = []
for j in range(self.num_gp_samples):
self.sensitivity.evalute_function(self.samples[j])
first_order_idcs = self.sensitivity.first_order_indices()
first_order_idcs_list.append(first_order_idcs.unsqueeze(0))
self.first_order_idcs_list: torch.Tensor = torch.cat(
first_order_idcs_list, dim=0
)
if not (self.bootstrap):
first_order_idcs_mean_var_se = []
for i in range(self.dim):
first_order_idcs_mean_var_se.append(
torch.tensor(
[
torch.mean(self.first_order_idcs_list[:, i]),
torch.var(self.first_order_idcs_list[:, i]),
torch.sqrt(
torch.var(self.first_order_idcs_list[:, i])
/ self.num_gp_samples
),
]
).unsqueeze(0)
)
first_order_idcs_mean_var_se = torch.cat(
first_order_idcs_mean_var_se, dim=0
)
return first_order_idcs_mean_var_se
else:
var_per_bootstrap = torch.var(self.first_order_idcs_list, dim=0)
gp_var = torch.mean(var_per_bootstrap, dim=0)
gp_se = torch.sqrt(gp_var / self.num_gp_samples)
var_per_gp_sample = torch.var(self.first_order_idcs_list, dim=1)
mc_var = torch.mean(var_per_gp_sample, dim=0)
mc_se = torch.sqrt(mc_var / (self.num_bootstrap_samples + 1))
total_mean = self.first_order_idcs_list.reshape(-1, self.dim).mean(dim=0)
first_order_idcs_mean_vargp_segp_varmc_segp = torch.cat(
[
torch.tensor(
[total_mean[i], gp_var[i], gp_se[i], mc_var[i], mc_se[i]]
).unsqueeze(0)
for i in range(self.dim)
],
dim=0,
)
return first_order_idcs_mean_vargp_segp_varmc_segp
[docs]
def total_order_indices(self) -> Tensor:
r"""Computes the total Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values, var_gp, stderr_gp, var_mc, stderr_mc) x dim
else
Tensor: (values, var, stderr) x dim
"""
total_order_idcs_list = []
for j in range(self.num_gp_samples):
self.sensitivity.evalute_function(self.samples[j])
total_order_idcs = self.sensitivity.total_order_indices()
total_order_idcs_list.append(total_order_idcs.unsqueeze(0))
total_order_idcs_list = torch.cat(total_order_idcs_list, dim=0)
if not (self.bootstrap):
total_order_idcs_mean_var = []
for i in range(self.dim):
total_order_idcs_mean_var.append(
torch.tensor(
[
torch.mean(total_order_idcs_list[:, i]),
torch.var(total_order_idcs_list[:, i]),
torch.sqrt(
torch.var(total_order_idcs_list[:, i])
/ self.num_gp_samples
),
]
).unsqueeze(0)
)
total_order_idcs_mean_var = torch.cat(total_order_idcs_mean_var, dim=0)
return total_order_idcs_mean_var
else:
var_per_bootstrap = torch.var(total_order_idcs_list, dim=0)
gp_var = torch.mean(var_per_bootstrap, dim=0)
gp_se = torch.sqrt(gp_var / self.num_gp_samples)
var_per_gp_sample = torch.var(total_order_idcs_list, dim=1)
mc_var = torch.mean(var_per_gp_sample, dim=0)
mc_se = torch.sqrt(mc_var / (self.num_bootstrap_samples + 1))
total_mean = total_order_idcs_list.reshape(-1, self.dim).mean(dim=0)
total_order_idcs_mean_vargp_segp_varmc_segp = torch.cat(
[
torch.tensor(
[total_mean[i], gp_var[i], gp_se[i], mc_var[i], mc_se[i]]
).unsqueeze(0)
for i in range(self.dim)
],
dim=0,
)
return total_order_idcs_mean_vargp_segp_varmc_segp
[docs]
def second_order_indices(self) -> Tensor:
r"""Computes the Second order Sobol indices:
Returns:
if num_bootstrap_samples>1
Tensor: (values, var_gp, stderr_gp, var_mc, stderr_mc) x dim(dim-1) / 2
else
Tensor: (values, var, stderr) x dim(dim-1) / 2
"""
if not (self.bootstrap):
second_order_idcs_list = []
second_order_idcs: Tensor = torch.empty(0)
for j in range(self.num_gp_samples):
self.sensitivity.evalute_function(self.samples[j])
second_order_idcs = self.sensitivity.second_order_indices(
self.first_order_idcs_list[j]
)
second_order_idcs_list.append(second_order_idcs.unsqueeze(0))
second_order_idcs_list = torch.cat(second_order_idcs_list, dim=0)
second_order_idcs_mean_var = []
for i in range(len(second_order_idcs)):
second_order_idcs_mean_var.append(
torch.tensor(
[
torch.mean(second_order_idcs_list[:, i]),
torch.var(second_order_idcs_list[:, i]),
torch.sqrt(
torch.var(second_order_idcs_list[:, i])
/ self.num_gp_samples
),
]
).unsqueeze(0)
)
second_order_idcs_mean_var = torch.cat(second_order_idcs_mean_var, dim=0)
return second_order_idcs_mean_var
else:
second_order_idcs_list = []
for j in range(self.num_gp_samples):
self.sensitivity.evalute_function(self.samples[j])
second_order_idcs = self.sensitivity.second_order_indices(
self.first_order_idcs_list[j][0],
self.first_order_idcs_list[j][1:],
)
second_order_idcs_list.append(second_order_idcs.unsqueeze(0))
second_order_idcs_list = torch.cat(second_order_idcs_list, dim=0)
var_per_bootstrap = torch.var(second_order_idcs_list, dim=0)
gp_var = torch.mean(var_per_bootstrap, dim=0)
gp_se = torch.sqrt(gp_var / self.num_gp_samples)
var_per_gp_sample = torch.var(second_order_idcs_list, dim=1)
mc_var = torch.mean(var_per_gp_sample, dim=0)
mc_se = torch.sqrt(mc_var / (self.num_bootstrap_samples + 1))
num_second_order = second_order_idcs_list.shape[-1]
total_mean = second_order_idcs_list.reshape(-1, num_second_order).mean(
dim=0
)
second_order_idcs_mean_vargp_segp_varmc_segp = torch.cat(
[
torch.tensor(
[total_mean[i], gp_var[i], gp_se[i], mc_var[i], mc_se[i]]
).unsqueeze(0)
for i in range(num_second_order)
],
dim=0,
)
return second_order_idcs_mean_vargp_segp_varmc_segp
[docs]
def compute_sobol_indices_from_model_list(
model_list: list[GPyTorchModel],
bounds: Tensor,
order: str = "first",
discrete_features: list[int] | None = None,
fixed_features: dict[int, float] | None = None,
**sobol_kwargs: Any,
) -> Tensor:
"""
Computes Sobol indices of a list of models on a bounded domain.
Args:
model_list: A list of botorch.models.model.Model types for which to compute
the Sobol indices.
bounds: A 2 x d Tensor of lower and upper bounds of the domain of the models.
order: A string specifying the order of the Sobol indices to be computed.
Supports "first", "second" and "total" and defaults to "first". "total"
computes the importance of a variable considering its main effect and
all of its higher-order interactions, whereas "first" and "second"
the variance when altering the variable in isolation or with one other
variable, respectively.
discrete_features: If specified, the inputs associated with the indices in
this list are generated using an integer-valued uniform distribution,
rather than the default (pseudo-)random continuous uniform distribution.
fixed_features: If specified, a dictionary mapping feature indices to fixed
values. These features will be held constant during sensitivity analysis,
and their sensitivity will not be computed. The bounds tensor should
still include all features (including fixed ones).
sobol_kwargs: keyword arguments passed on to SobolSensitivityGPMean.
Returns:
With m GPs, returns a (m x d') tensor of `order`-order Sobol indices,
where d' is the number of non-fixed features.
"""
if order not in ["first", "total", "second"]:
raise NotImplementedError(
f"Order {order} is not supported. Plese choose one of"
" 'first', 'total' or 'second'."
)
# Handle fixed features by reducing bounds and wrapping models
models_to_use: list[GPyTorchModel] | list[FixedFeatureModel] = model_list
if fixed_features is not None and len(fixed_features) > 0:
models_to_use, bounds, discrete_features = prepare_fixed_feature_inputs(
model_list=model_list,
bounds=bounds,
discrete_features=discrete_features,
fixed_features=fixed_features,
)
indices = []
method = getattr(SobolSensitivityGPMean, f"{order}_order_indices")
second_order = order == "second"
for model in models_to_use:
sens_class = SobolSensitivityGPMean(
model=model, # pyre-ignore[6]: FixedFeatureModel wraps GPyTorchModel
bounds=bounds,
discrete_features=discrete_features,
second_order=second_order,
**sobol_kwargs,
)
indices.append(method(sens_class))
return torch.stack(indices)
[docs]
def ax_parameter_sens(
adapter: TorchAdapter,
metrics: list[str] | None = None,
order: str = "first",
signed: bool = True,
exclude_map_key: bool = True,
exclude_task: bool = False,
**sobol_kwargs: Any,
) -> dict[str, dict[str, npt.NDArray]]:
"""
Compute sensitivity for all metrics on an TorchAdapter.
Sobol measures are always positive regardless of the direction in which the
parameter influences f. If `signed` is set to True, then the Sobol measure for each
parameter will be given as its sign the sign of the average gradient with respect to
that parameter across the search space. Thus, important parameters that, when
increased, decrease f will have large and negative values; unimportant parameters
will have values close to 0.
Args:
adapter: A Adapter object with models that were fit.
metrics: The names of the metrics and outcomes for which to compute
sensitivities. This should preferably be metrics with a good model fit.
Defaults to adapter.outcomes.
order: A string specifying the order of the Sobol indices to be computed.
Supports "first" and "total" and defaults to "first".
signed: A bool for whether the measure should be signed.
exclude_map_key: If True (default), the MAP_KEY ("step") feature will be
excluded from sensitivity analysis by fixing it at the maximum step value.
This makes the sensitivity analysis more interpretable for users who
care about the effect of parameters on final performance.
exclude_task: If True, task parameters (those with ``is_task=True``, e.g.
synthetic parameters from the TrialAsTask transform) will be excluded
from the sensitivity results.
sobol_kwargs: keyword arguments passed on to SobolSensitivityGPMean, and if
signed, GpDGSMGpMean.
Returns:
Dictionary {'metric_name': {'parameter_name' or
(parameter_name_1, 'parameter_name_2'): sensitivity_value}}, where the
`sensitivity` value is cast to a Numpy array in order to be compatible with
`plot_feature_importance_by_feature`.
"""
if order not in ["first", "total", "second"]:
raise NotImplementedError(
f"Order {order} is not supported. Plese choose one of"
" 'first', 'total' or 'second'."
)
if metrics is None:
metrics = adapter.outcomes
generator, digest = _get_generator_and_digest(adapter=adapter)
model_list = _get_model_per_metric(generator=generator, metrics=metrics)
# get device and dtype of the first model
first_model = next(model_list[0].parameters())
device = first_model.device
bounds = torch.tensor(
digest.bounds,
device=device,
).T # transposing to make it 2 x d
feature_names = digest.feature_names
# Determine if we need to fix the MAP_KEY feature
fixed_features: dict[int, float] | None = None
output_feature_names = feature_names
discrete_features = digest.categorical_features + digest.ordinal_features
if exclude_map_key and MAP_KEY in feature_names:
step_idx = feature_names.index(MAP_KEY)
# Use upper bound (maximum step value in normalized space)
step_value = float(bounds[1, step_idx])
fixed_features = {step_idx: step_value}
# Remove MAP_KEY from output feature names
output_feature_names = [f for f in feature_names if f != MAP_KEY]
# Exclude task parameters (e.g. TRIAL_PARAM from TrialAsTask transform)
# by fixing them at their target values.
if exclude_task and digest.task_features:
if fixed_features is None:
fixed_features = {}
for task_idx in digest.task_features:
if task_idx < len(feature_names):
fixed_features[task_idx] = float(
digest.target_values.get(task_idx, bounds[1, task_idx])
)
task_name = feature_names[task_idx]
output_feature_names = [
f for f in output_feature_names if f != task_name
]
# for second order indices, we need to compute first order indices first
# which is what is done here. With the first order indices, we can then subtract
# appropriately using the first-order indices to extract the second-order indices.
ind = compute_sobol_indices_from_model_list(
model_list=model_list,
bounds=bounds,
order="first" if order == "second" else order,
discrete_features=discrete_features,
fixed_features=fixed_features,
**sobol_kwargs,
)
if signed:
ind_deriv = compute_derivatives_from_model_list(
model_list=model_list,
bounds=bounds,
discrete_features=discrete_features,
fixed_features=fixed_features,
**sobol_kwargs,
)
# categorical features don't have a direction, so we set the derivative to 1.0
# in order not to zero out their sensitivity. We treat categorical features
# separately in the sensitivity analysis plot as well, to make clear that they
# are affecting the metric, but neither increasing nor decreasing. Note that the
# original variables have a well defined direction, so we do not need to treat
# them differently here.
for i in digest.categorical_features:
ind_deriv[:, i] = 1.0
ind *= torch.sign(ind_deriv).to(device)
indices = array_with_string_indices_to_dict(
rows=metrics, cols=output_feature_names, A=ind.cpu().numpy()
)
if order == "second" and len(output_feature_names) >= 2:
second_order_values = compute_sobol_indices_from_model_list(
model_list=model_list,
bounds=bounds,
order="second",
discrete_features=discrete_features,
fixed_features=fixed_features,
**sobol_kwargs,
)
second_order_feature_names = [
f"{f1} & {f2}" for f1, f2 in itertools.combinations(output_feature_names, 2)
]
second_order_dict = array_with_string_indices_to_dict(
rows=metrics,
cols=second_order_feature_names,
A=second_order_values.cpu().numpy(),
)
for metric in metrics:
indices[metric].update(second_order_dict[metric])
return indices
def _get_generator_and_digest(
adapter: TorchAdapter,
) -> tuple[BoTorchGenerator, SearchSpaceDigest]:
"""Returns the generator of the adapter and the SearchSpaceDigest
that was used to fit the adapter.
"""
if not isinstance(adapter, TorchAdapter):
raise NotImplementedError(
f"{type(adapter)=}, but only TorchAdapter is supported."
)
generator = adapter.generator
if not isinstance(generator, (BoTorchGenerator)):
raise NotImplementedError(
f"{type(generator)=}, but only BoTorchGenerator is supported."
)
return generator, generator.search_space_digest
def _get_model_per_metric(
generator: BoTorchGenerator, metrics: list[str]
) -> list[GPyTorchModel]:
"""For a given BoTorchGenerator model, returns a list of botorch.models.model.Model
objects corresponding to - and in the same order as - the given metrics.
"""
surrogate = generator.surrogate
outcomes = surrogate.outcomes
model_list = []
for m in metrics: # for each metric, find a corresponding surrogate
i = outcomes.index(m)
metric_model = surrogate.model
# since model is a BoTorchGenerator, metric_model will be a
# `botorch.models.model.Model` object, which have the `num_outputs`
# property and `subset_outputs` method.
if metric_model.num_outputs > 1: # subset to relevant output
metric_model = metric_model.subset_output([i])
model_list.append(metric_model)
return model_list
[docs]
def array_with_string_indices_to_dict(
rows: list[str],
cols: list[str],
A: npt.NDArray,
) -> dict[str, dict[str, npt.NDArray]]:
"""
Args:
rows: A list of strings with which to index rows of A.
cols: A list of strings with which to index columns of A.
A: A matrix, with `len(rows)` rows and `len(cols)` columns.
Returns:
A dictionary dict that satisfies dict[rows[i]][cols[j]] = A[i, j].
"""
return {r: dict(zip(cols, a)) for r, a in zip(rows, A)}