#!/usr/bin/env python3
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
r"""
Utilities for optimization.
"""
from __future__ import annotations
import warnings
from collections import OrderedDict
from inspect import signature
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import numpy as np
import torch
from gpytorch.mlls.exact_marginal_log_likelihood import ExactMarginalLogLikelihood
from gpytorch.mlls.marginal_log_likelihood import MarginalLogLikelihood
from gpytorch.mlls.sum_marginal_log_likelihood import SumMarginalLogLikelihood
from gpytorch.utils.errors import NanError
from torch import Tensor
from ..exceptions.errors import BotorchError
from ..exceptions.warnings import BotorchWarning
from ..models.gpytorch import GPyTorchModel
from .numpy_converter import TorchAttr, set_params_with_array
[docs]def sample_all_priors(model: GPyTorchModel) -> None:
r"""Sample from hyperparameter priors (in-place).
Args:
model: A GPyTorchModel.
"""
for _, prior, closure, setting_closure in model.named_priors():
if setting_closure is None:
raise RuntimeError(
"Must provide inverse transform to be able to sample from prior."
)
try:
setting_closure(prior.sample(closure().shape))
except NotImplementedError:
warnings.warn(
f"`rsample` not implemented for {type(prior)}. Skipping.",
BotorchWarning,
)
[docs]class ConvergenceCriterion:
r"""Basic class for evaluating optimization convergence.
"""
def __init__(
self,
maxiter: int = 15000,
ftol: float = 2.220446049250313e-09,
minimize: bool = True,
) -> None:
r"""Constructor for ConvergenceCriterion.
Args:
maxiter: maximum number of iterations.
ftol: Function value relative tolerance for termination.
minimize: boolean indicating the optimization direction.
"""
# by default, use the scipy defaults
self.maxiter = maxiter
self.ftol = ftol
self.minimize = minimize
self.prev_fvals = None
self.iter = 0
[docs] def evaluate(self, fvals: Tensor) -> bool:
r"""Evaluate convergence criterion.
Args:
fvals: tensor containing function values for the current iteration.
If `fvals` contains more than one element, then the relative
tolerance criterion is evaluated element-wise and True is returned
if all elements have converged.
TODO: add support for utilizing gradient information
Returns:
bool: convergence indicator
"""
self.iter += 1
if self.iter == self.maxiter:
return True
elif self.prev_fvals is not None:
fmax = torch.stack(
[self.prev_fvals.abs(), fvals.abs(), torch.ones_like(fvals)], dim=0
).max(dim=0)[0]
f_delta = (self.prev_fvals - fvals).div(fmax)
if not self.minimize:
f_delta *= -1
if torch.all(f_delta <= self.ftol):
return True
self.prev_fvals = fvals
return False
[docs]def columnwise_clamp(
X: Tensor,
lower: Optional[Union[float, Tensor]] = None,
upper: Optional[Union[float, Tensor]] = None,
raise_on_violation: bool = False,
) -> Tensor:
r"""Clamp values of a Tensor in column-wise fashion (with support for t-batches).
This function is useful in conjunction with optimizers from the torch.optim
package, which don't natively handle constraints. If you apply this after
a gradient step you can be fancy and call it "projected gradient descent".
This funtion is also useful for post-processing candidates generated by the
scipy optimizer that satisfy bounds only up to numerical accuracy.
Args:
X: The `b x n x d` input tensor. If 2-dimensional, `b` is assumed to be 1.
lower: The column-wise lower bounds. If scalar, apply bound to all columns.
upper: The column-wise upper bounds. If scalar, apply bound to all columns.
raise_on_violation: If `True`, raise an exception when the elments in `X`
are out of the specified bounds (up to numerical accuracy). This is
useful for post-processing candidates generated by optimizers that
satisfy imposed bounds only up to numerical accuracy.
Returns:
The clamped tensor.
"""
min_bounds = _expand_bounds(lower, X)
max_bounds = _expand_bounds(upper, X)
if min_bounds is not None and max_bounds is not None:
if torch.any(min_bounds > max_bounds):
raise ValueError("Minimum values must be <= maximum values")
Xout = X
if min_bounds is not None:
Xout = Xout.max(min_bounds)
if max_bounds is not None:
Xout = Xout.min(max_bounds)
if raise_on_violation and not torch.allclose(Xout, X):
raise BotorchError("Original value(s) are out of bounds.")
return Xout
[docs]def fix_features(
X: Tensor, fixed_features: Optional[Dict[int, Optional[float]]] = None
) -> Tensor:
r"""Fix feature values in a Tensor.
The fixed features will have zero gradient in downstream calculations.
Args:
X: input Tensor with shape `... x p`, where `p` is the number of features
fixed_features: A dictionary with keys as column indices and values
equal to what the feature should be set to in `X`. If the value is
None, that column is just considered fixed. Keys should be in the
range `[0, p - 1]`.
Returns:
The tensor X with fixed features.
"""
if fixed_features is None:
return X
else:
return torch.cat(
[
X[..., i].unsqueeze(-1)
if i not in fixed_features
else _fix_feature(X[..., i].unsqueeze(-1), fixed_features[i])
for i in range(X.shape[-1])
],
dim=-1,
)
def _fix_feature(Z: Tensor, value: Optional[float]) -> Tensor:
r"""Helper function returns a Tensor like `Z` filled with `value` if provided."""
if value is None:
return Z.detach()
return torch.full_like(Z, value)
def _expand_bounds(
bounds: Optional[Union[float, Tensor]], X: Tensor
) -> Optional[Tensor]:
r"""Expands a tensor representing bounds.
Expand the dimension of bounds if necessary such that the last dimension of
bounds is the same as the last dimension of `X`.
Args:
bounds: a bound (either upper or lower) of each column (last dimension)
of `X`. If this is a single float, then all columns have the same bound.
X: `... x d` tensor
Returns:
A tensor of bounds expanded to be compatible with the size of `X` if
bounds is not None, and None if bounds is None.
"""
if bounds is not None:
if not torch.is_tensor(bounds):
bounds = torch.tensor(bounds)
if len(bounds.shape) == 0:
ebounds = bounds.expand(1, X.shape[-1])
elif len(bounds.shape) == 1:
ebounds = bounds.view(1, -1)
else:
ebounds = bounds
if ebounds.shape[1] != X.shape[-1]:
raise RuntimeError(
"Bounds must either be a single value or the same dimension as X"
)
return ebounds.to(dtype=X.dtype, device=X.device)
else:
return None
def _get_extra_mll_args(
mll: MarginalLogLikelihood,
) -> Union[List[Tensor], List[List[Tensor]]]:
r"""Obtain extra arguments for MarginalLogLikelihood objects.
Get extra arguments (beyond the model output and training targets) required
for the particular type of MarginalLogLikelihood for a forward pass.
Args:
mll: The MarginalLogLikelihood module.
Returns:
Extra arguments for the MarginalLogLikelihood.
Returns an empty list if the mll type is unknown.
"""
if isinstance(mll, ExactMarginalLogLikelihood):
return list(mll.model.train_inputs)
elif isinstance(mll, SumMarginalLogLikelihood):
return [list(x) for x in mll.model.train_inputs]
return []
def _filter_kwargs(function: Callable, **kwargs: Any) -> Any:
r"""Filter out kwargs that are not applicable for a given function.
Return a copy of given kwargs dict with only the required kwargs."""
return {k: v for k, v in kwargs.items() if k in signature(function).parameters}
def _scipy_objective_and_grad(
x: np.ndarray, mll: MarginalLogLikelihood, property_dict: Dict[str, TorchAttr]
) -> Tuple[float, np.ndarray]:
r"""Get objective and gradient in format that scipy expects.
Args:
x: The (flattened) input parameters.
mll: The MarginalLogLikelihood module to evaluate.
property_dict: The property dictionary required to "unflatten" the input
parameter vector, as generated by `module_to_array`.
Returns:
2-element tuple containing
- The objective value.
- The gradient of the objective.
"""
mll = set_params_with_array(mll, x, property_dict)
train_inputs, train_targets = mll.model.train_inputs, mll.model.train_targets
mll.zero_grad()
try: # catch linear algebra errors in gpytorch
output = mll.model(*train_inputs)
args = [output, train_targets] + _get_extra_mll_args(mll)
loss = -mll(*args).sum()
except RuntimeError as e:
if isinstance(e, NanError) or "singular" in e.args[0]:
return float("nan"), np.full_like(x, "nan")
else:
raise e # pragma: nocover
loss.backward()
param_dict = OrderedDict(mll.named_parameters())
grad = []
for p_name in property_dict:
t = param_dict[p_name].grad
if t is None:
# this deals with parameters that do not affect the loss
grad.append(np.zeros(property_dict[p_name].shape.numel()))
else:
grad.append(t.detach().view(-1).cpu().double().clone().numpy())
mll.zero_grad()
return loss.item(), np.concatenate(grad)
