# ----------------------------------------------------------------------------
# SymForce - Copyright 2022, Skydio, Inc.
# This source code is under the Apache 2.0 license found in the LICENSE file.
# ----------------------------------------------------------------------------
from __future__ import annotations
import dataclasses
import warnings
from dataclasses import dataclass
from functools import cached_property
import numpy as np
from lcmtypes.sym._index_entry_t import index_entry_t
from lcmtypes.sym._lambda_update_type_t import lambda_update_type_t
from lcmtypes.sym._levenberg_marquardt_solver_failure_reason_t import (
levenberg_marquardt_solver_failure_reason_t,
)
from lcmtypes.sym._optimization_iteration_t import optimization_iteration_t
from lcmtypes.sym._optimization_status_t import optimization_status_t
from lcmtypes.sym._optimizer_params_t import optimizer_params_t
from lcmtypes.sym._sparse_matrix_structure_t import sparse_matrix_structure_t
from lcmtypes.sym._values_t import values_t
from symforce import cc_sym
from symforce import typing as T
from symforce.opt.factor import Factor
from symforce.opt.numeric_factor import NumericFactor
from symforce.values import Values
[docs]class Optimizer:
"""
A nonlinear least-squares optimizer
Typical usage is to construct an Optimizer from a set of factors and keys to optimize, and then
call :meth:`optimize` repeatedly with a :class:`Values <symforce.values.values.Values>`.
Example creation with a single :class:`Factor <.factor.Factor>`::
factor = Factor(
[my_key_0, my_key_1, my_key_2], my_residual_function
)
optimizer = Optimizer(
factors=[factor],
optimized_keys=[my_key_0, my_key_1],
)
And usage::
initial_guess = Values(...)
result = optimizer.optimize(initial_guess)
print(result.optimized_values)
Example creation with an :class:`.optimization_problem.OptimizationProblem` using
:meth:`make_numeric_factors() <.optimization_problem.OptimizationProblem.make_numeric_factors>`.
The linearization functions are generated in ``make_numeric_factors()`` and are linearized with
respect to
:meth:`problem.optimized_keys() <.optimization_problem.OptimizationProblem.optimized_keys>`::
problem = OptimizationProblem(subproblems=[...], residual_blocks=...)
factors = problem.make_numeric_factors("my_problem")
optimizer = Optimizer(factors)
Example creation with an :class:`.optimization_problem.OptimizationProblem` using
:meth:`make_symbolic_factors() <.optimization_problem.OptimizationProblem.make_symbolic_factors>`.
The symbolic factors are converted into numeric factors when the optimizer is created, and are
linearized with respect to the "optimized keys" passed to the optimizer. The linearization
functions are generated when converting to numeric factors when the optimizer is created::
problem = OptimizationProblem(subproblems=[...], residual_blocks=...)
factors = problem.make_symbolic_factors("my_problem")
optimizer = Optimizer(factors, problem.optimized_keys())
Wraps the C++ ``sym::Optimizer`` class in ``opt/optimizer.h``, so the API is mostly the same and
optimization results will be identical.
Args:
factors: A sequence of either Factor or NumericFactor objects representing the
residuals in the problem. If (symbolic) Factors are passed, they are convered to
NumericFactors by generating linearization functions of the residual with respect to the
keys in ``optimized_keys``.
optimized_keys: A set of the keys to be optimized. Only required if symbolic factors are
passed to the optimizer.
params: Params for the optimizer
"""
[docs] @dataclass
class Params:
"""
Parameters for the Python Optimizer
Mirrors the ``optimizer_params_t`` LCM type, see documentation there for information on each
parameter.
Note: For the Python optimizer, verbose defaults to True
"""
verbose: bool = True
debug_stats: bool = False
check_derivatives: bool = False
include_jacobians: bool = False
debug_checks: bool = False
initial_lambda: float = 1.0
lambda_lower_bound: float = 0.0
lambda_upper_bound: float = 1000000.0
lambda_update_type: lambda_update_type_t = lambda_update_type_t.STATIC
lambda_up_factor: float = 4.0
lambda_down_factor: float = 1 / 4.0
dynamic_lambda_update_beta: float = 2.0
dynamic_lambda_update_gamma: float = 3.0
dynamic_lambda_update_p: int = 3
use_diagonal_damping: bool = False
use_unit_damping: bool = True
keep_max_diagonal_damping: bool = False
diagonal_damping_min: float = 1e-6
iterations: int = 50
early_exit_min_reduction: float = 1e-6
enable_bold_updates: bool = False
Status = optimization_status_t
FailureReason = levenberg_marquardt_solver_failure_reason_t
[docs] @dataclass
class Result:
"""
The result of an optimization, with additional stats and debug information
Attributes:
initial_values:
The initial guess used for this optimization
optimized_values:
The best Values achieved during the optimization (Values with the smallest error)
iterations:
Per-iteration stats, if requested, like the error per iteration. If debug stats are
turned on, also the Values and linearization per iteration.
best_index:
The index into iterations for the iteration that produced the smallest error. I.e.
``result.iterations[best_index].values == optimized_values``. This is not
guaranteed to be the last iteration, if the optimizer tried additional steps which
did not reduce the error
status:
What was the result of the optimization? (did it converge, fail, etc.)
failure_reason:
If ``status == FAILED``, why?
best_linearization:
The linearization at best_index (at optimized_values), filled out if
``populate_best_linearization=True``
jacobian_sparsity:
The sparsity pattern of the jacobian, filled out if ``debug_stats=True`` and
``include_jacobians=True``
linear_solver_ordering:
The ordering used for the linear solver, filled out if ``debug_stats=True``
cholesky_factor_sparsity:
The sparsity pattern of the cholesky factor L, filled out if ``debug_stats=True``
"""
initial_values: Values
optimized_values: Values
# Private field holding the original stats - we expose fields of this through properties,
# since some of the conversions out of this are expensive
_stats: cc_sym.OptimizationStats
@cached_property
def iterations(self) -> T.List[optimization_iteration_t]:
return self._stats.iterations
@cached_property
def best_index(self) -> int:
return self._stats.best_index
@cached_property
def status(self) -> optimization_status_t:
return self._stats.status
@cached_property
def failure_reason(self) -> levenberg_marquardt_solver_failure_reason_t:
return Optimizer.FailureReason(self._stats.failure_reason)
@cached_property
def best_linearization(self) -> T.Optional[cc_sym.Linearization]:
return self._stats.best_linearization
@cached_property
def jacobian_sparsity(self) -> sparse_matrix_structure_t:
return self._stats.jacobian_sparsity
@cached_property
def linear_solver_ordering(self) -> np.ndarray:
return self._stats.linear_solver_ordering
@cached_property
def cholesky_factor_sparsity(self) -> sparse_matrix_structure_t:
return self._stats.cholesky_factor_sparsity
[docs] def error(self) -> float:
"""
The lowest error achieved by the optimization (the error at optimized_values)
"""
return self.iterations[self.best_index].new_error
def __init__(
self,
factors: T.Iterable[T.Union[Factor, NumericFactor]],
optimized_keys: T.Optional[T.Sequence[str]] = None,
params: T.Optional[Optimizer.Params] = None,
debug_stats: T.Optional[bool] = None,
include_jacobians: T.Optional[bool] = None,
):
if optimized_keys is None:
# This will be filled with the optimized keys of the numeric factors
self.optimized_keys = []
else:
self.optimized_keys = list(optimized_keys)
assert len(optimized_keys) == len(
set(optimized_keys)
), f"Duplicates in optimized keys: {optimized_keys}"
optimized_keys_set = set(self.optimized_keys)
numeric_factors = []
for factor in factors:
if isinstance(factor, Factor):
if optimized_keys is None:
raise ValueError(
"You must specify `optimized_keys` when passing symbolic factors."
)
# We compute the linearization in the same order as `factor.keys` so that using the
# same factor with different problem keys will produce the same generated function
factor_opt_keys = [
opt_key for opt_key in factor.keys if opt_key in optimized_keys_set
]
if not factor_opt_keys:
raise ValueError(
f"Factor {factor.name} has no arguments (keys: {factor.keys}) in "
+ f"optimized_keys ({optimized_keys})."
)
numeric_factors.append(factor.to_numeric_factor(factor_opt_keys))
else:
# Add unique keys to optimized keys
self.optimized_keys.extend(
opt_key
for opt_key in factor.optimized_keys
if opt_key not in self.optimized_keys
)
numeric_factors.append(factor)
# Set default params if none given
if params is None:
self.params = Optimizer.Params()
else:
self.params = params
if debug_stats is not None:
self.params.debug_stats = debug_stats
warnings.warn(
"debug_stats argument is deprecated, use params.debug_stats", FutureWarning
)
if include_jacobians is not None:
self.params.include_jacobians = include_jacobians
warnings.warn(
"include_jacobians argument is deprecated, use params.include_jacobians",
FutureWarning,
)
self._initialized = False
# Create a mapping from python identifier string keys to fixed-size C++ Key objects
# Initialize the keys map with the optimized keys, which are needed to construct the factors
# This works because the factors maintain a reference to this, so everything is fine as long
# as the unoptimized keys are also in here before we attempt to linearize any of the factors
self._cc_keys_map = {key: cc_sym.Key("x", i) for i, key in enumerate(self.optimized_keys)}
# create the mapping from cc_keys back into python keys
self._py_keys_from_cc_keys_map = {v: k for k, v in self._cc_keys_map.items()}
# This stores the list of keys in the python Values, which are necessary for reconstructing
# a Python Values from C++, in particular for methods that don't otherwise have a Python
# Values available. It's filled out in `_initialize`. The order is important here, which
# why we can't just pull the keys out of `_cc_keys_map`, which is constructed out-of-order.
self.values_keys_ordered: T.Optional[T.List[str]] = None
# Construct the C++ optimizer
self._cc_optimizer = cc_sym.Optimizer(
optimizer_params_t(**dataclasses.asdict(self.params)),
[factor.cc_factor(self._cc_keys_map) for factor in numeric_factors],
)
def _initialize(self, values: Values) -> None:
# Add unoptimized keys into the keys map
for i, key in enumerate(values.keys_recursive()):
if key not in self._cc_keys_map:
# Give these a different name (`v`) so we don't have to deal with numbering
self._cc_keys_map[key] = cc_sym.Key("v", i)
self.values_keys_ordered = list(values.keys_recursive())
self._initialized = True
def _cc_values(self, values: Values) -> cc_sym.Values:
"""
Create a cc_sym.Values from the given Python Values
This uses the stored cc_keys_map, which will be initialized if it does not exist yet.
"""
values = values.to_numerical()
if not self._initialized:
self._initialize(values)
cc_values = cc_sym.Values()
for key, cc_key in self._cc_keys_map.items():
cc_values.set(cc_key, values[key])
return cc_values
[docs] def compute_all_covariances(self, optimized_value: Values) -> T.Dict[str, np.ndarray]:
"""
Compute the covariance matrix (J^T@J)^-1 for all optimized keys about a given linearization point
Args:
optimized_value: A value containing the linearization point to compute the covariance matrix about
Returns:
A dict of {optimized_key: numerical covariance matrix}
"""
cc_covariance_dict = self._cc_optimizer.compute_all_covariances(
linearization=self.linearize(optimized_value)
)
return {self._py_keys_from_cc_keys_map[k]: v for k, v in cc_covariance_dict.items()}
[docs] def compute_covariances(
self, optimized_value: Values, keys: T.Sequence[str]
) -> T.Dict[str, np.ndarray]:
"""
Get covariances for the given subset of keys at the given linearization
This version is potentially much more efficient than computing the covariances for all
keys in the problem.
Currently requires that `keys` corresponds to a set of keys at the start of the list of keys
for the full problem, and in the same order. It uses the Schur complement trick, so will be
most efficient if the hessian is of the following form, with C block diagonal::
A = ( B E )
( E^T C )
Args:
optimized_value: A value containing the linearization point to compute the covariance matrix about
keys: The subset of keys to compute covariances for
Returns:
A dict of {optimized_key: numerical covariance matrix}
"""
cc_covariance_dict = self._cc_optimizer.compute_covariances(
linearization=self.linearize(optimized_value),
keys=[self._cc_keys_map[key] for key in keys],
)
return {self._py_keys_from_cc_keys_map[k]: v for k, v in cc_covariance_dict.items()}
[docs] def compute_full_covariance(self, optimized_value: Values) -> np.ndarray:
"""
Get the full problem covariance at the given linearization
Unlike compute_covariance and compute_all_covariances, this includes the off-diagonal
blocks, i.e. the cross-covariances between different keys.
The ordering of entries here is the same as the ordering of the keys in the linearization,
which can be accessed via linearization_index().
May not be called before either optimize or linearize has been called.
"""
return self._cc_optimizer.compute_full_covariance(self.linearize(optimized_value))
[docs] def optimize(self, initial_guess: Values, **kwargs: T.Any) -> Optimizer.Result:
"""
Optimize from the given initial guess, and return the optimized Values and stats
Args:
initial_guess: A Values containing the initial guess, should contain at least all the
keys required by the ``factors`` passed to the constructor
num_iterations: If < 0 (the default), uses the number of iterations specified by the
params at construction
populate_best_linearization: If true, the linearization at the best values will be
filled out in the stats
Returns:
The optimization results, with additional stats and debug information. See the
:class:`Optimizer.Result` documentation for more information
"""
cc_values = self._cc_values(initial_guess)
try:
stats = self._cc_optimizer.optimize(cc_values, **kwargs)
except ZeroDivisionError as ex:
raise ZeroDivisionError("ERROR: Division by zero - check your use of epsilon!") from ex
optimized_values = Values(
**{
key: cc_values.at(self._cc_keys_map[key])
for key in initial_guess.dataclasses_to_values().keys_recursive()
}
)
return Optimizer.Result(
initial_values=initial_guess, optimized_values=optimized_values, _stats=stats
)
[docs] def linearize(self, values: Values) -> cc_sym.Linearization:
"""
Compute and return the linearization at the given Values
"""
return self._cc_optimizer.linearize(self._cc_values(values))
[docs] def load_iteration_values(self, values_msg: values_t) -> Values:
"""
Load a ``values_t`` message into a Python :class:`Values <symforce.values.values.Values>`
by first creating a C++ Values, then converting back to the python key names.
"""
cc_values = cc_sym.Values(values_msg)
assert self.values_keys_ordered is not None
py_values = Values(
**{key: cc_values.at(self._cc_keys_map[key]) for key in self.values_keys_ordered}
)
return py_values
[docs] def linearization_index(self) -> T.Dict[str, index_entry_t]:
"""
Get the index mapping keys to their positions in the linearized state vector. Useful for
extracting blocks from the problem jacobian, hessian, or RHS
Returns: The index for the Optimizer's problem linearization
"""
index = self._cc_optimizer.linearization_index()
return {key: index[self._cc_keys_map[key]] for key in self.optimized_keys}
[docs] def linearization_index_entry(self, key: str) -> index_entry_t:
"""
Get the index entry for a given key in the linearized state vector. Useful for extracting
blocks from the problem jacobian, hessian, or RHS
Args:
key: The string key for a variable in the Python Values
Returns: The index entry for the variable in the Optimizer's problem linearization
"""
return self._cc_optimizer.linearization_index_entry(self._cc_keys_map[key])