Source code for symforce.opt.numeric_factor

# ----------------------------------------------------------------------------
# SymForce - Copyright 2022, Skydio, Inc.
# This source code is under the Apache 2.0 license found in the LICENSE file.
# ----------------------------------------------------------------------------
from __future__ import annotations

from pathlib import Path

import numpy as np

from symforce import cc_sym
from symforce import typing as T
from symforce.codegen import codegen_util
from symforce.values import Values


[docs]class NumericFactor: """ A class used to wrap linearization functions such that they can be used by the optimizer. Args: keys: The set of keys that are inputs to the linearization function. optimized_keys: A subset of ``keys`` representing the keys which the given linearization function computes the jacobian with respect to. linearization_function: A function that returns the residual, jacobian, hessian approximation, and right-hand-side used with the levenberg marquardt optimizer. """ def __init__( self, keys: T.Sequence[str], optimized_keys: T.Sequence[str], linearization_function: T.Callable[ ..., T.Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray] ], ) -> None: self.keys = keys self.optimized_keys = optimized_keys self.linearization_function = linearization_function
[docs] @classmethod def from_file_python( cls, keys: T.Sequence[str], optimized_keys: T.Sequence[str], output_dir: T.Openable, namespace: str, name: str, ) -> NumericFactor: """ Returns a NumericFactor constructed from the python function ``name`` from the module located at ``output_dir / "python" / "symforce" / namespace / f"{name}.py"`` (this matches the directory structure created by :meth:`.factor.Factor.generate`). This can be used after generating a linearization function from a symbolic factor as follows: Create a symbolic factor and generate the linearization function:: output_dir = "my_output_dir" namespace = "my_namespace" name = "my_custom_factor" sym_factor = Factor( keys=my_keys, residual=my_func, name=name, ) sym_factor.generate(my_optimized_keys, output_dir, namespace) Load the generated linearization function:: num_factor = NumericFactor.from_file_python( my_keys, my_optimized_keys, output_dir, namespace, name ) Args: keys: The set of keys that are inputs to the linearization function. optimized_keys: A subset of ``keys`` representing the keys which the given linearization function computes the jacobian with respect to. output_dir: The top-level output directory of the linearization function. namespace: The namespace of the linearization function. name: The name of the linearization function. """ assert all(opt_key in keys for opt_key in optimized_keys) function_dir = Path(output_dir) / "python" / "symforce" / namespace linearization_function = codegen_util.load_generated_function(name, function_dir) return cls( keys=keys, optimized_keys=optimized_keys, linearization_function=linearization_function )
[docs] def linearize(self, inputs: Values) -> T.Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]: """ Evaluates the linearization function for the given inputs. Returns the residual, jacobian, hessian approximation, and right hand side used with the levenberg marquardt optimizer. Args: inputs: Values object that does not contain any symbolic members and is ordered the same as the arguments to the linearization function. """ if inputs.keys_recursive() != self.keys: raise ValueError("Keys in inputs must match keys used to construct the factor.") residual, jacobian, hessian, rhs = self.linearization_function( *inputs.to_numerical().values_recursive() ) # Check dimensions. We require the jacobian has 2 dimensions, because otherwise it may be a # row vector that gets converted to a column vector by pybind if jacobian.ndim != 2: raise ValueError( f"Jacobian must have 2 dimensions, got {jacobian.ndim}. If the linearization " "function was generated by SymForce, make sure to set return_2d_vectors=True" ) return residual, jacobian, hessian, rhs
[docs] def cc_factor(self, cc_key_map: T.Mapping[str, cc_sym.Key]) -> cc_sym.Factor: """ Create a C++ Factor from this symbolic Factor, for use with the C++ Optimizer Note that while this is a C++ Factor object, the linearization function may be a compiled C++ function or a Python function passed into C++ through pybind, depending on the language the linearization function was generated in. Args: cc_key_map: Mapping from Python keys (strings, like returned by :meth:`Values.keys_recursive <symforce.values.values.Values.keys_recursive>` ) to C++ keys Returns: A C++ wrapped Factor object """ def wrapped( values: cc_sym.Values, _: T.Any ) -> T.Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]: return self.linearization_function(*[values.at(cc_key_map[key]) for key in self.keys]) return cc_sym.Factor(wrapped, [cc_key_map[key] for key in self.optimized_keys])