The autodiff code generator is extended in the following ways:
- Evaluate the complete jacobian into a temporary array
- Copy the temporary jacobian to the result if != NULL
Change-Id: If5ed0bff3b31e93a2d7d5dc718862cb648dba474
Before this patch the return type was implicitly defined by the
ExpressionType. This patch separates this connection and allows
each Expression to have one of the predefined types (scalar,
boolean, void).
This patch is required to add support for the functions isfinite,
isinf, isnan, and isnormal. These are function taking a double and
returning a bool.
This also moves some complexity of the code generator to the
Expression, because the generator can direclty get the c++ type.
Change-Id: I8b32bab1bfab2f668875e506d6f3b789a5d1f3fd
- Move codegen tests to a sub directory
- Add tests for all functions of Expression, ExpressionRef, and ExpressionGraph
- Respect dependencies during tests: The ExpressionGraph test doesn't
use ExpressionRef anymore.
The new tests revealed a few bugs so the following changes were made:
- Expression::MakeNop now resets the current expression with the default
constructed NOP expression
- ExpressionGraph::Insert now updates the lhs_id the same way as
InsertBack()
Change-Id: I6a18925c1e4d972c29ec1219f2073b4eaf2df737
Objects of 'Expression' can now be freely created and copied around.
This creation does NOT add them to the active ExpressionGraph any
more. The insertion into the graph is now done by ExpressionRef,
which explicitly call graph.add(...) in each operation.
This change brings the following advantages:
1. 'Expression' is now stand-alone and side-effect free
- Remove the dependency Expression->ExpressionGraph
- Expressions can be created by the optimizer
2. Explicit graph insertion
- Previously CreateCompileTimeConstant not only created an
expression, but also inserted it into the active graph. Now
this insertion is done explicitly by ExpressionRef
3. It is now easier to insert new types and members
- Should be straight forward now, because we got rid of the
3-way dependency
- This is a preparation patch for the new return-type member in
https://ceres-solver-review.googlesource.com/c/ceres-solver/+/16224
Change-Id: Icef2fe529a4db001a10d1fb6816c9dc681b14ff2
The public API for the Autodiff Codegen system consist of a single
function, GenerateCodeForFunctor. This function takes as template
argument a cost functor type and the residual/parameter structure.
The output is the C++ for the residual and derivative. This class
mainly serves as a wrapper for the different codegen modules.
Change-Id: I8ee974199219805d54eed5e07c0e8d1394940779
User defined cost functors have to include this new header
instead of internal/expression_ref.h. This hides some
complexity of ExpressionRef and reduces compile time outside
of code generation mode.
This patch also removes the dependency ExpressionRef->Jet.
Change-Id: Ie3f93648775e14881dc5cfab213bbc983c6cfeee
The move constructor of ExpressionRef now implements the copy
elision, which was incorrectly done by the copy constructor.
This patch also updates the CMakeList to only add the
Expression tests for gcc. This is currently required, because
the AutoDiffCodeGen system relies on implementation defined
behaviour. (Evalution order + copy elision)
Change-Id: Ib17aeb22bd4d81de6d838b55c1b6497fb3740d0e
Currently there is a bug in the copy constructor of ExpressionRef.
Fixing this bug, will create new expressions each time a copy is
executed. To reduce the amount of copy expression we pass
ExpressionRef by const reference now.
Change-Id: I0f81fb3d08aa4c66b9191f3138650f17023af67e
This functions allows the insertion of new expression anywhere into
the graph. All references are updated. This will be used by semantic
expression testing and by the optimizer.
Change-Id: Ieafcfec7672f6328106dc0511c85d6fb5bd64d97
- Add operator== for comparing Expressions and ExpressionGraphs.
- Remove the broken Jet-test for now. (Will be added again later)
Change-Id: I3ee1599e8a0b50fa61c16f30812c78db7333f273
Due to the removed definition, calling this method will result
in a compile-time error instead of a run-time error.
Change-Id: Iac2514c05d79a66fcbad124587d08eb9309aa6a5
- Remove TERNARY (replaced by function call)
- Combine PARAMETER and RUNTIME_CONSTANT into INPUT_ASSIGNMENT
Change-Id: Id35fb45da08a9f80c1ac0dab7008deb0b3922835
- Add missing copy constructor
- Add make functions for input/output expressions
- Add missing ExpressionRef overload for MakeRuntimeConstant
- Add ExpressionRef overloads for common functions (cos,exp,...)
- Make ExpressionRef constructor from double implicit
- Add simple test for using ExpressionRef in Jets.
- Rename expression test macro
Change-Id: I071dc6717e4034a281662021998a91e80636da27
- Expression types for if/else blocks
- CERES_IF/ELSE macros
- Multiple assignment to the same variable
Change-Id: If529516f243f31823d1ef7b8827bb6f2390e418d
This patch adds the 'Expression' class, which is a fundamental
building block of automatic code generation. The expressions can
be used as scalar types for cost functors as well as Jets.
Dynamic branching is not yet supported.
Change-Id: I8c61bee5c307e0eec20fd39382683ea90f720dff
Now that we are using c++11, it is safer to use override in subclasses
instead of virtual.
This CL does it for the interface, a follow up CL will do it for
other parts of the code base.
Change-Id: Ice8d0f4355cb700019d7a9c1566fbff0099e97d6
https://github.com/ceres-solver/ceres-solver/issues/270
Detailed list of changes:
1. Add SUBSET to the PreconditionerType enum.
2. Add Solver::Options::residual_blocks_for_subset_preconditioner
3. Integrate SubsetPreconditioner into the CGNR solver.
4. Add the reordering logic needed for this to TrustRegionPreprocessor.
5. Expect CreateJacobianBlockTranspose to take the starting row block
so that we can work with subparts of the Jacobian matrix.
6. Extend the denoising example to use this preconditioner.
As an illustration of its performance, we consider the performance of
denoising -input ../data/ceres_noisy.pgm --foe_file ../data/5x5.foe
tl;dr
For the same cost,
SPARSE_NORMAL_CHOLESKY - 81s
CGNR + JACOBI - 718s
CGNR + SUBSET - 57s
SPARSE_NORMAL_CHOLESKY
======================
Cost:
Initial 2.317806e+05
Final 2.232323e+04
Change 2.094574e+05
Minimizer iterations 10
Successful steps 10
Unsuccessful steps 0
Time (in seconds):
Preprocessor 2.999746
Residual only evaluation 2.306811 (10)
Jacobian & residual evaluation 7.421727 (10)
Linear solver 65.517273 (10)
Minimizer 78.731011
Postprocessor 0.026079
Total 81.756836
Termination: CONVERGENCE (Function tolerance reached. |cost_change|/cost: 8.573046e-04 <= 1.000000e-03)
CGNR + JACOBI
=============
Cost:
Initial 2.317806e+05
Final 2.232344e+04
Change 2.094572e+05
Minimizer iterations 10
Successful steps 10
Unsuccessful steps 0
Time (in seconds):
Preprocessor 0.648814
Residual only evaluation 2.297607 (10)
Jacobian & residual evaluation 7.327886 (10)
Linear solver 699.601248 (10)
Minimizer 712.419493
Postprocessor 0.024014
Total 713.092321
Termination: CONVERGENCE (Function tolerance reached. |cost_change|/cost: 8.528538e-04 <= 1.000000e-03)
CGNR + SUBSET (random 20% residuals used for the preconditioner)
===============================================================
Cost:
Initial 2.317806e+05
Final 2.232327e+04
Change 2.094574e+05
Minimizer iterations 10
Successful steps 10
Unsuccessful steps 0
Time (in seconds):
Preprocessor 1.472743
Residual only evaluation 2.428315 (10)
Jacobian & residual evaluation 7.367796 (10)
Linear solver 42.585999 (10)
Minimizer 55.664459
Postprocessor 0.024098
Total 57.161301
Termination: CONVERGENCE (Function tolerance reached. |cost_change|/cost: 8.538277e-04 <= 1.000000e-03)
Change-Id: Ifb011408bd53edbb9439b0b7345649a38f999e18
- Accelerate currently does not support dynamic sparsity, whereas the
other sparse linear algebra libraries do (outstanding issue to update)
- Previously we preferred Accelerate to all but SuiteSparse if it was
available, which breaks the dynamic_sparsity_test if SuiteSparse is
*not* available (even if others are) as Accelerate does not support
dynamic sparsity.
Change-Id: Ibc2dd2f14f83cffbecca38097d02bb2188aaaa05
Some methods in Problem do not modify the parameter block
and those methods now allow the user to call them with const double*.
The methods are
RemoveParameterBlock
SetParameterBlockConstant
IsParameterBlockConstant
GetParameterization
GetParameterLowerBound
GetParameterUpperBound
https://github.com/ceres-solver/ceres-solver/issues/479
Change-Id: I59dcb77134f59576dd498bd732e29aae9abd28b1
The implicitly-declared copy constructor and copy assignment operator
of ProductParameterization are not deleted since the corresponding
functions of the member vector are defined, though ill-formed if they
are instantiated.
With dllexport, the compiler must generate all of the member functions
of the class, including those are implicitly-declared, so we need
explicitly delete the ill-formed ones.
Change-Id: Iaa763dfaefdc74a9840f244ba05c6bfd0bd99492
This change only affects the TRUST_REGION minimizer and has no effect
on the LINE_SEARCH minimizer.
This options controls the number of iterations that the line search
algorithm performs. The line search algorithm is as the name implies,
used in the LINE_SEARCH minimizer. It is also used by the TRUST_REGION
minimizer when solving bounds constrained optimization problems.
In some bounds constrained problems, it is enough to project each step
onto the bounds constraints and not perform the line search. This can
have a significant impact on runtime. Setting
Solver::Options::max_num_line_search_step_size_iterations = 0 enables
this functionality.
Changchang Wu (ccwu@google.com) came up with the idea for this
implementation.
https://github.com/ceres-solver/ceres-solver/issues/477
Change-Id: Ifbe0bc5c48eedb2c1231d43cd98e4be7316c0682
1. ProblemEvaluateResidualTest was leaking the loss_function in cases
where it was not being used.
2. Fix a grammo in rotation.h
Change-Id: If94ae1624033c8f6d1934dc2f40fa0dfe4e025c3
This method gives the user the ability to evaluate a given residual
block.
A couple of minor cleanups.
Problem::problem_impl_ -> Problem::impl_
NULL -> nullptr
https://github.com/ceres-solver/ceres-solver/issues/417
Change-Id: I6dd94762c475fa264c387b8c93d516f6e06fe832
This commit includes the following:
- Changes to CMake to make it safer to see which compiler flags are supported,
so this way we do not need to worry about version checks in CMake.
- Unix platforms (which includes both Linux and Apple as far as i can tell)
will now enable -Wmissing-declarations warning for the whole Ceres.
- Changes in all sources to solve missing declaration warning. In most cases
it was either matter of using static qualifier or moving functions to an
anonymous namespace.
In one case the function got removed, since it seems to be unused.
Additionally, in slam examples there was a non-inlined function implementation
in a header, which is a direct way to cause linking errors if other .cc file
will include that helper header.
- All third party sources (which is currently only gmock) has this extra
paranoid warning disabled.
This warning is important in the following cases:
- Detect helper functions which are not needed anymore.
- Avoid unnoticed pollution of namespace.
- Avoid bad level calls.
- Avoid missing updates in header files after changes in implementation file.
- Helps integrating Ceres into software where paranoid warnings are important.
Change-Id: I9b1044aced3910d8c6b2356cfe2bf57f3c8c58db
This CL modernizes ProductParameterization in the following ways:
- It uses std::unique_ptr for memory handling instead of using raw
pointers and handmade memory management.
- Replaces the constructors with a variadic template.
Change-Id: I5c9fe42ac935b6c26e867dbd3369a4c766623047
Some rotation functions like UnitQuaternionRotatePoint,
QuaternionRotatePoint, etc. will calculate bad results if the input
point and output point points to the same memory (inplace operation).
This CL adds checks in debug mode to guard against it.
Change-Id: Id0a30e9a0286b340757f0790d417d9f9a3409810
For proper alignment on the heap Eigen needs to have a custom
allocator. There are two forms, new and in-place new. To make sure
that memory is aligned using new, one needs to overload new by
adding EIGEN_MAKE_ALIGNED_OPERATOR_NEW to any struct which contains a
fixed size Eigen type either through inheritance or as a direct or
indirect member. For the in-place new one need to use the
Eigen::aligned_allocator (e.g. for std::vector, std::list,
FixedArray, etc.). For more details see:
https://eigen.tuxfamily.org/dox/group__DenseMatrixManipulation__Alignement.html
This CL adds EIGEN_MAKE_ALIGNED_OPERATOR_NEW to all structs, which
contain fixed-size Eigen types and uses the Eigen::aligned_allocator
for containers which stores structs of fixed-size Eigen types.
Change-Id: I06c6c4fc74a6835918d5d1c571b7814a14c029d8
This PR changes the implementation of the current fixed array to the
abseil one, which has proper allocator support.
Some minor changes are made to make the fixed array implementation
self-contained (no dependent to abseil):
- No address sanitizer support (red zones, etc.)
- Remove of noexecpt specified for copy and move constructor.
- Remove of 'at' function as Ceres does not use exceptions.
- Use std::tuple instead of absl::CompressedTuple as it uses the abseil
utility header which includes a whole bunch of other headers.
Change-Id: I43445b42c37f944509b5353a587d0efce74cbccf
This is to FirstOrderFunction, what AutoDiffCostFunction is to CostFunction.
This allows users of GradientSolver to be able to define objective functions
without requiring them to define the derivatives.
The implementation uses the same Jet objects for computing the gradient as
is used by AutoDiffCostFunction.
Change-Id: Ide6e60532a3adab9be9899ba9b368dc267fd2dbb
This CL changes the implementation of AddResidualBlock() in
ceres::Problem using variadic templates. Also one new overload for
AddResidualBlock() is added using a double** and the number of
parameter blocks.
Change-Id: I007a82a06897335a117213a0d12fedb4a77076a0
This CL fixes a regression bug in numeric differentiation where the
differentiation is called even if the parameter block is hold constant.
This resulted into a write to a nullptr.
A unit test is added for Jacobian evaluation of constant parameters.
Change-Id: Ia0f7c6cc7ef18f0f2cd6d758a839729b8ff606a0
This CL simplifies the implementation of cost function to functor using
variadic templates. The cost functor can now be called with an
arbitrary number of parameter pointers.
Change-Id: I0d9df2e9055a4842f40c232d4474508f010089a4
This PR changes the interface of sized_cost_fucntion,
autodiff_cost_function and numeric_diff_costfunction from using ten
hardcoded parameter blocks to a variable number of parameter blocks
using variadic templates.
Trailing parameter blocks of size zero are now considered as error.
Change-Id: I37b9a0a420ef0eda6476a46672bbf6bd57e19760
Sameer Agarwal