TinyAD is a C++ header-only library for second-order automatic differentiation. Small dense problems are differentiated in forward mode, which allows unrestricted looping and branching. An interface for per-element functions allows convenient differentiation of large sparse problems, which are typical in geometry processing on meshes. For more details see our paper or watch our talk.
TinyAD has been tested on Linux, Mac, and Windows (VS >= 2017). It only requires:
- A C++17 compiler
- Eigen (e.g.
sudo apt-get install libeigen3-dev)
To use TinyAD in your existing project, include either TinyAD/Scalar.hh, TinyAD/ScalarFunction.hh, or TinyAD/VectorFunction.hh.
A minimal example project using TinyAD with libigl is available here.
We provide the scalar type TinyAD::Double<k> as a drop-in replacement for double. For small problems, simply choose the number of variables k and generate a vector of active variables. Then, perform computations as usual (e.g. using Eigen) and query the gradient and Hessian of any intermediate variable:
#include <TinyAD/Scalar.hh>
// Choose autodiff scalar type for 3 variables
using ADouble = TinyAD::Double<3>;
// Init a 3D vector of active variables and a 3D vector of passive variables
Eigen::Vector3<ADouble> x = ADouble::make_active({0.0, -1.0, 1.0});
Eigen::Vector3<double> y(2.0, 3.0, 5.0);
// Compute angle using Eigen functions and retrieve gradient and Hessian w.r.t. x
ADouble angle = acos(x.dot(y) / (x.norm() * y.norm()));
Eigen::Vector3d g = angle.grad;
Eigen::Matrix3d H = angle.Hess;All derivative computations are inlined and thus available for compiler optimization. As no taping is needed in forward mode, any kind of run time branching is possible.
Sparse problems on meshes can be implemented using our ScalarFunction or VectorFunction interfaces. Just pass a set of variable handles, a set of element handles, and a lambda function to be evaluated for each element. For example, in a planar parametrization problem, the variables are 2D positions per vertex, and the summands of the objective function are defined per face, each accessing 3 vertices:
#include <TinyAD/ScalarFunction.hh>
// Set up a function with 2D vertex positions as variables
auto func = TinyAD::scalar_function<2>(mesh.vertices());
// Add an objective term per triangle. Each connecting 3 vertices
func.add_elements<3>(mesh.faces(), [&] (auto& element)
{
// Element is evaluated with either double or TinyAD::Double<6>
using T = TINYAD_SCALAR_TYPE(element);
// Get variable 2D vertex positions of triangle t
OpenMesh::SmartFaceHandle t = element.handle;
Eigen::Vector2<T> a = element.variables(t.halfedge().to());
Eigen::Vector2<T> b = element.variables(t.halfedge().next().to());
Eigen::Vector2<T> c = element.variables(t.halfedge().from());
return ...
});
// Evaluate the funcion using any of these methods:
double f = func.eval(x);
auto [f, g] = func.eval_with_gradient(x);
auto [f, g, H] = func.eval_with_derivatives(x);
auto [f, g, H_proj] = func.eval_with_hessian_proj(x);
...Handle types from multiple mesh data structures are supported, e.g., OpenMesh, polymesh, geometry-central, or libigl-style matrices. Support for new types can be added by overloading a single function (see TinyAD/Support/Common.hh).
To get started, take a look at one of our TinyAD-Examples. We implement objective functions and basic solvers for typical geometry processing tasks using various mesh libraries.
We compute a piecewise linear map from a disk-topology triangle mesh to the plane and optimize the symmetric Dirichlet energy via a Projected-Newton solver. This can be the basis to experiment with more specialized algorithms or more complex objective functions. We provide examples using different mesh representations:
parametrization_openmesh.cc
parametrization_polymesh.cc
parametrization_geometrycentral.cc
parametrization_libigl.cc
In this example, we compute a 3D deformation of a tetrahedral mesh by optimizing different distortion energies subject to position constraints:
Here, we show how to re-implement the non-linear frame field optimization algorithm presented in Integrable PolyVector Fields [Diamanti et al. 2015], using very little code. Given an input frame field (two tangent vectors per triangle), the algorithm optimizes an objective based on complex polynomials via a Gauss-Newton method:
We optimize a map from a genus 0 surface to the sphere using a technique from manifold optimization. Vertex trajectories on the sphere are parametrized via tangent vectors and a retraction operator:
In this example, we optimize the 3D vertex positions of a quad mesh for face planarity. We implement one of the objective terms from Geometric Modeling with Conical Meshes and Developable Surfaces [Liu 2006]:
- Internal floating point types other than
doublecan be used viaTinyAD::Scalar<k, T>. - A gradient-only mode is availabe via
TinyAD::Scalar<k, T, false>. - Use
to_passive(...)to explicitly cast an active variable back to its scalar type without derivatives. E.g. to implement assertions or branching which should not be differentiated. - Avoid using the
autokeyword when when working with Eigen expressions. This is a limitation of Eigen and can produce unexpected results due to the deleted temporary objects. - Use e.g.
cos(...)instead ofstd::cos(...). - A common source for errors in the implementation of objective functions (per-element lambdas passed to
func.add_elements(...)) are multiple return statements of different types. This may lead to a compiler error, but can be prevented by explicitly stating the correct return type via:
func.add_elements<...>(..., [&] (auto& element) -> TINYAD_SCALAR_TYPE(element) { return ... }); - Note that calls to math functions involving TinyAD types are only legal if the derivatives exist and are finite for the given function argument. E.g. it is illegal to call
acos(x)withx==1.0since the derivative of acos is unbounded at 1.0.
When contributing to TinyAD, please run and extend the unit tests located in TinyAD/tests.
You can build and run the tests via:
mkdir build
cd build
cmake -DTINYAD_UNIT_TESTS=ON ..
make -j4
./TinyAD-Tests
Alternatively, you can use the TinyAD-Examples project which builds the unit tests by default.
We thank all test users and contributors. In particular: Alexandra Heuschling, Anton Florey, Dörte Pieper, Joe Jakobi, Philipp Domagalski, and David Jourdan.
If you use TinyAD in your academic work, please cite our paper:
@article{schmidt2022tinyad,
title={{TinyAD}: Automatic Differentiation in Geometry Processing Made Simple},
author={Schmidt, Patrick and Born, Janis and Bommes, David and Campen, Marcel and Kobbelt, Leif},
year={2022},
journal={Computer Graphics Forum},
volume={41},
number={5},
}
TinyAD is released under the MIT license.