diffcp
is a Python package for computing the derivative of a convex cone program, with respect to its problem data. The derivative is implemented as an abstract linear map, with methods for its forward application and its adjoint.
The implementation is based on the calculations in our paper Differentiating through a cone program.
diffcp
is available on PyPI, as a source distribution. Install it with
pip install diffcp
You will need a C++11-capable compiler to build diffcp
.
diffcp
requires:
- NumPy >= 1.15
- SciPy >= 1.10
- SCS >= 2.0.2
- pybind11 >= 2.4
- threadpoolctl >= 1.1
- Python 3.x
diffcp
uses Eigen; Eigen operations can be automatically vectorized by compilers. To enable vectorization, install with
MARCH_NATIVE=1 pip install diffcp
OpenMP can be enabled by passing extra arguments to your compiler. For example, on linux, you can tell gcc to activate the OpenMP extension by specifying the flag "-fopenmp":
OPENMP_FLAG="-fopenmp" pip install diffcp
To enable both vectorization and OpenMP (on linux), use
MARCH_NATIVE=1 OPENMP_FLAG="-fopenmp" pip install diffcp
diffcp
differentiates through a primal-dual cone program pair. The primal problem must be expressed as
minimize c'x
subject to Ax + s = b
s in K
where x
and s
are variables, A
, b
and c
are the user-supplied problem data, and K
is a user-defined convex cone. The corresponding dual problem is
minimize b'y
subject to A'y + c == 0
y in K^*
with dual variable y
.
diffcp
exposes the function
solve_and_derivative(A, b, c, cone_dict, warm_start=None, **kwargs).
This function returns a primal-dual solution x
, y
, and s
, along with
functions for evaluating the derivative and its adjoint (transpose).
These functions respectively compute right and left multiplication of the derivative
of the solution map at A
, b
, and c
by a vector.
In the case that the problem is not solved, i.e. SCS returns something
other than "Solved" or "Solved/Innacurate" for status, we raise
a SolverError
Exception.
The arguments A
, b
, and c
correspond to the problem data of a cone program.
A
must be a SciPy sparse CSC matrix.b
andc
must be NumPy arrays.cone_dict
is a dictionary that defines the convex coneK
.warm_start
is an optional tuple(x, y, s)
at which to warm-start SCS.**kwargs
are keyword arguments to forward to SCS (e.g.,verbose=True
).
These inputs must conform to the SCS convention for problem data. The keys in cone_dict
correspond to the cones, with
diffcp.ZERO
for the zero cone,diffcp.POS
for the positive orthant,diffcp.SOC
for a product of SOC cones,diffcp.PSD
for a product of PSD cones, anddiffcp.EXP
for a product of exponential cones.
The values in cone_dict
denote the sizes of each cone; the values of diffcp.SOC
, diffcp.PSD
, and diffcp.EXP
should be lists. The order of the rows of A
must match the ordering of the cones given above. For more details, consult the SCS documentation.
The function solve_and_derivative
returns a tuple
(x, y, s, derivative, adjoint_derivative)
-
x
,y
, ands
are a primal-dual solution. -
derivative
is a function that applies the derivative at(A, b, c)
to perturbationsdA
,db
,dc
. It has the signaturederivative(dA, db, dc) -> dx, dy, ds
, wheredA
is a SciPy sparse CSC matrix with the same sparsity pattern asA
, anddb
anddc
are NumPy arrays.dx
,dy
, andds
are NumPy arrays, approximating the change in the primal-dual solution due to the perturbation. -
adjoint_derivative
is a function that applies the adjoint of the derivative to perturbationsdx
,dy
,ds
. It has the signatureadjoint_derivative(dx, dy, ds) -> dA, db, dc
, wheredx
,dy
, andds
are NumPy arrays.
import numpy as np
from scipy import sparse
import diffcp
cone_dict = {
diffcp.ZERO: 3,
diffcp.POS: 3,
diffcp.SOC: [5]
}
m = 3 + 3 + 5
n = 5
A, b, c = diffcp.utils.random_cone_prog(m, n, cone_dict)
x, y, s, D, DT = diffcp.solve_and_derivative(A, b, c, cone_dict)
# evaluate the derivative
nonzeros = A.nonzero()
data = 1e-4 * np.random.randn(A.size)
dA = sparse.csc_matrix((data, nonzeros), shape=A.shape)
db = 1e-4 * np.random.randn(m)
dc = 1e-4 * np.random.randn(n)
dx, dy, ds = D(dA, db, dc)
# evaluate the adjoint of the derivative
dx = c
dy = np.zeros(m)
ds = np.zeros(m)
dA, db, dc = DT(dx, dy, ds)
For more examples, including the SDP example described in the paper, see the examples
directory.
If you wish to cite diffcp
, please use the following BibTex:
@article{diffcp2019,
author = {Agrawal, A. and Barratt, S. and Boyd, S. and Busseti, E. and Moursi, W.},
title = {Differentiating through a Cone Program},
journal = {Journal of Applied and Numerical Optimization},
year = {2019},
volume = {1},
number = {2},
pages = {107--115},
}
@misc{diffcp,
author = {Agrawal, A. and Barratt, S. and Boyd, S. and Busseti, E. and Moursi, W.},
title = {{diffcp}: differentiating through a cone program, version 1.0},
howpublished = {\url{https://github.com/cvxgrp/diffcp}},
year = 2019
}
The following thesis concurrently derived the mathematics behind differentiating cone programs.
@phdthesis{amos2019differentiable,
author = {Brandon Amos},
title = {{Differentiable Optimization-Based Modeling for Machine Learning}},
school = {Carnegie Mellon University},
year = 2019,
month = May,
}