raulpl/cola

Compositional Linear Algebra

Compositional Linear Algebra (CoLA)

CoLA is a framework for scalable linear algebra, automatically exploiting the structure often found in machine learning problems and beyond. CoLA supports both PyTorch and JAX.

Installation

pip install cola-ml

Features in CoLA

• Large scale linear algebra routines for solve(A,b), eig(A), logdet(A), exp(A), trace(A), diag(A), sqrt(A).
• Provides (user extendible) compositional rules to exploit structure through multiple dispatch.
• Has memory-efficient autodiff rules for iterative algorithms.
• Works with PyTorch or JAX, supporting GPU hardware acceleration.
• Supports operators with complex numbers and low precision.
• Provides linear algebra operations for both symmetric and non-symmetric matrices.

See https://cola.readthedocs.io/en/latest/ for our full documentation and many examples.

Quick start guide

1. LinearOperators. The core object in CoLA is the LinearOperator. You can add and subtract them +, -, multiply by constants *, /, matrix multiply them @ and combine them in other ways: kron, kronsum, block_diag etc.

import jax.numpy as jnp
import cola

A = cola.ops.Diagonal(jnp.arange(5) + .1)
B = cola.ops.Dense(jnp.array([[2., 1.], [-2., 1.1], [.01, .2]]))
C = B.T @ B
D = C + 0.01 * cola.ops.I_like(C)
E = cola.ops.Kronecker(A, cola.ops.Dense(jnp.ones((2, 2))))
F = cola.ops.BlockDiag(E, D)

v = jnp.ones(F.shape[-1])
print(F @ v)

[0.2       0.2       2.2       2.2       4.2       4.2       6.2
 6.2       8.2       8.2       7.8121004 2.062    ]

2. Performing Linear Algebra. With these objects we can perform linear algebra operations even when they are very big.

print(cola.linalg.trace(F))
Q = F.T @ F + 1e-3 * cola.ops.I_like(F)
b = cola.linalg.inv(Q) @ v
print(jnp.linalg.norm(Q @ b - v))
print(cola.linalg.eig(F)[0][:5])
print(cola.sqrt(A))

31.2701
0.0010193728
[ 2.0000000e-01+0.j  0.0000000e+00+0.j  2.1999998e+00+0.j
 -1.1920929e-07+0.j  4.1999998e+00+0.j]
diag([0.31622776 1.0488088  1.4491377  1.7606816  2.0248456 ])

For many of these functions, if we know additional information about the matrices we can annotate them to enable the algorithms to run faster.

Qs = cola.SelfAdjoint(Q)
%timeit cola.linalg.inv(Q) @ v
%timeit cola.linalg.inv(Qs) @ v

3. JAX and PyTorch. We support both ML frameworks.

import torch

A = cola.ops.Dense(torch.Tensor([[1., 2.], [3., 4.]]))
print(cola.linalg.trace(cola.kron(A, A)))

import jax.numpy as jnp
A = cola.ops.Dense(jnp.array([[1., 2.], [3., 4.]]))
print(cola.linalg.trace(cola.kron(A, A)))

tensor(25.)
25.0

and both support autograd (and jit):

from jax import grad, jit, vmap

def myloss(x):
    A = cola.ops.Dense(jnp.array([[1., 2.], [3., x]]))
    return jnp.ones(2) @ cola.linalg.inv(A) @ jnp.ones(2)


g = jit(vmap(grad(myloss)))(jnp.array([.5, 10.]))
print(g)

[-0.06611571 -0.12499995]

Citing us

If you use CoLA, please cite the following paper:

Andres Potapczynski, Marc Finzi, Geoff Pleiss, and Andrew Gordon Wilson. "CoLA: Exploiting Compositional Structure for Automatic and Efficient Numerical Linear Algebra." 2023.

@article{potapczynski2023cola,
  title={{CoLA: Exploiting Compositional Structure for Automatic and Efficient Numerical Linear Algebra}},
  author={Andres Potapczynski and Marc Finzi and Geoff Pleiss and Andrew Gordon Wilson},
  journal={arXiv preprint arXiv:2309.03060},
  year={2023}
}

Features implemented

Linear Algebra	inverse	eig	diag	trace	logdet	exp	sqrt	f(A)	SVD	pseudoinverse
Implementation	✓	✓	✓	✓	✓	✓	✓	✓

LinearOperators	Diag	BlockDiag	Kronecker	KronSum	Sparse	Jacobian	Hessian	Fisher	Concatenated	Triangular	FFT	Tridiagonal
Implementation	✓	✓	✓	✓	✓	✓	✓		✓	✓		✓

Annotations	SelfAdjoint	PSD	Unitary
Implementation	✓	✓	✓

Contributing

See the contributing guidelines docs/CONTRIBUTING.md for information on submitting issues and pull requests.

CoLA is Apache 2.0 licensed.

Support and contact

Please raise an issue if you find a bug or slow performance when using CoLA.