v0.2.0
Sparse Jacobians are frequently encountered in the simulation of physical systems. Jax tranformations jacfwd
and jacrev
make it easy to compute dense Jacobians, but these are wasteful when the Jacobian is sparse. sparsejac
provides a function to more efficiently compute the Jacobian if its sparsity is known. It makes use of the recently-introduced jax.experimental.sparse
module.
pip install sparsejac
A trivial example with a diagonal Jacobian follows:
fn = lambda x: x**2
x = jax.random.uniform(jax.random.PRNGKey(0), shape=(10000,))
@jax.jit
def sparse_jacrev_fn(x):
with jax.ensure_compile_time_eval():
sparsity = jax.experimental.sparse.BCOO.fromdense(jnp.eye(10000))
jacrev_fn = sparsejac.jacrev(fn, sparsity=sparsity)
return jacrev_fn(x)
dense_jacrev_fn = jax.jit(jax.jacrev(fn))
assert jnp.all(sparse_jacrev_fn(x).todense() == dense_jacrev_fn(x))
%timeit sparse_jacrev_fn(x).block_until_ready()
%timeit dense_jacrev_fn(x).block_until_ready()
And, the performance improvement can easily be seen:
93.1 µs ± 17.2 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
182 ms ± 26.9 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
- In general, it is preferable to directly provide the sparsity, rather than obtaining it from a dense matrix.
- GPU may show minimal or no performance advantage over CPU.
- Users are encouraged to test
jacrev
andjacfwd
on their specific problem to select the most performant option.