sparho

Nonsmooth bilevel hyperparameter optimization via implicit differentiation.

A maintained, performant successor to sparse-ho (ICML 2020, dormant since 2022). Tunes hyperparameters of non-smooth estimators (Lasso, ElasticNet, weighted Lasso, sparse logistic regression) by computing the hypergradient via implicit differentiation rather than grid/random search.

Status: pre-alpha. Public API may change between minor versions until v1.0.

Why

Implicit-differentiation HP optimization can be orders of magnitude faster than LassoCV-style grid search when you have a held-out criterion, but the existing libraries are dormant (sparse-ho) or no longer maintained (JAXopt). sparho is a clean-break, scipy-stack-native Rust+Python implementation built for the same target audience.

Perf summary

v0.2 numbers (HOAG + warm-start + celer inner solver), single-threaded on an Apple M-series; see benchmarks/README.md for the methodology and the v0.1 historical row.

What v0.2 delivers on top of v0.1:

• hoag_search outer loop (Pedregosa 2016): adaptive step from a Lipschitz proxy, +C·tol slack acceptance, exponentially-decreasing inner-tol schedule. Replaces LineSearch.
• Inner-solver warm-starting threaded through the Solver Protocol + every adapter + CrossVal(warm_start=True).
• celer adapter recommended for the high-d regime — compounding the HOAG/warm-start win into 32.8× on leukemia and 1.65× faster than sklearn on rcv1.binary.
• Dense-matvec fix in implicit_forward (no coo_tocsr round-trip on dense designs): 8.4× faster hypergradient solve on leukemia.

Everything v0.1 delivered still holds: gradient-based outer loop with full FD parity, vector-α (WeightedL1) which LassoCV cannot do, ridge-stabilized hypergradient-CG on ill-conditioned sparse designs, clean Protocol-based API, mypy strict + clippy clean, single wheel via maturin.

The rcv1 story. Implicit differentiation lets sparho search past LassoCV's discrete grid: on rcv1.binary, sparho's outer loop drives α down to 2.1·10⁻⁵, well below LassoCV's grid floor of 1·10⁻⁴, and lands on a 14 % better held-out MSE (0.194 vs 0.225). The wall-time gap halved at v0.2 (433 s → 211 s) thanks to warm-start + celer; the remaining gap is irreducible inner-solver work at very small α / large active set. sparho's win on this dataset is quality per outer iter, not raw wall time.

Install (planned)

pip install sparho               # release wheel, no Rust toolchain needed
pip install "sparho[celer]"      # add celer as a fast Lasso adapter

Quickstart (planned API)

from sparho import Problem, grad_search
from sparho.adapters import SklearnLasso
from sparho.criteria import held_out_mse
from sparho.optimizer import grad_descent
from sparho.hypergrad import implicit_forward

problem = Problem.lasso(X, y)
result = grad_search(
    problem,
    hp0=1e-3,
    solver=SklearnLasso(),
    hypergrad=implicit_forward,
    criterion=held_out_mse(idx_train, idx_val, X, y),
    optimizer=grad_descent(lr=1.0),
)
print(result.best_hyperparam, result.best_coef)

See docs/migration_from_sparse_ho.md for translation from sparse-ho's API.

Design

• One Problem dataclass. No abstract base class tower. Algorithms are free functions over Problem. Typing via typing.Protocol.
• Implicit-only at v0.1. ImplicitForward is the only hypergradient mode.
• Sparse-X first class. CSC iterated directly in Rust; no densification.
• Rust kernels via PyO3 + maturin + ABI3. Single binary wheel, no numba.
• Clean break from sparse-ho. Migration guide rather than compat shim.

Roadmap

See ROADMAP.md. v0.1 shipped sklearn + celer + callable adapters with verified correctness and dense-high-d parity vs LassoCV. v0.2 closes the inner-solver warm-starting and hypergradient-stability gaps and lands the HOAG outer loop — 32.8× on leukemia, 2× wall on rcv1.binary. v0.3 lands the sklearn-ecosystem wrappers (LassoHO, ElasticNetHO, LogisticRegressionHO) so sparho slots into Pipeline / EconML / MLflow, the SURE / GSURE criterion for unsupervised tuning, a MultiTaskLasso / Group-L1 penalty, and adapters for skein (nonconvex weighted/group) and skglm (MCP / SCAD / SLOPE / Group / Huber / Poisson). See docs/feature_research.md for the 2026-05-20 landscape synthesis behind these picks.

License

BSD 3-Clause.