Installation

pip install gradhpo

Requires Python ≥ 3.9. JAX, optax, scikit-learn and the rest of the runtime dependencies are pulled in automatically.

Source install:

git clone https://github.com/intsystems/gradhpo.git
pip install ./gradhpo/src

Editable / dev install (recommended for contributors):

git clone https://github.com/intsystems/gradhpo.git
cd gradhpo
pip install -e ./src
pip install pytest pytest-cov flake8

Quick start

import jax
import jax.numpy as jnp
from gradhpo import OnlineHypergradientOptimizer

def loss_fn(params, hyperparams, batch):
    x, y = batch
    pred = x @ params['w']
    mse = jnp.mean((pred - y) ** 2)
    reg = jax.nn.softplus(hyperparams['log_lam']) * jnp.sum(params['w'] ** 2)
    return mse + reg

def update_fn(w, lam, batch):
    g = jax.grad(loss_fn)(w, lam, batch)
    return jax.tree.map(lambda p, gp: p - 0.01 * gp, w, g)

opt = OnlineHypergradientOptimizer(
    update_fn=update_fn, gamma=0.99, estimation_period=10, T=20,
)
state = opt.init({'w': jnp.zeros(10)}, {'log_lam': jnp.array(0.0)})

state = opt.run(
    state, M=30,
    get_train_batch=get_train, get_val_batch=get_val,
    train_loss_fn=loss_fn, val_loss_fn=loss_fn,
    lr_hyper=1e-3,
)

The same interface works for T1T2Optimizer, GreedyOptimizer, FOOptimizer and OneStepOptimizer. See the documentation for a side-by-side comparison and a full notebook.

Documentation

• Full docs: https://intsystems.github.io/gradhpo/

• API reference: BilevelOptimizer, BilevelState, all algorithms, pytree/VJP utilities.

• Tutorial with a 2-layer MLP and a per-parameter learning rate vector.

Project information

• Source: https://github.com/intsystems/gradhpo

• Issue tracker: https://github.com/intsystems/gradhpo/issues

• License: MIT

Citation

If you use gradhpo in academic work, please cite:

Eynullayev, A., Rubtsov, D., & Karpeev, G. (2026).
gradhpo: Gradient-Based Hyperparameter Optimization.
MIPT Intelligent Systems.