Merge pull request #65 from ziatdinovmax/optimize_acq

Acquisition function optimization

gpax/acquisition/__init__.py

@@ -1,4 +1,5 @@
 from .acquisition import UCB, EI, POI, UE, Thompson, KG
 from .batch_acquisition import qEI, qPOI, qUCB, qKG
+from .optimize import optimize_acq
 
-__all__ = ["UCB", "EI", "POI", "UE", "KG", "Thompson", "qEI", "qPOI", "qUCB", "qKG"]
+__all__ = ["UCB", "EI", "POI", "UE", "KG", "Thompson", "qEI", "qPOI", "qUCB", "qKG", "optimize_acq"]

gpax/acquisition/optimize.py

@@ -0,0 +1,97 @@
+"""
+optimize.py
+==============
+
+Optimize continuous acquisition functions
+
+Created by Maxim Ziatdinov (email: [email protected])
+"""
+
+from typing import Type, Callable, Union, List, Tuple
+
+import jax.numpy as jnp
+import jax.random as jra
+import numpy as onp
+
+from ..models.gp import ExactGP
+
+
+def optimize_acq(rng_key: jnp.ndarray,
+                 model: Type[ExactGP],
+                 acq_fn: Callable,
+                 num_initial_guesses: int,
+                 lower_bound: Union[List, Tuple, float, onp.ndarray, jnp.ndarray],
+                 upper_bound: Union[List, Tuple, float, onp.ndarray, jnp.ndarray],
+                 **kwargs) -> jnp.ndarray:
+    """
+    Optimizes an acquisition function for a given Gaussian Process model using the JAXopt library.
+
+    This function finds the point that maximizes the acquisition function within the specified bounds.
+    It uses L-BFGS-B algorithm through ScipyBoundedMinimize from JAXopt.
+
+    Args:
+        rng_key: A JAX random key for stochastic processes.
+        model: The Gaussian Process model to be used.
+        acq_fn: The acquisition function to be maximized.
+        num_initial_guesses: Number of random initial guesses for the optimization.
+        lower_bound: Lower bounds for the optimization.
+        upper_bound: Upper bounds for the optimization.
+        **kwargs: Additional keyword arguments to be passed to the acquisition function.
+
+    Returns:
+        Parameter(s) that maximize the acquisition function within the specified bounds.
+
+    Note:
+        Ensure JAXopt is installed to use this function (`pip install jaxopt`).
+        The acquisition function is minimized using its negative value to find the maximum.
+
+    Examples:
+
+        Optimize EI given a trained GP model for 1D problem
+
+        >>> acq_fn = gpax.acquisition.EI
+        >>> num_initial_guesses = 10
+        >>> lower_bound = -2.0
+        >>> upper_bound = 2.0
+        >>> x_next = gpax.acquisition.optimize_acq(
+        >>>    rng_key, gp_model, acq_fn,
+        >>>    num_initial_guesses, lower_bound, upper_bound,
+        >>>    maximize=False, noiseless=True)
+    """
+
+    try:
+        import jaxopt  # noqa: F401
+    except ImportError as e:
+        raise ImportError(
+            "You need to install `jaxopt` to be able to use this feature. "
+            "It can be installed with `pip install jaxopt`."
+        ) from e
+
+    def acq(x):
+        x = jnp.array([x])
+        x = x[None] if x.ndim == 0 else x
+        obj = -acq_fn(rng_key, model, x, **kwargs)
+        return jnp.reshape(obj, ())
+
+    lower_bound = ensure_array(lower_bound)
+    upper_bound = ensure_array(upper_bound)
+
+    initial_guesses = jra.uniform(
+        rng_key, shape=(num_initial_guesses, lower_bound.shape[0]),
+        minval=lower_bound, maxval=upper_bound)
+    initial_acq_vals = acq_fn(rng_key, model, initial_guesses, **kwargs)
+    best_initial_guess = initial_guesses[initial_acq_vals.argmax()].squeeze()
+
+    minimizer = jaxopt.ScipyBoundedMinimize(fun=acq, method='l-bfgs-b')
+    result = minimizer.run(best_initial_guess, bounds=(lower_bound, upper_bound))
+
+    return result.params
+
+
+def ensure_array(x):
+    if not isinstance(x, jnp.ndarray):
+        if isinstance(x, (list, tuple, float, onp.ndarray)):
+            x = jnp.array([x]) if isinstance(x, float) else jnp.array(x)
+        else:
+            raise TypeError(f"Expected input to be a list, tuple, float, or jnp.ndarray, got {type(x)} instead.")
+    return x

requirements.txt

@@ -4,4 +4,5 @@ matplotlib>=3.1.1
 jax>=0.2.21
 numpyro>=0.8.0
 dm-haiku>=0.0.5
+jaxopt>0.8.0
 typing-extensions>=4.4.0

tests/test_acq.py

@@ -3,8 +3,6 @@
 import numpy as onp
 import jax
 import jax.numpy as jnp
-import numpyro
-import numpyro.distributions as dist
 from numpy.testing import assert_equal, assert_
 
 sys.path.insert(0, "../gpax/")

tests/test_optimize_acq.py

@@ -0,0 +1,36 @@
+import sys
+import pytest
+import numpy as onp
+import jax.numpy as jnp
+from numpy.testing import assert_
+
+sys.path.insert(0, "../gpax/")
+
+from gpax.models.gp import ExactGP
+from gpax.acquisition.optimize import optimize_acq
+from gpax.acquisition.acquisition import UCB, EI
+from gpax.utils import get_keys
+
+
+def get_inputs():
+    X = onp.random.uniform(-2, 2, size=(4,))
+    y = X**3
+    return X, y
+
+
+@pytest.mark.parametrize("acq_fn", [UCB, EI])
+def test_optimize_acq(acq_fn):
+    lower_bound = -2.0
+    upper_bound = 2.0
+    num_initial_guesses = 3
+    key1, key2 = get_keys()
+    X, y = get_inputs()
+    model = ExactGP(1, 'RBF')
+    model.fit(key1, X, y, num_warmup=50, num_samples=50)
+    x_next = optimize_acq(
+        key2, model, acq_fn, num_initial_guesses, lower_bound, upper_bound)
+    assert_(isinstance(x_next, jnp.ndarray))
+
+
+
+