gfsopt

Convenient scaffolding for the excellent Global Function Search hyperparameter optimizer from the Dlib library. (See: 'A Global Optimization Algorithm Worth Using')

Provides the following features:

• Parallel optimization: Run multiple hyperparameter searches in parallel
• Save and restore progress: Save/restore settings and optimization progress to/from file.
• Average over multiple runs: Run a stochastic objective function using the same parameters multiple times and report the average to Dlib's Global Function Search. Useful in highly stochastic domains to avoid biasing the search towards lucky runs.

Example usage

A basic example where we maximize 'obj_func' with respect to 'y' over 10 runs, with as many parallel processes as there are logical cores, and save progress to file.

def obj_func(x, y, pid):
    a = (1.5 - x + x * y)**2
    b = (2.25 - x + x * y * y)**2
    c = (2.625 - x + x * y * y * y)**2
    return -(a + b + c)
# For this example we pretend that we want to keep 'x' fixed at 0.5
# while optimizing 'y' in the range -4.5 to 4.5
space = {'y': [False, -4.5, 4.5]}  # False since 'y' is not an Int
pp = {'x': 0.5}
optimizer = GFSOptimizer(pp, space, fname="test.pkl")
optimizer.run(obj_func, n_sims=10)

For a more extensive example, see example.py.

Installation & Requirements

Requires Python >=3.6 and the following libraries:

datadiff
dlib
numpy

To install, do: pip install gfsopt

Documentation

See example.py for an example and http://gfsopt.readthedocs.io/ for documentation.