Examples

The following examples demonstrate how to use gplearn-CUDA for various tasks. For high-performance GPU-specific examples and benchmarks, please see the examples/ directory in the source repository.

High-Performance Examples

Detailed examples for CUDA acceleration can be found in the repository:

Regressor Tutorial: examples/regressor_cuda.py A step-by-step GPU example using SymbolicRegressor.
Classifier Tutorial: examples/classifier_cuda.py A GPU example using SymbolicClassifier.
Transformer Tutorial: examples/transformer_cuda.py A GPU example using SymbolicTransformer for feature generation.
Benchmark Suite: examples/benchmark.py A comprehensive performance comparison between multi-core CPU and GPU runs across various dataset scales.

Notebook Examples

A detailed set of examples for the original gplearn functionality (all of which are supported by gplearn-CUDA) can be found here in the original repository.

Tutorial 1: Symbolic Regressor

This example is the same as the first example in the LISP GP book and seeks to find the function \(y = X_0^{2} - X_1^{2} + X_1 - 1\).

from gplearn.genetic import SymbolicRegressor
import numpy as np

# Generate synthetic data
x0 = np.arange(-1, 1, .1)
x1 = np.arange(-1, 1, .1)
x0, x1 = np.meshgrid(x0, x1)
y_truth = x0**2 - x1**2 + x1 - 1

X_train = np.vstack([x0.ravel(), x1.ravel()]).T
y_train = y_truth.ravel()

# Fit the model
est = SymbolicRegressor(population_size=5000,
                       generations=20, stopping_criteria=0.01,
                       p_crossover=0.7, p_subtree_mutation=0.1,
                       p_hoist_mutation=0.05, p_point_mutation=0.1,
                       max_samples=0.9, verbose=1,
                       parsimony_coefficient=0.01, random_state=0)
est.fit(X_train, y_train)

# Inspect the results
print(est._program)

Output:

sub(add(-0.999, X1), mul(sub(X1, X0), add(X0, X1)))

The result is almost identical to the truth, with some rounding of the constant. To run this same example on the GPU for massive speedups on larger datasets, simply add device='cuda' to the estimator initialization.