Implementation:Rapidsai Cuml Symreg Example

Knowledge Sources	Rapidsai_Cuml
Domains	Machine_Learning, Genetic_Programming
Last Updated	2026-02-08 12:00 GMT

Overview

A standalone C++ example demonstrating GPU-accelerated symbolic regression using the cuML genetic programming API.

Description

symreg_example.cpp provides a complete working example of the cuML symbolic regression pipeline. It reads training and test data from text files, configures genetic programming hyperparameters, trains a symbolic regression model on the GPU, and evaluates the resulting mathematical expression on a test dataset.

The program follows this pipeline:

1. Argument Parsing -- Command-line arguments configure population size, random state, number of generations, stopping criteria, mutation probabilities (crossover, subtree, hoist, point), parsimony coefficient, evaluation metric (MAE, MSE, or RMSE), dataset dimensions, and file paths for training/test data, labels, and optional sample weights.
2. Data Loading -- Reads column-major formatted text files into host vectors using parse_col_major. Missing weight files default to uniform weights.
3. GPU Memory Allocation -- Allocates device memory using RMM (rmm::device_uvector) for features, labels, weights, predictions, and program storage. Data is copied from host to device using async memcpy.
4. Training -- Calls cuml::genetic::symFit to evolve a population of programs over multiple generations. The training history is stored as a vector of program populations.
5. Best Program Selection -- Finds the program with the lowest raw fitness in the final generation and converts it to a human-readable equation using cuml::genetic::stringify.
6. Inference -- Calls cuml::genetic::symRegPredict to generate predictions on test data using the best program, and evaluates the fitness score using cuml::genetic::compute_metric.

The example uses CUDA events for timing each phase (allocation, training, inference) and outputs performance metrics.

Usage

Use this example as a reference for integrating cuML symbolic regression into C++ applications, or to test and benchmark the genetic programming implementation with custom datasets.

Code Reference

Source Location

• Repository: Rapidsai_Cuml
• File: cpp/examples/symreg/symreg_example.cpp

Signature

int main(int argc, char* argv[]);

template <typename T>
T get_argval(char** begin, char** end, const std::string& arg, const T default_val);

template <typename math_t = float>
int parse_col_major(const std::string fname, std::vector<math_t>& vec,
                    const int n_rows, const int n_cols);

Import

#include <cuml/genetic/common.h>
#include <cuml/genetic/genetic.h>
#include <cuml/genetic/program.h>
#include <cuml/common/logger.hpp>
#include <raft/util/cudart_utils.hpp>
#include <rmm/device_uvector.hpp>
#include <rmm/device_scalar.hpp>

I/O Contract

Inputs

Name	Type	Required	Description
-n_cols	int	Yes	Number of feature columns
-n_train_rows	int	Yes	Number of training samples
-n_test_rows	int	Yes	Number of test samples
-train_data	string	No	Path to training feature file (default: train_data.txt)
-train_labels	string	No	Path to training labels file (default: train_labels.txt)
-test_data	string	No	Path to test feature file (default: test_data.txt)
-test_labels	string	No	Path to test labels file (default: test_labels.txt)
-population_size	int	No	Size of the genetic program population
-generations	int	No	Number of evolution generations
-metric	string	No	Evaluation metric: mae, mse, or rmse (default: mae)
-p_crossover	float	No	Crossover probability
-p_subtree	float	No	Subtree mutation probability
-p_hoist	float	No	Hoist mutation probability
-p_point	float	No	Point mutation probability

Outputs

Name	Type	Description
Best equation	string	Human-readable mathematical expression discovered by the genetic program
Training time	float	GPU training time in milliseconds
Inference score	float	Fitness score on the test dataset
Predicted values	float[]	Predicted target values for the test set

Usage Examples

# Run the symbolic regression example
./symreg_example \
  -n_cols 5 \
  -n_train_rows 1000 \
  -n_test_rows 200 \
  -train_data my_train.txt \
  -train_labels my_labels.txt \
  -test_data my_test.txt \
  -test_labels my_test_labels.txt \
  -population_size 5000 \
  -generations 20 \
  -metric mse

Related Pages

• Environment:Rapidsai_Cuml_CUDA_GPU