Benchmark

Test device: E5 2680 v4, 14 core.

Package Comparison

This comparison for different packages ( gplearn, BindingGP) speed.

Test datasets: boston datasets from scikit-learn .

pop*gen times (s)	fastgplearn	gplearn	BindingGP
1000*10	0.15	4.91	29.09
10000*10	1.49	50.66	295.34
10000*10 (8 core)	1.45	57.01	126.20
100000*10	15.8	490.00	3053.01
100000*10 (8 core)	11.5	422.67	1452.58

Backend Comparison

This comparison for different backend and their parallelization performance.

Conclusion

1. For large samples datasets (more than 1000), torch > numpy.

2. For any population variation, numpy > torch.

3. For “c_torch”, keep n_jobs==1. c_torch is already well parallelized in c++ level, and does not need to be parallelized in python code.

Code

>>> np.random.seed(0)
>>> deps = generate_random(func_num=13, xs_num=10, pop_size=10000, depth_min=1, depth_max=3)
>>> deps = deps.tolist()

>>> "p_numpy"
>>> rs1 = p_np_score(deps, x, y, func_index=func_index)
>>> "p_numpy"
>>> rs2 = p_np_score_mp(deps, x, y, func_index=func_index, n_jobs=4)
>>> "c_numpy"
>>> rs3 = c_np_score(deps, x, y, func_index=func_index)
>>> "c_numpy"
>>> rs4 = c_np_score_mp(deps, x, y, func_index=func_index, n_jobs=4)

>>> "p_torch"
>>> rs1 = p_torch_score(deps, x2, y2, func_index=func_index)
>>> "p_torch"
>>> rs2 = p_torch_score_mp(deps, x2, y2, func_index=func_index, n_jobs=4)
>>> "c_torch"
>>> rs3 = c_torch_score(deps, x2, y2, func_index=func_index)
>>> "c_torch"
>>> rs4 = c_torch_score_mp(deps, x2, y2, func_index=func_index, n_jobs=4)