At least one option is reset when loading a GPX surrogate from file: print_global

[esnguyenvan]

Hey!

After playing with #732 I realised that at least one option is reset when loading a GPX from a saved file. In my case, it's the print_global option that gets reset, I haven't investigated further but it may be the case for other options.

I re-used Enrico's sample code:

from pathlib import Path

import pandas as pd
from smt.surrogate_models import GPX, KRG

PATH_DOE = Path(__file__).parent / "doe_example.csv"
PATH_SURROGATE = Path(__file__).parent / "surrogate.json"

data = pd.read_csv(PATH_DOE)

y_train = data["out1"].to_numpy()
x_train = data.drop(columns="out1").to_numpy()

surrogate = GPX(
    theta0=[1e-1],
    theta_bounds=[1e-6, 20e0],
    print_global = False,
)
surrogate.set_training_values(x_train, y_train)
surrogate.train()

# Check the value of print_global
print("Initial value of print_global: ", surrogate.options._dict['print_global'])

# Save
surrogate.save(PATH_SURROGATE.as_posix())

#Load
surrogate_loaded_smt = GPX.load(PATH_SURROGATE.as_posix())

print("Value of print_global, after reload from file: ", surrogate_loaded_smt.options._dict['print_global'])

# Retry with KRG
surrogate_krg = KRG(
    theta0=[1e-1],
    theta_bounds=[1e-6, 20e0],
    print_global = False,
)
surrogate_krg.set_training_values(x_train, y_train)
surrogate_krg.train()

surrogate_krg.save(PATH_SURROGATE.as_posix())
surrogate_krg_loaded = KRG.load(PATH_SURROGATE.as_posix())

print("Initial value of print_global with KRG: ", surrogate_krg.options._dict['print_global'])
print("Value of print_global, after reload from file with KRG: ", surrogate_krg_loaded.options._dict['print_global'])

With the same data in deo_example.csv file:

,in1,in2,in3,in4,in5,out1
0,17.29527629,0.299723272,0.119148789,0.121477063,0.089776587,0.5
1,12.31335104,0.063619377,0.127541093,0.11526527,0.110776079,0.6
2,17.02670356,0.013510148,0.123028355,0.131023021,0.116939302,0.7
3,14.44708333,0.14615784,0.126824464,0.123176709,0.122047058,0.8
4,12.67327884,0.164918608,0.131369514,0.105941044,0.09386734,0.9

Give the following output:

Initial value of print_global:  False
Value of print_global, after reload from file:  True
Initial value of print_global with KRG:  False
Value of print_global, after reload from file with KRG:  False

Current version of SMT: 2.9.3 and egobox: 0.29.1

[relf]

Good catch! Options are not saved at all by GPX and take their default values at the loading time.
It is an undocumented (!) limitation of the GPX wrapper which is just an adapter of the egobox::Gpx API and is primarily designed to call train/predict methods.

To properly fix this, it would be required to implement special save/load methods to pickle the GPX object and mix this with the current native format used by the egobox::Gpx implementation.