Loading a model can result in std::out_of_range exception

[jpkoponen]

The leaf values (leaf outputs) can become denormal doubles, meaning their absolute value is less than the minimum positive normal value of a double. (https://en.wikipedia.org/wiki/Denormal_number)

You can check the minimum value for normal and denormal doubles in C++ with std::numeric_limits<double>::min() and std::numeric_limits<double>::denorm_min().

This can cause that when loading the model from a string, you get an std::out_of_range exception because std::stod() cannot transform a string representing a denormal double to double.

I propose a solution where the leaf outputs are rounded to zero if they are denormal.

Unfortunately, I cannot give you an example of how the model was made because the data I used is sensitive. I can say that I encountered this problem with quantile regression and with many different parameter combinations.

Environment info

Operating System: Ubuntu 18.04.3 LTS

CPU/GPU model: Intel(R) Core(TM) i7-8650U CPU @ 1.90GHz

Python version: 3.7.4
g++ version: 7.4.0

LightGBM version: 2.3.2

Error message

[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
[LightGBM] [Warning] stod
terminate called without an active exception
Process finished with exit code 134

Reproducible examples

import lightgbm as lgb
booster = lgb.Booster(model_file='model_string.txt')

Steps to reproduce

1. Download the attached model_string.txt
2. Run the above Python code in the same directory as the downloaded file.

[guolinke]

thanks @jpkoponen . PR for fixing this is very welcomed.