[python-package] Inconsistent Parameter Handling Between String and File Loading Methods in LightGBM 4.5.0

[marcom3622]

Here's a filled out issue report:

Description

There appears to be an inconsistency in how parameters are handled when loading a model from string versus loading from file. When using model_from_string(), the .params attribute is not set, while it is properly set when loading from a file using the same model. This creates inconsistent behavior in the API.

Reproducible example

import lightgbm as lgb
import numpy as np

# Create a simple model first
X = np.random.rand(100, 4)
y = np.random.rand(100)
train_data = lgb.Dataset(X, label=y)
model = lgb.train({'objective': 'regression'}, train_data, num_boost_round=1)

# Save model to string
model_str = model.model_to_string()

# Load model from string
loaded_model_str = lgb.Booster(model_str=model_str)
print("Params when loaded from string:", loaded_model_str.params)  # Empty dict {}
print("Params using _get_loaded_param():", loaded_model_str._get_loaded_param())  # Shows parameters

# Save and load from file for comparison
model.save_model('temp.txt')
loaded_model_file = lgb.Booster(model_file='temp.txt')
print("Params when loaded from file:", loaded_model_file.params)  # Shows parameters

Environment info

LightGBM version or commit hash: 4.5.0

Additional Comments

Looking at the source code, this appears to be due to model_from_string() not setting the params attribute, while loading from file does set it via self._get_loaded_param(). This inconsistency could cause issues for users relying on the .params attribute to access model parameters, especially with libraries like shap

The fix would likely involve adding self.params = self._get_loaded_param() in the model_from_string() method to maintain consistency with file loading behavior.

Relevant code sections showing the inconsistency:

• When loading from file: params = self._get_loaded_param()
• When loading from string: no parameter setting occurs

This behavior difference doesn't seem intentional and could be considered a minor bug in the API design.

[jameslamb]

Thanks for your interest in LightGBM and for the thorough write-up!

This was already identified in #6851 and fixed in #6852

Those changes will be in the next release of lightgbm . With older versions of lightgbm, you'll have to manually reset the parameters following the advice in #6851 (comment). Sorry about that.

I'm going to close this as a duplicate of #6851.

Relevant code sections showing the inconsistency

In the future, please provide links to code you're referring to. That way, anyone reading this can just click a link and see what you're seeing, instead of having to go figure out for themselves what you were referring to.