Issue of loading a Python trained LGB regression model using C-API for prediction.

[BladeCoda]

Hello, I get a issue while using C API to load a LGB regression model which is trained in python(python 3.6, lightgbm 3.1.1).
The result predicted by C API loaded model is greatly differs from which is directly predicted by python library.

Here is how I train the regression model and do further prediction in python (brief code):

lgb_train = lgb.Dataset(train_feature, train_target)   
params = {
	'boosting_type': 'gbdt',
	'objective': 'regression',
	'max_depth': 8,
	'num_leaves': 255,
	'learning_rate': 0.05,
	'feature_fraction': 0.5, 
	'verbose': -1 
}
print('Starting training...')
gbm = lgb.train(params, lgb_train, num_boost_round=1500)
gbm.save_model(model_path)

gbm = lgb.Booster(model_file=model_path)  
result_predict = gbm.predict(test_feature, num_iteration=gbm.best_iteration)`

The result predicted by the python loaded model is just fine. However, when I load this saved model using C-API for prediction, the predicted result is totally not as my expected. Here is the brief code about how I load the model for prediction:

• I didn't post detailed features because there are too long, but I carefully check those features and ensured they are totally the same as what I used in python experiment.

#include "LightGBM/c_api.h"
#include <string>
#include <pthread.h>

class LightGbmAlgorithm
{
public:
    LightGbmAlgorithm()
    {
	    ;
    }

    ~LightGbmAlgorithm()
    {
        LGBM_BoosterFree(this->handle);
        LGBM_FastConfigFree(this->fastConfig);
    }
	
	int initConfig(int prediceType, int dataType, int inputLength, int startIteration, int numIteration)
	{
	    if (0 != LGBM_BoosterPredictForMatSingleRowFastInit(this->handle, prediceType, startIteration,
                                                            numIteration, dataType, inputLength, "", &(this->fastConfig)))
		{
			printf("lightgbm predict config failed [%s]......\n", LastErrorMsg());
			return -1;
		}
		return 0;
	}


    int handleData(void *ptrFeature, void *ptrOutput)
    {
        int64_t outLength = 0;
        INT32 dataType = 0;
        if (0 != LGBM_BoosterPredictForMatSingleRowFast(fastConfig, ptrFeature, &outLength, (double *)ptrOutput))
        {
            printf("lightgbm predict failed [%s]......\n", LastErrorMsg());
            return -1;
        }
        return (INT32)outLength;
    }

    INT32 loadModel(const char *modelPath)
    {
        return LGBM_BoosterCreateFromModelfile(modelPath, &(this->num_iterations), &(this->handle));
    }

private:
    BoosterHandle handle;
    INT32 num_iterations;        // Number of iterations of this booster
    FastConfigHandle fastConfig;
};


int main()
{
    string path = "path to model";
    LightGbmAlgorithm *algObject = new LightGbmAlgorithm();
	algObject->loadModel(path.c_str());
	if ( 0 != algObject->initConfig(C_API_PREDICT_NORMAL, C_API_DTYPE_FLOAT64, 108, 0, -1))
	{
	    _exit(1);
	}
	double data = "features to send";
	double result = "result to save";
	handleData(&data, &result);
	printf("result = {%lf}", result);
}

In fact, I already successfully test a classification task using C-API loaded python model, and the predicted result is corrected. But when I use the same code to load a regression model, it goes wrong.

Is there something I omitted to load a python trained regression model using C-API?

[JackYangzg]

I have the same problem. using python to train model and predict using c api for regression task.
the c api output is different with the python output.

It's a Bug?

[shiyu1994]

@BladeCoda @JackYangzg Thanks for using LightGBM. I think this is related to #3778. If the model is produced with the released python package (version 3.1.1), and loaded by C_API from latest master branch, there can be a problem. Could you please try if the fix in #4056 resolves your problems?

[StrikerRUS]

@BladeCoda The above-mentioned fix is in 3.2.0 release, so you can check it.

[no-response]

This issue has been automatically closed because it has been awaiting a response for too long. When you have time to to work with the maintainers to resolve this issue, please post a new comment and it will be re-opened. If the issue has been locked for editing by the time you return to it, please open a new issue and reference this one. Thank you for taking the time to improve LightGBM!

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.