pass fit_params to the cross_val_score in sklearn, it always return nan

Question

when I pass fit_params to the cross_val_score in sklearn, it always return nan

parameters = {'n_estimators':100}

model = XGBRegressor()

cv_results = cross_val_score(model,
                             X_train_,
                             y_train_,
                             cv=split,
                             scoring="neg_mean_squared_error",
                             fit_params=parameters,
                             n_jobs=-1)

result is: [nan nan nan nan nan]

when I not use the fit_params or pass an empty dict to the fit_params, the result is okay.
But return nan, when pass any parameter to fit_params.
Anyone knows the answer, why?

Answer 1

You get an error because the parameter is not used in .fit() method, and if you don't provide parameters you don't run into an error.

n_estimators is a parameter used in initializing the model XGBRegressor(), that is you can do:

model = XGBRegressor(n_estimators = 100)

cv_results = cross_val_score(model,
                             X_train_,
                             y_train_,
                             cv=5,
                             scoring="neg_mean_squared_error")

The parameters are passed to the .fit() method (see help page) are eval_metric, early_stopping_rounds etc, so for example:

model = XGBRegressor(n_estimators = 100)
parameters = {'eval_metric': 'mae'}
    
cv_results = cross_val_score(model,
                             X_train_,
                             y_train_,
                             cv=5,
                             scoring="neg_mean_squared_error",
                             fit_params = parameters)

Answer 2

@StupidWolf nailed it ...

I was having a similar issue when I was passing a pipeline to cross_val_score, and solved it by passing the model directly, like in the example here by @StupidWolf.

xgb_model = XGBClassifier(**hyperparams,use_label_encoder =False, n_jobs=4)

fit_params={'early_stopping_rounds': 5, 
            'eval_metric': ['logloss'],
            'verbose': False,
            'eval_set': [(X_valid, y_valid)]
           }

# Multiply by -1 since sklearn calculates *negative* RMSE
scores = -1 * cross_val_score(xgb_model, X_train, y_train,
                              cv=5,
                              scoring='neg_root_mean_squared_error',
                              fit_params = fit_params,
                              verbose=1)