# python3.6+sklearn0.22
#===========datasets.make_hastie_10_2===================================================================
from sklearn import datasets
def loglikehood(labels, preds):

    #labels = train_data.get_label()
    preds = 1.0 / (1.0 + np.exp(-preds))
    grad = (preds - labels)
    hess = preds * (1. - preds)
    return grad, hess

X, y = datasets.make_hastie_10_2(n_samples=12000, random_state=1)
X = X.astype(np.float32)

# map labels from {-1, 1} to {0, 1}
labels, y = np.unique(y, return_inverse=True)

X_train, X_valid = X[:2000], X[2000:]
y_train, y_valid = y[:2000], y[2000:]
tree_nums = 100
xgb_params = {'learning_rate': 0.01, "n_estimators": tree_nums,
              "max_depth": 5,'min_child_weight': 1, 'seed': 10,
              'subsample': 0.9, 'colsample_bytree': 0.9, 'gamma': 0, 'reg_alpha': 0, 'reg_lambda': 1,
              "tree_method": "gpu_hist", "n_jobs": 8}

xgb_clf0 = xgb.XGBClassifier(**xgb_params, objective="binary:logistic")

xgb_clf1 = xgb.XGBClassifier(**xgb_params, objective=loglikehood)

xgb_clf0.fit(X_train, y_train) 
xgb_clf1.fit(X_train, y_train)

def getAuc(clf):

    y_pred_prob = clf.predict_proba(X_valid)
    y_pred = y_pred_prob.argmax(1)

    auc_score = roc_auc_score(y_valid, y_pred_prob[:,1])
    
    return auc_score
auc_scores_0 = getAuc(xgb_clf0)

auc_scores_1 = getAuc(xgb_clf1)
print("binary:logistic => auc score: ",auc_scores_0)   # 0.89597577539
print("loglikelood => auc score: ",auc_scores_1)  # 0.834210247555
#===============================================================================    

def loglikelood(labels, preds):

    #labels = train_data.get_label()
    #preds = 1.0 / (1.0 + np.exp(-preds))
    grad = (preds - labels)
    hess = preds * (1. - preds)
    return grad, hess