在pickle.load()之后，XGBRegressor.predict()返回错误

Question

我已经使用sklearn界面训练了XGBRegressor模型。相关代码如下：

def xgb_regressor_wrapper(X_train, y_train):
    xgb_regressor = XGBRegressor(objective='reg:linear', n_estimators=1000, learning_rate=0.01, base_score=0.005)
    xgb_regressor.fit(X=X_train, y=y_train) #, eval_set=[(X_test, y_test)], verbose=True)
    return xgb_regressor

def save_regressor(station, feature, regressor):
    fname = generate_regressor_fname(station, feature)
    pickle.dump(regressor, open(fname, "wb" ))

# regressor_list dict contains wrapper functions
# I currently have XGBRegressor and CatBoostRegressor in the list.
regressor_wrapper = regressor_list.get(name) 

# Create and fit XGBRegressor
regressor = regressor_wrapper(X_train, y_train)

# Save regressor
save_regressor(station_id, feature, best_regressor)

一段时间后，我使用以下代码重新加载回归器，并进行预测：

def load_regressor(station, feature):
    fname = generate_regressor_fname(station, feature)
    return pickle.load(open(fname, "rb" ))

# Load the regressor
regressor = load_regressor(station_id, feature)

# Do the prediction
y_predict = regressor.predict(X_test)

我得到以下错误：

  File "regressor_stuff.py", line 169, in regressor_check_for_station_feature
    y_predict = regressor.predict(X_test)
  File "D:\Anaconda\envs\Deep\lib\site-packages\xgboost\sklearn.py", line 268, in predict
    return self.booster().predict(test_dmatrix,
TypeError: 'str' object is not callable

经过一些调试，我发现self.booster实际上存储了字符串'gbtree‘。在训练了回归器的特性后(顺便说一句，这花了几天时间)，这并不酷。

对于为什么会发生这种情况，有什么建议吗？

我目前的解决方法是按如下方式重建XGBBooster：

# Load the regressor
if isinstance(regressor, XGBRegressor):
    regressor = XGBRegressor()
    r = pickle.load(open(fname, "rb" ))
    print r.get_xgb_params()
    regressor._Booster = r._Booster
    regressor.set_params(**r.get_xgb_params())

# Do the prediction
y_predict = regressor.predict(X_test)

谢谢

库尔萨特

Answer 1

我认为在您的训练和评分环境中，您可能会遇到xgboost版本不匹配的问题。我遇到了同样的问题，并发现我使用xgboost==0.6进行训练，而不是使用xgboost==0.6a2进行评分。