Introduction and use of automatic machine learning framework (flaml, H2O)

One 、 Introduce

Two 、 Data is introduced

3、 ... and 、flaml frame

        3.1、flaml brief introduction

        3.2、 Use flaml

                3.2.1、 download flaml library

                3.2.2、 Import related libraries

                3.2.3、 Data processing

                3.2.4、 call flaml

Four 、h2o frame

        4.1、h2o brief introduction

        4.2、h2o Use

                4.2.1、 download h2o

                4.2.2、 Import related libraries

                4.2.3、 Data processing

                4.2.4、 start-up h2o Of jar package

                4.2.5、 call h2o

5、 ... and 、 summary

---

One 、 Introduce

         Automatic machine learning （Automl）, It is a way to turn traditional machine learning into Automation ,start—end Fully automatic . At present, the popular automatic machine learning frameworks in the market include ：Flaml、H20 wait . This chapter records how these two frameworks are used .

Two 、 Data is introduced

        The bank transaction flow created by the data refers to . share 23 Column data , Among them is 18 Column characteristic data ,1 Column label data ,4 Column user information data .

3、 ... and 、flaml frame

        3.1、flaml brief introduction

                Flaml It is an automatic machine learning framework launched by Microsoft , Support custom learners and parameters , It also provides a fast automatic adjustment tool .flaml You can find the accuracy with low computational resources in the customized learner ML Model . It frees users from choosing learners and super parameters . It is very convenient to use .        

        3.2、 Use flaml

                3.2.1、 download flaml library

pip install flaml

                3.2.2、 Import related libraries

from flaml import AutoML
from sklearn.datasets import load_iris
from sklearn.datasets import load_boston
import pandas as pd
import sys,logging
from sklearn.metrics import confusion_matrix,classification_report,recall_score,accuracy_score,f1_score,precision_score
from sklearn.preprocessing import MinMaxScaler
from imblearn.over_sampling import RandomOverSampler
import time

                3.2.3、 Data processing

#  Data path 
data_path = r"source/data_jianhang.csv"
input_data_all = pd.read_csv(data_path,encoding="gbk",index_col=0)
#  Extract forecast customer information   【 Customer name 、 Customer number 、 Customer account number 】
customer_info = input_data_all.iloc[:,:3]
print(customer_info)

#  features 
input_data_target = input_data_all[" label "]
input_data_feature = input_data_all.iloc[:, 3:-1]
#  selection input_data_all All of the line , From the third column to the last column （ The head is not the tail ）
input_data = input_data_all.iloc[:, 3:]
#  Fill the blank value with 0
input_data.fillna(0, inplace=True)
#  Output the first five lines to view 
print(input_data.head())


#  Random oversampling for the imbalance of positive and negative samples 
f = RandomOverSampler(random_state=0)
data, target = f.fit_resample(input_data.iloc[:,:-1], input_data.iloc[:,-1])
#  Data maximum and minimum normalization 
data = MinMaxScaler().fit_transform(data)
#  Output sample quantity 
print(target.to_frame().value_counts())

#  Segmentation data   X： Feature set   y： label 
X,y = input_data.iloc[:,:-1],input_data.iloc[:,-1]
X = MinMaxScaler().fit_transform(X)

                3.2.4、 call flaml

                        Looking at the log printed by the program, we can find , Through parameters estimator_list, Calling flaml In the process of ,flaml Automatically compare for us lgbm、rf、xgboost The effect of classifier , Finally, print and use the optimal classifier and parameters to train the model . The whole process is automated , There is no need for artificial operation comparison . This is also a feature of automated machine learning .

t1 = time.time()
#  initialization flaml Automated modeling framework 
flaml_automl = AutoML()
#  Pass in training data x and y Conduct fit Training 
flaml_automl.fit(data,target,task='classification',log_file_name="xxx.log",metric="f1",estimator_list = ['lgbm', 'rf', 'xgboost'])
# fit Introduction to common parameters 
'''
    #  X_train=None,    Training data feature set 
    #  y_train=None,    Training data tag set 
    #  estimator_list = ['lgbm', 'rf', 'xgboost', 'extra_tree', 'xgb_limitdepth', 'lrl1']
    #  metric: 'accuracy', 'roc_auc', 'roc_auc_ovr', 'roc_auc_ovo','f1', 'micro_f1', 'macro_f1', 'log_loss', 'mae', 'mse', 'r2','mape'.
    #  n_jobs: Pass in an integer , Enable multithreading 
    #  n_splits： Pass in an integer , Number of folds cross verified 
    #  log_file_name: Log output , If you don't want to output logs , Pass in an empty string  ’’  that will do 
    #  estimator_list： Model list , Optional 【‘lgbm’,’xgboot’,’xgb_limitdepth’,’catboost’,’rf’,’extra_tree’】, It will eventually output best Model .
    #  time_budget： The time limit , In seconds . If restricted 10s, Then the optimal model is output in ten seconds . There is no time limit for incoming -1
    #  sample： Boolean value , Default False. Whether to sample the incoming data .
    #  early_stop: Boolean value , Default False. If the model search converges , Stop ahead of time .
'''


# flaml Print the optimal model and parameters 
'''
[flaml.automl: 03-09 14:52:24] {2694} INFO - retrain lgbm for 1.3s
[flaml.automl: 03-09 14:52:24] {2699} INFO - retrained model: LGBMClassifier(colsample_bytree=0.5716563773446997,
               global_max_steps=9223372036854775807,
               learning_rate=0.7886932330930241, max_bin=511,
               min_child_samples=7, n_estimators=181, num_leaves=1006,
               reg_alpha=0.007095760722363662, reg_lambda=0.3005614400342159,
               verbose=-1)
[flaml.automl: 03-09 14:52:24] {2077} INFO - fit succeeded
[flaml.automl: 03-09 14:52:24] {2079} INFO - Time taken to find the best model: 23.60042953491211
'''


#  Print some results 
print(" Elapsed time ： ",time.time()-t1)
print(flaml_automl.estimator_list)
print(" The optimal model ",flaml_automl.model)
print(" Optimal parameters ",flaml_automl.best_config)
print(" Training time ",flaml_automl.best_config_train_time)
print(" classifier ",flaml_automl.best_estimator)
print(" Loss ",flaml_automl.best_loss)

#  call predict forecast X
y_pred = flaml_automl.predict(X)
#  Output forecast results 
print(y_pred)

#  Print indicators 、 Classification effect 
print(" Confusion matrix :\n",confusion_matrix(y,y_pred))
print(" Classified reports :\n",classification_report(y,y_pred))
print(" Recall rate :",recall_score(y,y_pred))
print(" Accuracy rate :",accuracy_score(y,y_pred))
print("f1 The score is :",f1_score(y,y_pred))
print(" Accuracy :",precision_score(y,y_pred))
print(" Total time consumed ： ",time.time()-t1)

Four 、h2o frame

        4.1、h2o brief introduction

                 h2o The framework is an open source , Distributed based on java The framework of machine learning .h2o By （h2o.AI） The company developed and released . The website of their company ：H2O.ai | AI Cloud Platform.h20 It also supports visual analysis of user tasks .

        4.2、h2o Use

                4.2.1、 download h2o

pip install h20

                4.2.2、 Import related libraries

import h2o
from h2o.automl import H2OAutoML
from h2o.estimators.random_forest import H2ORandomForestEstimator
from h2o.estimators.gbm import H2OGradientBoostingEstimator
from h2o.estimators.stackedensemble import H2OStackedEnsembleEstimator
from h2o.grid.grid_search import H2OGridSearch
import numpy as np
import pandas as pd
import time
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import confusion_matrix,classification_report,recall_score,accuracy_score,f1_score,precision_score
from imblearn.over_sampling import RandomOverSampler

                4.2.3、 Data processing

                        Same as 3.2.3 The chapters are exactly the same , I'm not going to repeat it here , Direct copy 3.2.3 Chapter data processing code .

                4.2.4、 start-up h2o Of jar package

                        1） Download free ： Automatic machine learning h2o start-up jar package - Machine learning document class

                        2） open cmd, Input java -jar h2o.jar  start-up

                        3） visit web page ： http://localhost:54321

 

                4.2.5、 call h2o

##``"DRF"``,``"GLM"``,``"XGBoost"``,``"GBM"``,``"DeepLearning"``,``"StackedEnsemble"``.
#  initialization 
automl_estimator = H2OAutoML(max_runtime_secs=50,balance_classes=True,exclude_algos=["DeepLearning"],stopping_metric="auc",sort_metric="auc")
#  Training 
automl_estimator.train(x=train_data_h2o.names[0:-1],y="target",training_frame=train_data_h2o)
print(" Time 2：",time.time()-t1)
# predict To make predictions , And output the prediction results 
h2o_result = automl_estimator.predict(test_data_h2o[:-1])[:,0]
print(h2o_result)

# Print indicators 
print(" Confusion matrix :\n",confusion_matrix(test_data_h2o[:,-1].as_data_frame(),h2o_result.as_data_frame()))
print(" Classified reports :\n",classification_report(test_data_h2o[:,-1].as_data_frame(),h2o_result.as_data_frame()))
print(" Recall rate :",recall_score(test_data_h2o[:,-1].as_data_frame(),h2o_result.as_data_frame()))
print(" Accuracy rate :",accuracy_score(test_data_h2o[:,-1].as_data_frame(),h2o_result.as_data_frame()))
print("f1 The score is :",f1_score(test_data_h2o[:,-1].as_data_frame(),h2o_result.as_data_frame()))
print(" Accuracy :",precision_score(test_data_h2o[:,-1].as_data_frame(),h2o_result.as_data_frame()))
print(" Time 3：",time.time()-t1)

5、 ... and 、 summary

        Because of h2o Less understanding , And so on h2o The code of the framework is not explained too much . Personally, I prefer to use flaml Do automated machine learning . As for the effect, it's still very good , No screenshots are posted here . Those who are interested can practice offline .