对画图组件做了一些bug修正，新增了多个子图的支持

.idea/misc.xml

+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectRootManager" version="2" project-jdk-name="Python 3.6 (model_mvp)" project-jdk-type="Python SDK" />
+</project>
\ No newline at end of file

.idea/model_mvp.iml

+<?xml version="1.0" encoding="UTF-8"?>
+<module type="PYTHON_MODULE" version="4">
+  <component name="NewModuleRootManager">
+    <content url="file://$MODULE_DIR$" />
+    <orderEntry type="jdk" jdkName="Python 3.6 (model_mvp)" jdkType="Python SDK" />
+    <orderEntry type="sourceFolder" forTests="false" />
+  </component>
+  <component name="TestRunnerService">
+    <option name="PROJECT_TEST_RUNNER" value="Unittests" />
+  </component>
+</module>
\ No newline at end of file

.idea/modules.xml

+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="ProjectModuleManager">
+    <modules>
+      <module fileurl="file://$PROJECT_DIR$/.idea/model_mvp.iml" filepath="$PROJECT_DIR$/.idea/model_mvp.iml" />
+    </modules>
+  </component>
+</project>
\ No newline at end of file

.idea/vcs.xml

+<?xml version="1.0" encoding="UTF-8"?>
+<project version="4">
+  <component name="VcsDirectoryMappings">
+    <mapping directory="$PROJECT_DIR$" vcs="Git" />
+  </component>
+</project>
\ No newline at end of file

models_kit/general_methods.py

-def topN_feature_importance(classifier, clf, topN=20, model=lgb):
+import matplotlib.pyplot as plt
+
+
+
+def topN_feature_importance(classifier, clf ,mode , topN=20):
     '''
     plot feature importance squence
     '''
@@ -11,3 +15,4 @@ def topN_feature_importance(classifier, clf, topN=20, model=lgb):
     plt.title("Feature Importances")
     plt.show()
 
+

models_kit/lightgbm.py

@@ -44,8 +44,7 @@ def returnAUC(clf, training_set, validation_set, features, target='target'):
     return train_auc, val_auc
 
 
-def train_lgbm(params, df_train, df_val, features, adds_on=None, target='target',
-               featureImportance_path = '../mvp/plots/', topN_featureImportance=20, featureImportance_title='lightgbm'):
+def train_lgbm(params, df_train, df_val, features, adds_on=None, target='target'):
     '''
     instructions : training lightgbm model with specified params
 
@@ -70,8 +69,6 @@ def train_lgbm(params, df_train, df_val, features, adds_on=None, target='target'
 
     lgbm = lgb.train(params, lgb_train, valid_sets=lgb_val, verbose_eval=False)
     train_auc, val_auc = returnAUC(lgbm, df_train, df_val, features)
-    matplot.topN_feature_importance(lgb, lgbm, title=featureImportance_title,
-                                    save_path = featureImportance_path, topN=topN_featureImportance)
     # auc = roc_auc_score(dev['target'],gbm.predict(dev[features]))
     return train_auc, val_auc, lgbm
 
@@ -102,12 +99,12 @@ def lgb_params_tuning(params, features, train, val, target='target', topN=3, cv_
     print('Memory Occupancy Rate: ' + (str)(psutil.virtual_memory().percent) + '%')
     optimal_para = list(topn)
 
-    for deepth in np.arange(2, 7, 1):
-        for leaves in np.arange(2, 2 ** deepth, 2):
+    for deepth in np.arange(2, 4, 1):
+        for leaves in np.arange(2, 2 ** deepth, 4):
             params['max_depth'] = deepth
             params['num_leaves'] = leaves
             print("parameter combination : ", 'max_depth ', deepth, 'num_leaves ', leaves)
-            cv_result = lgb.cv(params, lgb_train, seed=7, nfold=cv_fold, verbose_eval=False)
+            cv_result = lgb.cv(params, lgb_train, seed=7, nfold=cv_fold, verbose_eval=30)
             # return max auc(best performance)
             auc_score = pd.Series(cv_result['auc-mean']).max()
             print('auc ', auc_score)
@@ -122,7 +119,7 @@ def lgb_params_tuning(params, features, train, val, target='target', topN=3, cv_
                 para['max_depth'] = deepth
                 para['num_leaves'] = leaves
                 optimal_para[topn.argmin()] = para
-    return optimal_para, topn
+    return optimal_para, list(topn)
 
 
 #        training_curve.append(train_auc)
@@ -168,7 +165,7 @@ def lgb_params_tuning(params, features, train, val, target='target', topN=3, cv_
 
 def predict(lgbm,df_test,features,target='target'):
     predictions = lgbm.predict(df_test[features])
-    auc = roc_auc_score(predictions,df_test[target])
+    auc = roc_auc_score(df_test[target],predictions)
     return predictions, auc
 
 

mvp/refit.py

@@ -6,85 +6,122 @@ from models_kit import xgboost
 import lightgbm as lgb
 from graph import matplot
 from tools import filetool
+from sklearn.metrics import roc_auc_score
 
 
 dhb = dhb_obj.dhb(features=None, sql=None, start_time_period=None, end_time_period=None,passdue_day=15)
-
 # 提取样本
 #df_sample = dhb.dhb_features_extract()
 
 
 ######### temp #############
 import pandas as pd
-df_sample = pd.read_csv('E:\\model\\model_mvp\\mvp\\dhb_loan_sample——2019-04-23.csv',engine='python')
+df_sample = pd.read_csv('E:\\model\\model_mvp\\mvp\\sample.csv',engine='python')
+target = 'target'
+score = 'score'
+prediction = 'predict'
 
 ############################
 
 
 # 备份df_sample
-df_sample.to_csv(str(datetime.date.today())+"dhb_samples.xlsx")
+#df_sample.to_csv(str(datetime.date.today())+"dhb_samples.xlsx")
 
 
 # 默认样本划分
-df_train, df_val, df_test = datacal.train_test_split_general(df_sample, val_size=0.2, test_size=0.2, stratify='target',
+df_train, df_val, df_test = datacal.train_test_split_general(df_sample, val_size=0.2, test_size=0.2, stratify=target,
                                                              random_state=7,split_methods='random',
                                                              time_label='applied_at')
 del df_sample
 # 用交叉验证获取最优参optimal_para和对应参数在CV验证集上最优AUC列表topn
-optimal_para,topn = lightgbm.lgb_params_tuning(lightgbm.params_lgb, dhb.features, df_train, df_val, target='target',
+optimal_para,topn = lightgbm.lgb_params_tuning(lightgbm.params_lgb, dhb.features, df_train, df_val, target=target,
                                                topN=3, cv_fold=5)
 print('topn 通过train交叉验证得到的auc ',topn)
-# 用新参数(optimal_para)训练模型,adds_on是需要修改的参数字典,输出feature Importance
-train_auc, val_auc, lgbm = lightgbm.train_lgbm(lightgbm.params_lgb, df_train, df_val, dhb.features,
-                                               adds_on=optimal_para, target='target')
 
-predictions ,test_auc = lightgbm.predict(lgbm,df_test,features=dhb.features)
-df_test['predict'] = predictions
+# model matrix
+model_matrix_index = ['name','Params','trainAUC','validationAUC']
+model_matrix = pd.DataFrame(['NULL','NULL',roc_auc_score(df_train[target],df_train[score]),roc_auc_score(df_train[target],df_train[score])],index=model_matrix_index,columns=['线上模型'])
 
+pointer = 0
+for param in optimal_para:
+    train_auc, val_auc, lgbm = lightgbm.train_lgbm(lightgbm.params_lgb, df_train, df_val, dhb.features,
+                                                   adds_on=param, target=target)
+    model_matrix = pd.concat([model_matrix, pd.DataFrame(['lightGBM', param, train_auc, val_auc], index=model_matrix_index, columns=[pointer])],axis=1)
+    pointer += 1
 
+# 简单选取一下validation set auc 最高的 params
+best_params = model_matrix.T.sort_values(by='validationAUC',ascending=False).iloc[0,:].loc['Params']
+
+# 用新参数(optimal_para)训练模型,adds_on是需要修改的参数字典,输出feature Importance
+train_auc, val_auc, lgbm = lightgbm.train_lgbm(lightgbm.params_lgb, df_train, df_val, dhb.features,
+                                               adds_on=best_params, target='target')
 
+# 用新模型预测结果
+predictions ,test_auc = lightgbm.predict(lgbm,df_test,dhb.features,target)
+# 把新的预测结果加入test
+df_test[prediction] = predictions
 
 ####### allocator cache ############
 applied_from = {'1,214,217,198': '内部', '333': '融360', '159537': '360金融'}
-applied_type = {'1,2':'首贷','1,2,3':'首付贷','1':'首申','2':'复申','3':'复贷'}
+applied_type = {'1,2':'首贷','1,2,3':'全量客群','1':'首申','2':'复申','3':'复贷'}
 ####################################
 
-
 ### report
 
 # plot feature importance
-path = matplot.topN_feature_importance(lgb, lgbm, title="untitled", save_path='./plots/', topN=20)
+topnfeat_path = matplot.topN_feature_importance(lgb, lgbm, title="untitled", save_path='./mvp/plots/', topN=20)
+
 # report file
-report_path = "E:\\bla\\"
+report_path = "E:/bla/model_mvp/"
 report_name = "lgb_report.docx"
 
+# 生成docx Documents
 document = filetool.buildDocument(report_path, report_name)
 
+# docx加入title
 document.add_heading('lightGBM 算法refit报告')
 
-filetool.Document.add_paragraph('特征权重图')
+# docx新增 特征权重段
+document.add_paragraph('特征权重图')
+
+# docx加入特征权重图像
+document.add_picture(topnfeat_path)
 
-filetool.add_picture(path)
+# 新增 univar_chart段
+document.add_paragraph('univar_chart')
 
-filetool.Document.add_paragraph('univar_chart')
+# 遍历目标features画出univarchart
+for i in dhb.features[:3]:
+    univar_train = datacal.cal_univar(df_train, i, target, qcut=10)
+    univar_val = datacal.cal_univar(df_val, i, target, qcut=10)
+    univar_test = datacal.cal_univar(df_test, i, target, qcut=10)
+    univarChart = matplot.plot_table_list([univar_train,univar_val,univar_test], [1,2,3], datalist_description=None, title= i +' univar Chart', X_label=None, y_label=None,
+                    tab_df_list=None, plot_tab=True,
+                    saved_path='./mvp/plots/cache/')
+    document.add_picture('./mvp/plots/cache/' + i +' univar Chart' + ".png")
 
-for i in dhb.features:
-    univar = datacal.cal_univar(df_train,score='raw_score')
-    univarChart = matplot.plot_table(univar,title= i +' univar Chart',saved_path='./plots/cache')
-    filetool.add_picture("./plots/cache" + i +' univar Chart')
+document.add_paragraph('PDP_chart')
+# 遍历目标features 画出对应PDP
+for i in dhb.features[:3]:
+    pdp = datacal.cal_pdp(df=df_test, score=prediction, feature=i, qcut=10)
+    pdpChart = matplot.plot_table(pdp,title= i +' PDP Chart',saved_path='./mvp/plots/cache/')
+    document.add_picture('./mvp/plots/cache/' + i +' PDP Chart' + ".png")
 
-for i in dhb.features:
-    pdp = datacal.cal_pdp(df_test,score='predict')
-    pdpChart = matplot.plot_table(pdp,title= i +' PDP Chart',saved_path='./plots/cache')
-    filetool.add_picture("./plots/cache" + i + ' PDP Chart')
 
-for i in dhb.features:
-    lift = datacal.cal_liftchart(df_test,score='predict')
-    liftChart = matplot.plot_table(lift, title=i +' lift Chart',saved_path='./plots/cache')
-    filetool.add_picture("./plots/cache" + i + ' lift Chart')
 
 filetool.saveDocument(document, report_path, report_name)
 
+document.add_paragraph('lift_chart')
+# 遍历给定渠道 & 客群 默认等频画出liftchart
+for channel in applied_from:
+    for type in applied_type:
+        df_sliced = df_test[df_test.applied_type.map(lambda x : True if str(x) in type.split(',') else False) & df_test.applied_from.map(lambda x : True if str(x) in channel.split(',') else False)]
+        lift = datacal.cal_liftchart(df_sliced,score=prediction)
+        liftChart = matplot.plot_table(lift, title=i +' lift Chart',saved_path='./mvp/plots/cache')
+        document.add_picture("./mvp/plots/cache" + i + ' lift Chart.png')
+
+
+filetool.saveDocument(document, report_path, report_name)
 
 
 

tools/datacal.py

@@ -4,25 +4,78 @@ import datetime
 from sklearn.model_selection import train_test_split
 
 
-
-def liftchart(df,target='target',qcut=10,retbins=True):
+def cal_lift(df_list, score, target='target', qcut=10, retbin=False):
     '''
     instructions : return liftchart dataframe with qcut & pivot 逾期率liftchart
     Params :
-        df - dataframe(注意一定是是放款集！！)
+        df - dataframe(注意一定是是放款集！！) list
+        score - 模型分数
         target - label column
         qcut - quantiles
         retbins - return bins interval when 'retbins' is True, else False
     :return:
-        liftchart dataframe
+        liftchart pivot
     '''
+    pivot = pd.DataFrame([])
+    if type(df_list) == pd.DataFrame:
+        df = df_list.copy()
+        # fillin missing with -1
+        df.fillna(value=-1,inplace=True)
+        df = df[[score, target]]
+        # create a bins column
+        df_noneNA = [df[score] < 0]
+
+
+        df['bins'] = pd.qcut(df[score], q=qcut, precision=6, retbins=retbin, duplicates='drop')
+        pivot_tmp = df[['bins', target]].groupby('bins').agg(['mean', 'count'])
+        pivot = pd.concat([pivot, pivot_tmp], axis=1)
+    if type(df_list) == list:
+        print('none')
+        for df in df_list:
             df = df.copy()
+            df = df[[score, target]]
             # create a bins column
-    df['bins'] = pd.qcut(df, q=10, precision=6, retbins=False, duplicates='drop')
-    pivot = df[['bins','target']].groupby('bins').agg(['mean','count'])
-    return pivot
+            df['bins'] = pd.qcut(df[score], q=qcut, precision=6, retbins=retbin, duplicates='drop')
+            pivot_tmp = df[['bins', target]].groupby('bins').agg(['mean', 'count'])
+            pivot = pd.concat([pivot, pivot_tmp], axis=1)
+    return pivot[target]
 
 
+def cal_univar(df, feature, target, qcut=10):
+    '''
+    instructions : return univar pivot
+    Params:
+        :param df: dataframe with unvariable & label target(overdue label)
+        :param feature: single feature to
+        :param target:
+        :param qcut: N bins in the same frequency
+
+    :return: univar pivot
+    '''
+    df = df.copy()
+    df = df[[feature, target]]
+    # fill missing with -1
+    df.fillna(value=-1,inplace=True)
+    df['bins'] = pd.qcut(df[feature], q=qcut, precision=6, retbins=False, duplicates='drop')
+    pivot = df[[target,'bins']].groupby('bins').sum() / df[[target,'bins']].groupby('bins').count()
+    return pivot[target]
+
+
+
+def cal_pdp(df, score, feature, qcut=10):
+    '''
+    instructions : return pdp pivot
+    :param df: dataframe of test set
+    :param score: score that predicts by model
+    :param feature:
+    :param qcut:
+    :return:
+    '''
+    df = df.copy()
+    df = df[[feature, score]]
+    df['bins'] = pd.qcut(df[feature], q=qcut, precision=6, retbins=False, duplicates='drop')
+    pivot = df[[score,'bins']].groupby('bins').sum() / df[[score,'bins']].groupby('bins').count()
+    return pivot[score]
 
 
 
@@ -184,45 +237,45 @@ def cal_accume(df,feature,target,bin=10,classes=[]):
     return df_out
 
 
-def cal_univar(df,feature,target,bin=10,classes=[]):
-    '''
-    groupby(classes) 分组,对feature 进行bin 分位，对各个分位进行 count,mean ,sum计算
-    :param df: dataframe
-    :param feature: feature in df.columns
-    :param target: in df.columns eg: count(target) mean(target)
-    :param bins:default =10
-    :param classes: 分组
-    :return:
-    '''
-    if df.shape[0]==0:
-        raise('no data')
-    columns=df.columns.tolist()
-    if target not in columns:
-        raise('not found %s' % target)
-    if feature not in columns:
-        raise('not found %s' % feature)
-
-    tmp=df.copy()
-    tmp[feature].fillna(-1, inplace=True)
-    # == bin 划分,feature 有可能 非数字
-    try:
-        tmp[feature] = tmp[feature].astype(float)
-        feature_grid = cal_feature_grid(tmp, feature, bin)
-        tmp['lbl'] = pd.cut(tmp[feature], feature_grid, include_lowest = True)
-        tmp['grid'] = tmp['lbl'].cat.codes
-    except ValueError:
-        tmp['lbl']=tmp[feature]
-        tmp['grid']=tmp[feature]
-
-    if len(classes) > 0:
-        df_gp = tmp.groupby(classes+['grid','lbl']).agg({target: ['count', 'mean','sum']}).reset_index()
-        df_gp.columns = classes+['grid','lbl', 'count', 'mean','sum']
-        df_out=df_gp
-    else:
-        df_all = tmp.groupby(['grid','lbl']).agg({target: ['count', 'mean','sum']}).reset_index()
-        df_all.columns = ['grid', 'lbl', 'count', 'mean', 'sum']
-        df_out = df_all
-    return df_out
+# def cal_univar(df,feature,target,bin=10,classes=[]):
+#     '''
+#     groupby(classes) 分组,对feature 进行bin 分位，对各个分位进行 count,mean ,sum计算
+#     :param df: dataframe
+#     :param feature: feature in df.columns
+#     :param target: in df.columns eg: count(target) mean(target)
+#     :param bins:default =10
+#     :param classes: 分组
+#     :return:
+#     '''
+#     if df.shape[0]==0:
+#         raise('no data')
+#     columns=df.columns.tolist()
+#     if target not in columns:
+#         raise('not found %s' % target)
+#     if feature not in columns:
+#         raise('not found %s' % feature)
+#
+#     tmp=df.copy()
+#     tmp[feature].fillna(-1, inplace=True)
+#     # == bin 划分,feature 有可能 非数字
+#     try:
+#         tmp[feature] = tmp[feature].astype(float)
+#         feature_grid = cal_feature_grid(tmp, feature, bin)
+#         tmp['lbl'] = pd.cut(tmp[feature], feature_grid, include_lowest = True)
+#         tmp['grid'] = tmp['lbl'].cat.codes
+#     except ValueError:
+#         tmp['lbl']=tmp[feature]
+#         tmp['grid']=tmp[feature]
+#
+#     if len(classes) > 0:
+#         df_gp = tmp.groupby(classes+['grid','lbl']).agg({target: ['count', 'mean','sum']}).reset_index()
+#         df_gp.columns = classes+['grid','lbl', 'count', 'mean','sum']
+#         df_out=df_gp
+#     else:
+#         df_all = tmp.groupby(['grid','lbl']).agg({target: ['count', 'mean','sum']}).reset_index()
+#         df_all.columns = ['grid', 'lbl', 'count', 'mean', 'sum']
+#         df_out = df_all
+#     return df_out