/
main.py
263 lines (201 loc) · 14 KB
/
main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
from lightgbm.callback import early_stopping, reset_parameter
from pandas.core.arrays import categorical
from pandas.io.pytables import Term
from sklearn import dummy
from make_var import make_variable
from util import load_data,preprocess,mk_trainset, metric
from clustering import clustering
from sklearn.model_selection import train_test_split
from lightgbm import LGBMRegressor
from xgboost import XGBRegressor
from dim_reduction import by_PCA
import pandas as pd
import seaborn as sns
import numpy as np
import joblib
from sklearn.model_selection import KFold
from lightgbm import LGBMClassifier
from sklearn.preprocessing import RobustScaler
robustScaler = RobustScaler()
def boosting(X,y,X_val,robustScaler,col_sample=0.6,lr=0.04,iter=1500,inference=True):
X_val1 = X_val[X_val['is_mcode']==1]
X_val0 = X_val[X_val['is_mcode']==0]
y_val1 = X_val1['sales']
y_val0 = X_val0['sales']
model_lgb = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample, learning_rate=lr,n_estimators=iter,random_state=2020)
model_lgb1 = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample, learning_rate=lr,n_estimators=iter,random_state=2020)
model_lgb0 = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample, learning_rate=lr,n_estimators=iter,random_state=2020)
if inference:
model_lgb1.fit(X.drop(drop1_+['is_mcode'],axis=1),y,verbose=False)
model_lgb0.fit(X.drop(drop0_+['is_mcode','mcode_freq','mcode_freq_gr','mcode_sales_mean','mcode_sales_std','mcode_sales_med','mcode_sales_rank','mcode_order_mean','mcode_order_med','mcode_order_rank','mcode_order_std'],axis=1),y,verbose=False)
pred_lgb1 = model_lgb1.predict(X_val1.drop(drop1_+['is_mcode','sales'],axis=1))
res1 = pd.concat([y_val1.reset_index(drop=True),pd.DataFrame(pred_lgb1,columns=['pred'])],axis=1)
real1 = robustScaler.inverse_transform(np.array(res1['sales']).reshape(-1,1))
pred1 = robustScaler.inverse_transform(np.array(res1['pred']).reshape(-1,1))
print(f'real1 : {real1.shape},pred1 : {pred1.shape}')
pred_lgb0 = model_lgb0.predict(X_val0.drop(drop0_+['is_mcode','sales','mcode_freq','mcode_freq_gr','mcode_sales_mean','mcode_sales_std','mcode_sales_med','mcode_sales_rank','mcode_order_mean','mcode_order_med','mcode_order_rank','mcode_order_std'],axis=1))
res0 = pd.concat([y_val0.reset_index(drop=True),pd.DataFrame(pred_lgb0,columns=['pred'])],axis=1)
real0 = robustScaler.inverse_transform(np.array(res0['sales']).reshape(-1,1))
pred0 = robustScaler.inverse_transform(np.array(res0['pred']).reshape(-1,1))
print(f'real0 : {real0.shape},pred0 : {pred0.shape}')
score = (metric(real1,pred1) + metric(real0 , pred0))/2
print(len(res1),round(metric(real1,pred1),2),len(res0),round(metric(real0 , pred0),2))
length = len(res0) + len(res1)
real = np.concatenate((real1,real0))
pred = np.concatenate((pred1,pred0))
return score, length, pd.DataFrame({'real':real.flatten(), 'pred':pred.flatten()},columns=['real','pred']),model_lgb0,model_lgb1
else:
model_lgb.fit(X.drop(['is_mcode'],axis=1),y)
pred_lgb = model_lgb.predict(X.drop(['is_mcode'],axis=1))
res = pd.concat([y.reset_index(drop=True),pd.DataFrame(pred_lgb,columns=['pred'])],axis=1)
real = robustScaler.inverse_transform(np.array(res['sales']).reshape(-1,1))
pred = robustScaler.inverse_transform(np.array(res['pred']).reshape(-1,1))
print(real.shape,pred.shape)
return metric(real,pred), len(res), pd.DataFrame({'real':real.flatten(), 'pred':pred.flatten()},columns=['real','pred']),model_lgb0, model_lgb1
# def boosting_2(pop,inference,robustScaler,col_sample=0.6,lr=0.04,iter=1500,test=True):
# y = pop['sales']
# X = pop.drop(['id','sales','kmeans'],axis=1)
# model_lgb = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample, learning_rate=lr,n_estimators=iter,random_state=2020)
# if test==True:
# model_lgb.fit(X,y)
# pred_lgb = model_lgb.predict(inference.drop(['id','sales','kmeans'],axis=1))
# res = pd.concat([inference.reset_index(drop=True),pd.DataFrame(pred_lgb,columns=['pred'])],axis=1)
# real = robustScaler.inverse_transform(np.array(res['sales']).reshape(-1,1))
# pred = robustScaler.inverse_transform(np.array(res['pred']).reshape(-1,1))
# return metric(real,pred), pd.DataFrame({'real':real.flatten(), 'pred':pred.flatten()},columns=['real','pred']), model_lgb
# else:
# X_train, X_val, y_train, y_val = train_test_split(X,y,random_state=2020)
# model_lgb.fit(X_train,y_train,early_stopping_rounds = 1000,eval_set = [(X_val,y_val)],verbose=False)
# pred_lgb = model_lgb.predict(X_val)
# res = pd.concat([y_val.reset_index(drop=True),pd.DataFrame(pred_lgb,columns=['pred'])],axis=1)
# real = robustScaler.inverse_transform(np.array(res['sales']).reshape(-1,1))
# pred = robustScaler.inverse_transform(np.array(res['pred']).reshape(-1,1))
# return metric(real,pred), pd.DataFrame({'real':real.flatten(), 'pred':pred.flatten()},columns=['real','pred']), model_lgb
def predict(X_train,val,k,robustScaler,col_sample=0.6,lr=0.04,iter=1500,inference=True):
"""
predict '취급액' score only using train set(perform)
return : RMAE score for each cluster
"""
origin, originlen, tmp, model0,model1 = boosting(X_train.drop(['id','sales','kmeans'],axis=1),X_train['sales'],val.drop(['id','kmeans'],axis=1),robustScaler,col_sample,lr,iter,inference)
print(f'origin error : {round(origin,2)}%\n')
sum = 0
total_len = 0
for i in range(k):
train_tem = X_train[X_train['kmeans']==i]
val_tem = val[val['kmeans']==i]
score,len,pred,_,_ = boosting(train_tem.drop(['sales','kmeans','id'],axis=1),train_tem['sales'],val_tem.drop(['kmeans','id'],axis=1),robustScaler,col_sample,lr,iter,inference)
if inference==True:
results = pd.concat([val_tem.reset_index(drop=True),pred],axis=1)
else:
results = pd.concat([train_tem.reset_index(drop=True),pred],axis=1)
results['MAPE'] = pred.apply(lambda x: metric(x['real'],x['pred']),axis=1)
if i == 0:
fin_results = results.copy()
else:
fin_results = pd.concat([fin_results,results])
sum += (score * len)
total_len += len
print(f'Cluster_{i} : {round(score,2)}%\n')
print(f'Total error : {round(sum/total_len,2)}%')
return fin_results, model0, model1
# def second_predict(input,robustScaler,train,val):
# neg = input.iloc[:5000,:]
# pos = input.iloc[5000:,:].sample(n=5000,random_state=2020)
# population = pd.concat([neg,pos]).reset_index(drop=True)
# result,model_l,_ = predict(train,val,3,robustScaler,inference=True,iter=1500)
# # result_0 = result[result['kmeans']==0]
# score,res,model_lgb = boosting_2(population.iloc[:,:-3],result.iloc[:,:-3],robustScaler,0.55,0.04,1500,True)
# result.reset_index(drop=True, inplace=True)
# result.reset_index(drop=True, inplace=True)
# result['pred_eva'] = (res['pred']*0.5 + result['pred']*0.5)
# print(metric(result['real'],result['pred_eva']))
# # print(f'second_fit cluster0 : {score}')
# # fin_score = (metric(result_0['pred_eva'], result_0['real'])*1467 + result[result['kmeans']!=0]['MAPE'].sum())/2746
# # print(f'final error : {round(fin_score,2)}%')
# return model_lgb
def final_test(train,test,k,robustScaler,col_sample=0.6,lr=0.04,iter=1500):
test0 = test[test['is_mcode']==0]
test1 = test[test['is_mcode']==1]
model_total0 = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample, learning_rate=lr,n_estimators=iter,random_state=2020)
model_total1 = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample, learning_rate=lr,n_estimators=iter,random_state=2020)
model_total0.fit(train.drop(drop0_+['id','sales','kmeans','is_mcode','mcode_freq','mcode_freq_gr','mcode_sales_mean','mcode_sales_std','mcode_sales_med','mcode_sales_rank','mcode_order_mean','mcode_order_med','mcode_order_rank','mcode_order_std'],axis=1),train['sales'],verbose=False)
pred_total0 = model_total0.predict(test0.drop(drop0_+['id','sales','kmeans','is_mcode','mcode_freq','mcode_freq_gr','mcode_sales_mean','mcode_sales_std','mcode_sales_med','mcode_sales_rank','mcode_order_mean','mcode_order_med','mcode_order_rank','mcode_order_std'],axis=1))
pred_total0 = robustScaler.inverse_transform(pred_total0.reshape(-1,1))
res0 = pd.concat([test0.reset_index(drop=True),pd.DataFrame(pred_total0,columns=['pred'])],axis=1)
model_total1.fit(train.drop(drop1_+['id','sales','kmeans','is_mcode'],axis=1),train['sales'],verbose=False)
pred_total1 = model_total1.predict(test1.drop(drop1_+['id','sales','kmeans','is_mcode'],axis=1))
pred_total1 = robustScaler.inverse_transform(pred_total1.reshape(-1,1))
res1 = pd.concat([test1.reset_index(drop=True),pd.DataFrame(pred_total1,columns=['pred'])],axis=1)
total_pred = pd.concat([res0,res1])
for i in range(k):
train_tem = train[train['kmeans']==i]
test_tem = test[test['kmeans']==i]
test_tem_0 = test_tem[test_tem['is_mcode']==0]
test_tem_1 = test_tem[test_tem['is_mcode']==1]
best_params = joblib.load(f'best_lgbm_params_{i+1}.pkl')
model_cluster0 = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample,**best_params,random_state=2020)
model_cluster1 = LGBMRegressor(subsample= 0.7, colsample_bytree= col_sample,**best_params,random_state=2020)
model_cluster0.fit(train_tem.drop(drop0_+['id','sales','kmeans','is_mcode','mcode_freq','mcode_freq_gr','mcode_sales_mean','mcode_sales_std','mcode_sales_med','mcode_sales_rank','mcode_order_mean','mcode_order_med','mcode_order_rank','mcode_order_std'],axis=1),train_tem['sales'],verbose=False)
pred_cluster0 = model_cluster0.predict(test_tem_0.drop(drop0_+['id','sales','kmeans','is_mcode','mcode_freq','mcode_freq_gr','mcode_sales_mean','mcode_sales_std','mcode_sales_med','mcode_sales_rank','mcode_order_mean','mcode_order_med','mcode_order_rank','mcode_order_std'],axis=1))
pred_cluster0 = robustScaler.inverse_transform(pred_cluster0.reshape(-1,1))
res_tem0 = pd.concat([test_tem_0.reset_index(drop=True),pd.DataFrame(pred_cluster0,columns=['pred'])],axis=1)
model_cluster1.fit(train_tem.drop(drop1_+['id','sales','kmeans','is_mcode'],axis=1),train_tem['sales'],verbose=False)
pred_cluster1 = model_cluster1.predict(test_tem_1.drop(drop1_+['id','sales','kmeans','is_mcode'],axis=1))
pred_cluster1 = robustScaler.inverse_transform(pred_cluster1.reshape(-1,1))
res_tem1 = pd.concat([test_tem_1.reset_index(drop=True),pd.DataFrame(pred_cluster1,columns=['pred'])],axis=1)
cluster_pred_tem = pd.concat([res_tem0,res_tem1])
if i ==0:
cluster_pred = cluster_pred_tem.copy()
else:
cluster_pred = pd.concat([cluster_pred,cluster_pred_tem])
return total_pred.sort_values(by='id').reset_index(drop=True), cluster_pred.sort_values(by='id').reset_index(drop=True)
# excution
if __name__=='__main__':
data_path = 'data/'
perform_raw, rating, test_raw = load_data(data_path,trend=False,weather=False,query=False)
train_var, test_var = make_variable(perform_raw,test_raw,rating)
raw_data, y_km, train_len= preprocess(train_var,test_var,0.03,3,inner=False)
data = mk_trainset(raw_data,categorical=True) # lgbm만 categorical = True, 나머지 모델은 False -> one-hot encoding
train, val, robustScaler = clustering(data,y_km,train_len,test=True) # test 할때만 test = True
# permutation으로 날릴 변수들
# lgbm 기준이라서 one-hot 안된 카테고리 변수들이 있음, 다른 모델 랜덤 서치 돌릴 때는
# 해당 변수들은 이름이 없을테니(ex. min -> min_0, min_1, min_2 ...)
# 알아서 에러나는거 보고 빼던가 미리 카테고리 변수는 drop 리스트에서 빼 놓으셈
# 그리고 변수 drop은 0,1 모델 기준으로 한거라 cluster 기준으로 랜덤서치하는 거랑 안 맞을 수 있음
# 혜린이한테는 일단 두 리스트 교집합으로 하라 했는데 더 좋은 방법 있음 생각해서 시도 ㄱㄱ
drop1_ = ['min_sales_med', 'min_sales_std', 'day_sales_rank', 'min_sales_rank',
'min_order_rank', 'cate_sales_rank', 'cate_order_rank', 'cate_order_med',
'cate_sales_med', 'prime', 'min_order_std', 'min_sales_mean',
'day_order_rank', 'cate_order_std', 'min_order_med', 'min_order_mean',
'cate_sales_mean', 'cate_sales_std', 'day_order_std', 'rating',
'day_sales_med', 'min', 'day_order_med']
drop0_ = ['min_order_med', 'day_order_rank', 'min_sales_rank', 'min_sales_med',
'day_sales_rank', 'min_order_rank', 'cate_order_std', 'cate_sales_rank',
'min_sales_std', 'min_order_std', 'min', 'cate_order_rank',
'min_order_mean', 'rating', 'min_sales_mean', 'prime', 'cate_order_med',
'day_order_med']
# val 코드(test할 땐 실행 X)
# tem_result,model0,model1 = predict(train,val,3,robustScaler,inference=True,iter=2000)
# 테스트 코드
"""
total : 클러스터 안나누고 한번에 돌린 결과
cluster : 클러스터별로 따로 모델돌린거 합친 결과
"""
total, cluster = final_test(train,val,3,robustScaler,0.6,0.04,2000)
#total.to_csv('final_predict.csv')
# def feature_select(val,mcode,var,model):
# val_sel = val[val['is_mcode']==mcode]
# val_sel.reset_index(drop=True,inplace=True)
# val_features = val_sel.drop(var,axis=1)
# r = permutation_importance(model, val_features, val_sel['sales'],
# n_repeats=30,
# random_state=0)
# drop = []
# for i in r.importances_mean.argsort():
# # print(i)
# # if r.importances_mean[i] - 2 * r.importances_std[i] > 0:
# drop.append(val_features.columns[i])
# print(f"{val_features.columns[i]:<8}"
# f"{r.importances_mean[i]:.3f}"
# f" +/- {r.importances_std[i]:.3f}")
# return drop