-
Notifications
You must be signed in to change notification settings - Fork 80
/
amazon_stacking.py
236 lines (170 loc) · 9.89 KB
/
amazon_stacking.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
# -*- coding: utf-8 -*-
"""
@author: marios
Script that does meta modelling level 1 as in taking the held-out predictions from the previous models and using as features in a new model.
This one uses Extra trees to do this.
"""
import numpy as np
import gc
from sklearn.externals import joblib
from sklearn.preprocessing import StandardScaler
from sklearn.utils import shuffle
from sklearn.metrics import roc_auc_score
from sklearn.cross_validation import StratifiedKFold
from sklearn.ensemble import ExtraTreesClassifier
#load a single file
def loadcolumn(filename,col=4, skip=1, floats=True):
pred=[]
op=open(filename,'r')
if skip==1:
op.readline() #header
for line in op:
line=line.replace('\n','')
sps=line.split(',')
#load always the last columns
if floats:
pred.append(float(sps[col]))
else :
pred.append(str(sps[col]))
op.close()
return pred
#functions to manipulate pickles
def load_datas(filename):
return joblib.load(filename)
def printfile(X, filename):
joblib.dump((X), filename)
def printfilcsve(X, filename):
np.savetxt(filename,X)
def bagged_set(X,y,model, seed, estimators, xt, update_seed=True):
# create array object to hold predictions
baggedpred=[ 0.0 for d in range(0, (xt.shape[0]))]
#loop for as many times as we want bags
for n in range (0, estimators):
#shuff;e first, aids in increasing variance and forces different results
X_t,y_c=shuffle(X,y, random_state=seed+n)
if update_seed: # update seed if requested, to give a slightly different model
model.set_params(random_state=seed + n)
model.fit(X_t,y_c) # fit model0.0917411475506
preds=model.predict_proba(xt)[:,1] # predict probabilities
# update bag's array
for j in range (0, (xt.shape[0])):
baggedpred[j]+=preds[j]
# divide with number of bags to create an average estimate
for j in range (0, len(baggedpred)):
baggedpred[j]/=float(estimators)
# return probabilities
return np.array(baggedpred)
def save_results(predictions, filename):
"""Given a vector of predictions, save results in CSV format."""
with open(filename, 'w') as f:
f.write("id,ACTION\n")
for i, pred in enumerate(predictions):
f.write("%d,%f\n" % (i + 1, pred))
def main():
load_data=True
SEED=15
Use_scale=False # if we want to use standard scaler on the data
meta_folder="" # this is how you name the foler that keeps the held-out and test predictions to be used later for meta modelling
# least of meta models. All the train held out predictions end with a 'train.csv' notation while all test set predictions with a 'test.csv'
meta=["main_xgboost","main_logit_2way","main_logit_3way","main_logit_3way_best","main_xgboos_count","main_xgboos_count_2D","main_xgboos_count_3D"]
bags=1 # helps to avoid overfitting. Istead of 1, we ran 10 models with differnt seed and different shuffling
######### Load files (...or not!) ############
y = np.loadtxt("train.csv", delimiter=',',usecols=[0], skiprows=1)
if load_data:
Xmetatrain=None
Xmetatest=None
for modelname in meta :
mini_xtrain=np.loadtxt(meta_folder + modelname + '.train.csv') # we load the held out prediction of the int'train.csv' model
mini_xtest=np.loadtxt(meta_folder + modelname + '.test.csv') # we load the test set prediction of the int'test.csv' model
mean_train=np.mean(mini_xtrain) # we calclaute the mean of the train set held out predictions for reconciliation purposes
mean_test=np.mean(mini_xtest) # we calclaute the mean of the test set predictions
# we print the AUC and the means and we still hope that everything makes sense. Eg. the mean of the train set preds is 1232314.34 and the test is 0.7, there is something wrong...
print("model %s auc %f mean train/test %f/%f " % (modelname,roc_auc_score(np.array(y),mini_xtrain) ,mean_train,mean_test))
if Xmetatrain==None:
Xmetatrain=mini_xtrain
Xmetatest=mini_xtest
else :
Xmetatrain=np.column_stack((Xmetatrain,mini_xtrain))
Xmetatest=np.column_stack((Xmetatest,mini_xtest))
# we combine with the stacked features
X=Xmetatrain
X_test=Xmetatest
# we print the pickles
printfile(X,"xmetahome.pkl")
printfile(X_test,"xtmetahome.pkl")
X=load_datas("xmetahome.pkl")
print("rows %d columns %d " % (X.shape[0],X.shape[1] ))
#X_test=load_datas("onegramtest.pkl")
#print("rows %d columns %d " % (X_test.shape[0],X_test.shape[1] ))
else :
X=load_datas("xmetahome.pkl")
print("rows %d columns %d " % (X.shape[0],X.shape[1] ))
#X_test=load_datas("onegramtest.pkl")
#print("rows %d columns %d " % (X_test.shape[0],X_test.shape[1] ))
outset="amazon_stacking" # Name of the model (quite catchy admitedly)
number_of_folds =5 # repeat the CV procedure 5 times and save the holdout predictions
print("len of target=%d" % (len(y))) # print the length of the target variable because we can
#model we are going to use
model=ExtraTreesClassifier(n_estimators=10000, criterion='entropy', max_depth=9, min_samples_leaf=1, max_features=6, n_jobs=30, random_state=1)
#model=LogisticRegression(C=0.01)
train_stacker=[ 0.0 for k in range (0,(X.shape[0])) ] # the object to hold teh held-out preds
# CHECK EVerything in five..it could be more efficient
#Will hold the average AUC value. I leave it as 'mean_kapa' as I copied the code from the Flower competition.
#(Always re-use old good code.) Pepperidge Farm remembers.
mean_auc = 0.0
# cross validation model we are going to use
kfolder=StratifiedKFold(y, n_folds=number_of_folds,shuffle=True, random_state=SEED)
i=0 # iterator counter
print ("starting cross validation with %d kfolds " % (number_of_folds)) # some words to keep you engaged
if number_of_folds>0: # this basically means "if we are doing cross validation ", sometimes my model crashed after cv and I had to skip this part in 2nd iteration
for train_index, test_index in kfolder:
# get train (set and target variable)
X_train = X[train_index]
y_train= np.array(y)[train_index]
#talk about it
print (" train size: %d. test size: %d, cols: %d " % (len(train_index) ,len(test_index) ,(X_train.shape[1]) ))
if Use_scale:
stda=StandardScaler()
X_train=stda.fit_transform(X_train)
# get validation (set and target variable)
X_cv= X[test_index]
y_cv = np.array(y)[test_index]
if Use_scale:
X_cv=stda.transform(X_cv)
# bag 10 models of the selected type :)
preds=bagged_set(X_train,y_train,model, SEED + i, bags, X_cv, update_seed=True)
# measure AUC, but keep confusing everyone by naming "kapa"
auc = roc_auc_score(y_cv,preds)
# talk about it, hoping people will not actually find out about it
print "size train: %d size cv: %d AUC (fold %d/%d): %f" % (len(train_index), len(test_index), i + 1, number_of_folds, auc)
mean_auc += auc
#update the held-out array for meta modelling
no=0
for real_index in test_index:
train_stacker[real_index]=(preds[no])
no+=1
i+=1 # update iterator
if (number_of_folds)>0: # if we did cross validation the print the results
mean_auc/=number_of_folds
print (" Average AUC: %f" % (mean_auc) ) # keep calling it AUC (because you can)
print (" printing train datasets in %s" % (meta_folder+ outset + "train.csv")) # print the hold out predictions
print ("start final modeling")
if Use_scale:
stda=StandardScaler()
X=stda.fit_transform(X)
#load test data
X_test=load_datas("xtmetahome.pkl")
if Use_scale:
X_test=stda.transform(X_test)
# giving stats never gets old
print("rows %d columns %d " % (X_test.shape[0],X_test.shape[1] ))
# bag 10 models of the selected type :)
preds=bagged_set(X, y,model, SEED, bags, X_test, update_seed=True)
X_test=None
gc.collect() # collect the garbage
#load the id variable to put it in the submission file (if we ant to submit this model in kaggle)
print (" printing test datasets in %s" % (meta_folder + outset + "test.csv"))
save_results(preds, outset+"_submission_" +str(mean_auc) + ".csv")
print("Done.")
if __name__=="__main__":
main()