質問編集履歴
1
コードの追加
test
CHANGED
File without changes
|
test
CHANGED
@@ -12,6 +12,396 @@
|
|
12
12
|
|
13
13
|
|
14
14
|
|
15
|
+
```
|
16
|
+
|
17
|
+
import numpy
|
18
|
+
|
19
|
+
import pandas
|
20
|
+
|
21
|
+
from keras.models import Sequential
|
22
|
+
|
23
|
+
from keras.layers import Dense
|
24
|
+
|
25
|
+
from keras import optimizers
|
26
|
+
|
27
|
+
from keras.wrappers.scikit_learn import KerasRegressor
|
28
|
+
|
29
|
+
from keras.wrappers.scikit_learn import KerasClassifier
|
30
|
+
|
31
|
+
from sklearn.model_selection import cross_val_score
|
32
|
+
|
33
|
+
from sklearn.model_selection import KFold
|
34
|
+
|
35
|
+
from sklearn.model_selection import GridSearchCV
|
36
|
+
|
37
|
+
from sklearn.preprocessing import StandardScaler
|
38
|
+
|
39
|
+
from sklearn.pipeline import Pipeline
|
40
|
+
|
41
|
+
from keras.models import load_model
|
42
|
+
|
43
|
+
import os
|
44
|
+
|
45
|
+
import argparse
|
46
|
+
|
47
|
+
|
48
|
+
|
49
|
+
#----------------------------
|
50
|
+
|
51
|
+
# get command line variables
|
52
|
+
|
53
|
+
#----------------------------
|
54
|
+
|
55
|
+
parser = argparse.ArgumentParser(description='Make models by keras. Place Y on the head column in the cleaned dataset with header names on the top row. Rows containing null values will be deleated.')
|
56
|
+
|
57
|
+
parser.add_argument('--mode', choices=['create', 'predict'], dest='mode', metavar='create/predict', type=str, nargs='+', required=True,
|
58
|
+
|
59
|
+
help='an integer for the accumulator')
|
60
|
+
|
61
|
+
parser.add_argument('--input_file', dest='input_file', type=str, nargs='+', required=True,
|
62
|
+
|
63
|
+
help='path to dataset or model')
|
64
|
+
|
65
|
+
parser.add_argument('--method', choices=['binary', 'multiple', 'regression'], metavar='binary/multiple/regression', dest='method', type=str, nargs='+', required=True,
|
66
|
+
|
67
|
+
help='Model type you solve')
|
68
|
+
|
69
|
+
parser.add_argument('--output_file', dest='output_file', default=False, required=False,
|
70
|
+
|
71
|
+
help='If you input output_file it will save result as directed path.')
|
72
|
+
|
73
|
+
parser.add_argument('--model_file', dest='model_file', default=False, nargs='*',
|
74
|
+
|
75
|
+
help='If you input model_file it will save or load a model.')
|
76
|
+
|
77
|
+
parser.add_argument('--definition', metavar='array of data type such as str, int and float with delimiter [,]', dest='definition', default=False, nargs='*',
|
78
|
+
|
79
|
+
help='If you define data type of columns, send array of full column definitions.')
|
80
|
+
|
81
|
+
|
82
|
+
|
83
|
+
args = parser.parse_args()
|
84
|
+
|
85
|
+
|
86
|
+
|
87
|
+
#----------------------------
|
88
|
+
|
89
|
+
# functions
|
90
|
+
|
91
|
+
#----------------------------
|
92
|
+
|
93
|
+
class MakeModel:
|
94
|
+
|
95
|
+
#init
|
96
|
+
|
97
|
+
def __init__(self, args):
|
98
|
+
|
99
|
+
self.X = self.Y = []
|
100
|
+
|
101
|
+
self.row_length = self.column_length = 0
|
102
|
+
|
103
|
+
self.method = args.method[0]
|
104
|
+
|
105
|
+
self.ifp = args.input_file[0]
|
106
|
+
|
107
|
+
|
108
|
+
|
109
|
+
if args.model_file != False:
|
110
|
+
|
111
|
+
self.mfp = args.model_file[0]
|
112
|
+
|
113
|
+
else:
|
114
|
+
|
115
|
+
self.mfp = False
|
116
|
+
|
117
|
+
|
118
|
+
|
119
|
+
if args.output_file != False:
|
120
|
+
|
121
|
+
self.ofp = args.output_file[0]
|
122
|
+
|
123
|
+
else:
|
124
|
+
|
125
|
+
self.ofp = False
|
126
|
+
|
127
|
+
|
128
|
+
|
129
|
+
if args.definition != False:
|
130
|
+
|
131
|
+
self.dfin = args.definition.split(",")
|
132
|
+
|
133
|
+
else:
|
134
|
+
|
135
|
+
self.dfin = False
|
136
|
+
|
137
|
+
|
138
|
+
|
139
|
+
#create layers
|
140
|
+
|
141
|
+
def create_model(self, evMethod, neurons, layers, act, learn_rate, cls, mtr):
|
142
|
+
|
143
|
+
# Criate model
|
144
|
+
|
145
|
+
model = Sequential()
|
146
|
+
|
147
|
+
model.add(Dense(neurons, input_dim=self.column_length, kernel_initializer='normal', activation='relu'))
|
148
|
+
|
149
|
+
for i in range(1, layers):
|
150
|
+
|
151
|
+
model.add(Dense(int(numpy.ceil(numpy.power(neurons,1/i)*2)), kernel_initializer='normal', activation='relu'))
|
152
|
+
|
153
|
+
model.add(Dense(cls, kernel_initializer='normal', activation=act))
|
154
|
+
|
155
|
+
# Compile model
|
156
|
+
|
157
|
+
adam = optimizers.Adam(lr=learn_rate)
|
158
|
+
|
159
|
+
model.compile(loss=evMethod, optimizer=adam, metrics=mtr)
|
160
|
+
|
161
|
+
return model
|
162
|
+
|
163
|
+
|
164
|
+
|
165
|
+
#load dataset
|
166
|
+
|
167
|
+
def load_dataset(self):
|
168
|
+
|
169
|
+
dataframe = pandas.read_csv(self.ifp, header=0, encoding="sjis").dropna()
|
170
|
+
|
171
|
+
if self.dfin != False:
|
172
|
+
|
173
|
+
dataframe[dataframe.columns].apply(lambda x: x.astype(self.dfin[dataframe.columns.get_loc(x.name)]))
|
174
|
+
|
175
|
+
dataframe_X = pandas.get_dummies(dataframe[dataframe.columns[1:]]) #create dummy variables
|
176
|
+
|
177
|
+
if self.method == 'multiple':
|
178
|
+
|
179
|
+
dataframe_Y = pandas.get_dummies(dataframe[dataframe.columns[0]]) #create dummy variables
|
180
|
+
|
181
|
+
else:
|
182
|
+
|
183
|
+
dataframe_Y = dataframe[dataframe.columns[0]]
|
184
|
+
|
185
|
+
#print(dataframe_Y.head())
|
186
|
+
|
187
|
+
#print(dataframe_X.head())
|
188
|
+
|
189
|
+
self.row_length, self.column_length = dataframe_X.shape
|
190
|
+
|
191
|
+
self.X = dataframe_X.values
|
192
|
+
|
193
|
+
self.Y = dataframe_Y.values
|
194
|
+
|
195
|
+
|
196
|
+
|
197
|
+
#train
|
198
|
+
|
199
|
+
def train_model(self):
|
200
|
+
|
201
|
+
#pipe to Grid Search
|
202
|
+
|
203
|
+
estimators = []
|
204
|
+
|
205
|
+
estimators.append(('standardize', StandardScaler()))
|
206
|
+
|
207
|
+
|
208
|
+
|
209
|
+
#rely on chosen method parameters
|
210
|
+
|
211
|
+
if self.method == 'binary':
|
212
|
+
|
213
|
+
evMethod = ['binary_crossentropy']
|
214
|
+
|
215
|
+
activation = ['sigmoid']
|
216
|
+
|
217
|
+
metr = [['accuracy']]
|
218
|
+
|
219
|
+
estimators.append(('mlp', KerasClassifier(build_fn=self.create_model, epochs=10, batch_size=200, verbose=1)))
|
220
|
+
|
221
|
+
cls = [1]
|
222
|
+
|
223
|
+
elif self.method == 'multiple':
|
224
|
+
|
225
|
+
evMethod = [['categorical_crossentropy']]
|
226
|
+
|
227
|
+
activation = ['softmax']
|
228
|
+
|
229
|
+
metr = [['accuracy']]
|
230
|
+
|
231
|
+
estimators.append(('mlp', KerasClassifier(build_fn=self.create_model, epochs=10, batch_size=200, verbose=1)))
|
232
|
+
|
233
|
+
cls = [self.Y.shape[1]]
|
234
|
+
|
235
|
+
else:
|
236
|
+
|
237
|
+
evMethod = ['mean_squared_error']
|
238
|
+
|
239
|
+
activation = [None]
|
240
|
+
|
241
|
+
metr = [None]
|
242
|
+
|
243
|
+
estimators.append(('mlp', KerasRegressor(build_fn=self.create_model, epochs=10, batch_size=200, verbose=1)))
|
244
|
+
|
245
|
+
cls = [1]
|
246
|
+
|
247
|
+
|
248
|
+
|
249
|
+
pipeline = Pipeline(estimators)
|
250
|
+
|
251
|
+
|
252
|
+
|
253
|
+
#test parameters
|
254
|
+
|
255
|
+
batch_size = list(set([int(numpy.ceil(self.row_length/i)) for i in [1000,300,100]]))
|
256
|
+
|
257
|
+
epochs = [10, 50, 100]
|
258
|
+
|
259
|
+
neurons = list(set([int(numpy.ceil(self.column_length/i)*2) for i in numpy.arange(1,3,0.4)]))
|
260
|
+
|
261
|
+
learn_rate = [0.001, 0.005, 0.01, 0.07]
|
262
|
+
|
263
|
+
layers = [1,2,3,4,5]
|
264
|
+
|
265
|
+
#test parameter
|
266
|
+
|
267
|
+
"""batch_size = [31]
|
268
|
+
|
269
|
+
epochs = [100]
|
270
|
+
|
271
|
+
neurons = [32]
|
272
|
+
|
273
|
+
learn_rate = [0.01]
|
274
|
+
|
275
|
+
layers = [5]"""
|
276
|
+
|
277
|
+
#execution
|
278
|
+
|
279
|
+
param_grid = dict(mlp__neurons = neurons, mlp__batch_size = batch_size, mlp__epochs=epochs, mlp__learn_rate=learn_rate, mlp__layers=layers, mlp__act=activation, mlp__evMethod=evMethod, mlp__cls=cls, mlp__mtr=metr)
|
280
|
+
|
281
|
+
grid = GridSearchCV(estimator=pipeline, param_grid=param_grid)
|
282
|
+
|
283
|
+
grid_result = grid.fit(self.X, self.Y)
|
284
|
+
|
285
|
+
|
286
|
+
|
287
|
+
#output best parameter condition
|
288
|
+
|
289
|
+
clf = []
|
290
|
+
|
291
|
+
clf = grid_result.best_estimator_
|
292
|
+
|
293
|
+
print(clf.get_params())
|
294
|
+
|
295
|
+
accuracy = clf.score(self.X, self.Y)
|
296
|
+
|
297
|
+
if self.method in ['binary', 'multiple']:
|
298
|
+
|
299
|
+
print("\nAccuracy: %.2f" % (accuracy))
|
300
|
+
|
301
|
+
else:
|
302
|
+
|
303
|
+
print("Results: %.2f (%.2f) MSE" % (accuracy.mean(), accuracy.std()))
|
304
|
+
|
305
|
+
|
306
|
+
|
307
|
+
#save model
|
308
|
+
|
309
|
+
if self.mfp != False:
|
310
|
+
|
311
|
+
clf.steps[1][1].model.save(self.mfp)
|
312
|
+
|
313
|
+
|
314
|
+
|
315
|
+
#predict dataset
|
316
|
+
|
317
|
+
def predict_ds(self):
|
318
|
+
|
319
|
+
model = load_model(self.mfp)
|
320
|
+
|
321
|
+
model.summary()
|
322
|
+
|
323
|
+
sc = StandardScaler()
|
324
|
+
|
325
|
+
self.X = sc.fit_transform(self.X)
|
326
|
+
|
327
|
+
pr_Y = model.predict(self.X)
|
328
|
+
|
329
|
+
if len([self.Y != '__null__']) > 0:
|
330
|
+
|
331
|
+
if self.method == 'binary':
|
332
|
+
|
333
|
+
predictions = [float(numpy.round(x)) for x in pr_Y]
|
334
|
+
|
335
|
+
accuracy = numpy.mean(predictions == self.Y)
|
336
|
+
|
337
|
+
print("Prediction Accuracy: %.2f%%" % (accuracy*100))
|
338
|
+
|
339
|
+
elif self.method == 'multiple':
|
340
|
+
|
341
|
+
predictions = []
|
342
|
+
|
343
|
+
for i in range(0, len(pr_Y)-1):
|
344
|
+
|
345
|
+
for j in range(0, len(pr_Y[i])-1):
|
346
|
+
|
347
|
+
predictions.append(int(round(pr_Y[i][j]) - self.Y[i][j]))
|
348
|
+
|
349
|
+
accuracy_total = len([x for x in predictions if x == 0])/len(predictions)
|
350
|
+
|
351
|
+
accuracy_tooneg = len([x for x in predictions if x == -1])/len(predictions)
|
352
|
+
|
353
|
+
accuracy_toopos = len([x for x in predictions if x == 1])/len(predictions)
|
354
|
+
|
355
|
+
print("Prediction Accuracy: %.2f%% (positive-error:%.2f%%/negative-error:%.2f%%)" % (accuracy_total*100, accuracy_tooneg*100, accuracy_toopos*100))
|
356
|
+
|
357
|
+
else:
|
358
|
+
|
359
|
+
accuracy = numpy.mean((self.Y - pr_Y)**2)
|
360
|
+
|
361
|
+
print("MSE: %.2f" % (numpy.sqrt(accuracy)))
|
362
|
+
|
363
|
+
|
364
|
+
|
365
|
+
#save predicted result
|
366
|
+
|
367
|
+
if self.ofp != False:
|
368
|
+
|
369
|
+
numpy.savetxt(self.ofp, pr_Y, fmt='%5s')
|
370
|
+
|
371
|
+
|
372
|
+
|
373
|
+
#----------------------------
|
374
|
+
|
375
|
+
# select mode
|
376
|
+
|
377
|
+
#----------------------------
|
378
|
+
|
379
|
+
m = MakeModel(args)
|
380
|
+
|
381
|
+
if args.mode == ['create']:
|
382
|
+
|
383
|
+
#make model
|
384
|
+
|
385
|
+
m.load_dataset()
|
386
|
+
|
387
|
+
m.train_model()
|
388
|
+
|
389
|
+
else:
|
390
|
+
|
391
|
+
#predict dataset
|
392
|
+
|
393
|
+
m.predict_ds()
|
394
|
+
|
395
|
+
m.load_dataset()
|
396
|
+
|
397
|
+
m.predict_ds()
|
398
|
+
|
399
|
+
|
400
|
+
|
401
|
+
```
|
402
|
+
|
403
|
+
|
404
|
+
|
15
405
|
|
16
406
|
|
17
407
|
|