forked from JanzYe/KDD_CUP_2012_Track2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
deep_fm_enough_memory.py
615 lines (493 loc) · 32.6 KB
/
deep_fm_enough_memory.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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
# -*- coding:utf-8 -*-
import pandas as pd
from sklearn.utils import shuffle
from sklearn.feature_extraction.text import sp
from deepctr.models import DeepFM
from deepctr.models import MDFM
from tensorflow.keras.callbacks import ModelCheckpoint, Callback
import auc
import numpy as np
import operator
import time
from constants import *
from tensorflow.keras import metrics
# total number of records in training.txt is almost 149600000, exactly 149639105
mode = TRAIN # train valid test
num_train = int(NUM_TRAINING * (10.0 / 11))
# num_train=300000
num_valid = int(NUM_TRAINING * (1.0 / 11))
# chunk_size = 10000000
chunk_size = 360000000
chunk_size_test = 36000000
seed = 1024
# 隐藏层单元数不是越高越好，中间有一个临界值达到最优.
# Dropout在数据量本来就很稀疏的情况下尽量不用，不同的数据集dropout表现差距比较大
embedding_size = 8 # embedding cost most of the memory, if OOM, reduce this
hidden_size = [512, 256, 256, 256, 128]
lr = 0.0005
batch_size = 18000
l2_reg_linear = 0.000001
l2_reg_embedding = 0.000001
l2_reg_deep = 0.001
def process_line(line):
records = line.strip().split(',')
y = (float(records[0]))
x = np.zeros((len(sparse_features)), dtype=np.int)
i = 0
for word in records[1:]:
x[i] = (int(word))
i += 1
return x, y
def get_dense_need_combine_feature(data, header, sum_):
ids = data[header].values
dense_need_combine_vals = sum_[ids]
# for i, id in enumerate(ids):
# if id >= sparseM.shape[0]:
# continue
# else:
# row = sparseM[int(id)]
# if row.count_nonzero() > 0:
# dense_need_combine_vals[i] = row.sum() / row.count_nonzero()
return np.array(dense_need_combine_vals)
def get_dense_need_combine_features(data):
dense_need_combine_vals = []
dense_need_combine_vals.append(get_dense_need_combine_feature(data, QUERY_ID, sum_query))
dense_need_combine_vals.append(get_dense_need_combine_feature(data, KEYWORD_ID, sum_keyword))
dense_need_combine_vals.append(get_dense_need_combine_feature(data, TITLE_ID, sum_title))
dense_need_combine_vals.append(get_dense_need_combine_feature(data, DESCRIPTION_ID, sum_description))
return dense_need_combine_vals
def get_multi_val_feature(data, header, sparseV):
# rows.nonzero()[1][rows.nonzero()[0]==0]
# ids = data[header].values
# cols = features_min_max[max_header][1]
# rows = len(ids)
# multi_vals = np.zeros((rows, cols), dtype='int')
# valid_len = np.ones((rows), dtype='int')
# for i, id in enumerate(ids):
# pos = sparseM[int(id)].nonzero()[1]
# if len(pos) < 1:
# # many id don't have relative tokens
# # print('empty tokens')
# continue
# valid_len[i] = len(pos)
# multi_vals[i, 0:valid_len[i]] = pos
ids = data[header].values
valid_len_multi_val = sparseV[ids].toarray()
valid_len = valid_len_multi_val[:, 0]
valid_len[valid_len == 0] = 1
return valid_len_multi_val[:, 1:], valid_len
def get_multi_val_features(data):
multi_vals = []
valid_lens = []
print('combining tokens_vector_query')
vals, len = get_multi_val_feature(data, QUERY_ID, tokens_multi_val_query)
multi_vals.append(vals)
valid_lens.append(len)
print('combining tokens_vector_keyword')
vals, len = get_multi_val_feature(data, KEYWORD_ID, tokens_multi_val_keyword)
multi_vals.append(vals)
valid_lens.append(len)
print('combining tokens_vector_title')
vals, len = get_multi_val_feature(data, TITLE_ID, tokens_multi_val_title)
multi_vals.append(vals)
valid_lens.append(len)
print('combining tokens_vector_description')
vals, len = get_multi_val_feature(data, DESCRIPTION_ID, tokens_multi_val_description)
multi_vals.append(vals)
valid_lens.append(len)
return multi_vals, valid_lens
def generate_arrays_from_file(path_combined_mapped):
reader = pd.read_csv(path_combined_mapped, chunksize=chunk_size)
print('generating start')
for df in reader:
print('\ndf size: %d' % df.shape[0])
df = shuffle(df, random_state=seed)
len_df=df.shape[0]
if len(multi_val_features_dim) > 0:
print('combining multi_val features ......')
multi_vals, valid_lens = get_multi_val_features(df)
else:
multi_vals = []
valid_lens = []
if len(dense_need_combine_features) > 0:
print('combining get_dense_need_combine_features ......')
# sum of idf
dense_need_combine_val = get_dense_need_combine_features(df)
# mean of idf
if len(valid_lens) > 0:
dense_need_combine_val.extend([dense_need_combine_val[i] / valid_lens[i]
for i in range(len(valid_lens))])
else:
dense_need_combine_val = []
print('scaling features ...')
for key in features_min_max:
min_val = features_min_max[key][0]
max_val = features_min_max[key][1]
df[key] = (df[key] - min_val) / (max_val - min_val)
print(df[key].values[0])
# print(multi_vals)
x_train = [df[key].values[0:num_train] for key in sparse_features] + \
[df[key].values[0:num_train] for key in dense_features] + \
[feat[0:num_train] for feat in dense_need_combine_val] + \
[vec[0:num_train] for vec in multi_vals] + [val_len[0:num_train] for val_len in valid_lens]
y_train = [df[key].values[0:num_train] for key in target]
x_valid = [df[key].values[num_train:] for key in sparse_features] + \
[df[key].values[num_train:] for key in dense_features] + \
[feat[num_train:] for feat in dense_need_combine_val] + \
[vec[num_train:] for vec in multi_vals] + [val_len[num_train:] for val_len in valid_lens]
y_valid = [df[key].values[num_train:] for key in target]
return x_train, y_train, x_valid, y_valid
print('generating complete')
class AucCallback(Callback):
def __init__(self, clicks, imps, datas, batch_size):
self.clicks = clicks
self.imps = imps
self.datas = datas
self.batch_size = batch_size
self.auc = []
def on_epoch_end(self, epoch, logs=None):
print('predicting valid ......')
preds = self.model.predict(self.datas, self.batch_size, verbose=1)
print('calculating auc ......')
Auc = auc.scoreClickAUC(self.clicks, self.imps, preds)
self.auc.append(str(Auc))
print('scoreClickAUC: %s' % (','.join(self.auc)))
def step_decay(epoch):
initial_lrate = lr
drop = 0.5
epochs_drop = 10.0
lrate = initial_lrate * math.pow(drop,math.floor((1+epoch)/epochs_drop))
return lrate
if __name__ == "__main__":
# dense_features = [
# # 1 2 3 4 5 6 7
# 'DisplayURL', 'AdID', 'AdvertiserID', 'Depth', 'Position', 'QueryID', 'KeywordID',
# # 8 9 10 12 13
# 'TitleID', 'DescriptionID', 'UserID', 'Gender', 'Age', 'RelativePosition',
#
# # 14 15 16 17 18 19 20
# 'group_Ad', 'group_Advertiser', 'group_Query', 'group_Keyword', 'group_Title', 'group_Description', 'group_User',
#
# # 21 22 23 24 25
# 'aCTR_Ad', 'aCTR_Advertiser', 'aCTR_Depth', 'aCTR_Position', 'aCTR_RPosition',
#
# # 26 27 28 29 30 31 32
# 'pCTR_Url', 'pCTR_Ad', 'pCTR_Advertiser', 'pCTR_Query', 'pCTR_Keyword', 'pCTR_Title', 'pCTR_Description',
# # 33 34 35 36
# 'pCTR_User', 'pCTR_Gender', 'pCTR_Age', 'pCTR_RPosition',
#
# # 37 38 39
# 'num_Depth', 'num_Position', 'num_RPosition',
# # 40 41 42 43
# 'num_Query', 'num_Keyword', 'num_Title', 'num_Description',
# # 44 45 46 47 48
# 'num_Imp__Ad', 'num_Imp__Advertiser', 'num_Imp_Depth', 'num_Imp_Position', 'num_Imp_RPosition'
# ]
dense_features = [
'DisplayURL', 'AdID', 'AdvertiserID', 'QueryID', 'KeywordID', 'TitleID', 'DescriptionID',
'Depth', 'Position', 'Gender', 'Age', 'RelativePosition',
#
# # 14 15 16 17 18 19 20
'group_Ad', 'group_Advertiser', 'group_Query', 'group_Keyword', 'group_Title', 'group_Description', 'group_User',
#
# # 21 22 23 24
#
# # 26 27 28 29 30 31 32
'pCTR_Url', 'pCTR_Ad', 'pCTR_Advertiser', 'pCTR_Query', 'pCTR_Keyword', 'pCTR_Title', 'pCTR_Description',
# # 33 34 35 36
'pCTR_User', 'pCTR_Gender', 'pCTR_Age', 'pCTR_RPosition',
'num_Depth', 'num_Position', 'num_RPosition',
'num_Query', 'num_Keyword', 'num_Title', 'num_Description',
]
dense_need_combine_features = [
'sum_Query', 'sum_Keyword', 'sum_Title', 'sum_Description', # sum of idf
'mean_Query', 'mean_Keyword', 'mean_Title', 'mean_Description', # mean of idf
]
# dense_need_combine_features = []
dense_features_complete = []
dense_features_complete.extend(dense_features)
dense_features_complete.extend(dense_need_combine_features)
min_max = pd.read_csv(PATH_MIN_MAX, dtype=float)
features_min_max = {key: min_max[key].values for key in [
'DisplayURL', 'AdID', 'AdvertiserID', 'QueryID', 'KeywordID', 'TitleID', 'DescriptionID',
'Depth', 'Position', 'Gender', 'Age',
#
# # 14 15 16 17 18 19 20
'group_Ad', 'group_Advertiser', 'group_Query', 'group_Keyword', 'group_Title', 'group_Description',
'group_User',
'num_Depth', 'num_Position', 'num_RPosition',
'num_Query', 'num_Keyword', 'num_Title', 'num_Description',
]}
sparse_features = [
# 49 50 51 52 53
'sparse_Url', 'sparse_Ad', 'sparse_Advertiser', 'sparse_Depth', 'sparse_Position',
# 54 55 56 57 58 59
'sparse_Query', 'sparse_Keyword', 'sparse_Title', 'sparse_Description', 'sparse_UserID', 'sparse_Gender', 'sparse_Age',
'sparse_PosDepth']
print('summing tokens_vector_title ......')
sum_title = pd.read_csv(PATH_SUM_TITLE, header=None, dtype=np.float)[0].values
print(sum_title.shape)
print('summing tokens_vector_query ......')
sum_query = pd.read_csv(PATH_SUM_QUERY, header=None, dtype=np.float)[0].values
print(sum_query.shape)
print('summing tokens_vector_keyword ......')
sum_keyword = pd.read_csv(PATH_SUM_KEYWORD, header=None, dtype=np.float)[0].values
print('summing tokens_vector_description ......')
sum_description = pd.read_csv(PATH_SUM_DESCRIPTION, header=None, dtype=np.float)[0].values
print(sum_description.shape)
multi_val_features_dim = {
'vec_Query': [int(features_min_max['num_Query'][1]), N_WORDS_QUERY],
'vec_Keyword': [int(features_min_max['num_Keyword'][1]), N_WORDS_KEYWORD],
'vec_Title': [int(features_min_max['num_Title'][1]), N_WORDS_TITLE],
'vec_Description': [int(features_min_max['num_Description'][1]), N_WORDS_DESCRIPTION],
}
# multi_val_features_dim = {}
print('loading tokens_multi_val ......')
tokens_multi_val_query = sp.load_npz(PATH_MUL_QUERY)
tokens_multi_val_title = sp.load_npz(PATH_MUL_TITLE)
tokens_multi_val_keyword = sp.load_npz(PATH_MUL_KEYWORD)
tokens_multi_val_description = sp.load_npz(PATH_MUL_DESCRIPTION)
exclude = ['']
target = ['Click', 'Impression']
# 2.count #unique features for each sparse field
sparse_feature_dim = {}
with open('./data/features_infos_combined.txt') as fr:
# with open('./data/sample/features_infos.txt') as fr:
for line in fr:
records = line.strip().split(':')
if records[0] in exclude:
continue
sparse_feature_dim[records[0]] = int(records[1])
fr.close()
# 4.Define Model,compile and train
# model = DeepFM({"sparse": sparse_feature_dim, "dense": dense_features}, embedding_size=embedding_size,
# hidden_size=hidden_size,
# final_activation='sigmoid')
model = MDFM({"sparse": sparse_feature_dim, "dense": dense_features_complete,
'multi_val': multi_val_features_dim},
embedding_size=embedding_size,
hidden_size=hidden_size,
final_activation='sigmoid',
l2_reg_linear=l2_reg_linear, l2_reg_embedding=l2_reg_embedding, l2_reg_deep=l2_reg_deep)
if mode == TRAIN:
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=[metrics.binary_crossentropy])
now_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
filepath = 'model_save/deep_fm_combined-ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}-bs' + str(batch_size)\
+ '-ee' + str(embedding_size) + '-hz' + str(hidden_size) \
+ '-l2l' + str(l2_reg_linear) + '-l2e' + str(l2_reg_embedding) + '-l2d' + str(l2_reg_deep) \
+ '-t' + now_time + '.h5'
x_train, y_train, x_valid, y_valid = generate_arrays_from_file(PATH_TRAIN)
checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='min',
save_weights_only=True)
auc_eval = AucCallback(y_valid[0], y_valid[1], x_valid, batch_size)
print((y_valid[0] / y_valid[1]))
#lrate = LearningRateScheduler(step_decay)
history = model.fit(x=x_train, y=y_train[0] / y_train[1],
steps_per_epoch=int(np.ceil(num_train/batch_size)),
callbacks=[checkpoint, auc_eval], epochs=100, verbose=1,
validation_data=(x_valid, y_valid[0] / y_valid[1]),
validation_steps=int(np.ceil(num_valid/batch_size)), shuffle=True)
elif mode == VALID:
# model.load_weights('model_save/deep_fm_fn-ep002-loss0.148-val_loss0.174.h5') # auc: 0.718467 batch_size=6000
#model.load_weights('model_save/deep_fm_fn-ep001-loss0.149-val_loss0.175.h5') # auc: 0.714243 batch_size = 2048
# model.load_weights('model_save/deep_fm_fn-ep005-loss0.147-val_loss0.173.h5') # auc: 0.722535 batch_size = 10000
# model.load_weights('model_save/deep_fm_fn_bs10000-ep001-loss0.155-val_loss0.153.h5') # auc: 0.738023
#model.load_weights('model_save/deep_fm_fn_bs15000-ep001-loss0.156-val_loss0.152.h5') # auc: 0.739935
#model.load_weights('model_save/deep_fm_fn-ep002-loss0.154-val_loss0.154-bs15000-ee20-hz[128, 128].h5') # auc: 0.741590
# model.load_weights('model_save/deep_fm_fn-ep020-loss0.153-val_loss0.153-bs15000-ee20-hz[5, 600].h5') # auc: 0.742558
#model.load_weights('model_save/deep_fm_combined-ep001-loss3.077-val_loss0.627-bs6000-ee8-hz[128, 128].h5') # auc: 0.49
# model.load_weights('model_save/deep_fm_combined-ep009-loss0.134-val_loss0.134-bs15000-ee20-hz[3, 600].h5') # auc: 0.876005
# add pctr, group, len of token, age, depth, position, rposition, num_imp
#model.load_weights('model_save/deep_fm_combined-ep004-loss0.141-val_loss0.141-bs15000-ee20-hz[3, 600]-t2018-12-18 18:35:59.h5') # auc: 0.834878
# add pctr, group, len of token, age, depth, position, rposition, num_occurs, tokens_vector
# model.load_weights('model_save/deep_fm_combined-ep007-loss0.140-val_loss0.140-bs18000-ee20-hz[3, 500]-t2018-12-25 18:44:01.h5') # auc: 0.837464
# model.load_weights('model_save/deep_fm_combined-ep004-loss0.139-val_loss0.139-bs18000-ee8-hz[512, 256, 256, 256, 128]-t2018-12-26 21:32:47.h5') # auc:838491
#model.load_weights('model_save/deep_fm_combined-ep005-loss0.139-val_loss0.139-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e1e-05-l2d0.0001-t2018-12-28 08:39:49.h5') # auc: 0.839753
#model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-02 20:18:42.h5' # auc: 0.869702 0.
#model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-03 03:45:50.h5' # auc: 0.869080 0.
model_name = 'deep_fm_combined-ep010-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-03 03:45:50.h5' # auc: 0.874691 0.784256
model.load_weights('model_save/' + model_name)
_, _, x_valid, y_valid = generate_arrays_from_file(PATH_TRAIN)
preds = model.predict(x_valid, batch_size, verbose=1)
print('calculating auc ......')
AUC = auc.scoreClickAUC(y_valid[0], y_valid[1], preds)
print('scoreClickAUC: %f' % AUC)
elif mode == TEST:
now_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print(now_time)
# if the loss changes less than 0.001, stops training, does not update model
# model.load_weights('model_save/deep_fm_fn_bs10000-ep001-loss0.155-val_loss0.153.h5') # auc: 0.714774
#model.load_weights('model_save/deep_fm_fn_bs15000-ep001-loss0.156-val_loss0.152.h5') # auc: 0.717083
#model.load_weights('model_save/deep_fm_fn-ep002-loss0.154-val_loss0.154-bs15000-ee20-hz[128, 128].h5') # auc: 0.718581
#model.load_weights('model_save/deep_fm_fn-ep020-loss0.153-val_loss0.153-bs15000-ee20-hz[5, 600].h5') # auc: 0.719317
#model.load_weights('model_save/deep_fm_fn-ep043-loss0.152-val_loss0.152-bs15000-ee20-hz[3, 600].h5') # auc: 0.722419
# add dense feature pCTR: over fitting
# model.load_weights('model_save/deep_fm_combined-ep009-loss0.134-val_loss0.134-bs15000-ee20-hz[3, 600].h5') # auc: 0.733984
# model.load_weights('model_save/deep_fm_combined-ep001-loss0.147-val_loss0.139-bs15000-ee20-hz[3, 600].h5') # auc: 0.744694
#model.load_weights('model_save/deep_fm_combined-ep003-loss0.135-val_loss0.135-bs15000-ee20-hz[3, 600].h5') # auc: 0.733826
#model.load_weights('model_save/deep_fm_combined-ep003-loss0.132-val_loss0.132-bs15000-ee8-hz[3, 600].h5') # auc: 0.735597
#model.load_weights('model_save/deep_fm_combined-ep001-loss0.144-val_loss0.135-bs15000-ee8-hz[3, 600].h5') # auc: 0.743677
# add pCTR and aCTR
#model.load_weights('model_save/deep_fm_combined-ep011-loss0.135-val_loss0.135-bs15000-ee20-hz[3, 600].h5') # auc: 0.738687
#model.load_weights('model_save/deep_fm_combined-ep001-loss0.150-val_loss0.141-bs15000-ee20-hz[3, 600].h5') # auc: 0.737412
#model.load_weights('model_save/deep_fm_combined-ep002-loss0.140-val_loss0.139-bs15000-ee20-hz[3, 600].h5') # auc: 0.736510
# add pCTR and group
#model.load_weights('model_save/deep_fm_combined-ep003-loss0.142-val_loss0.142-bs15000-ee20-hz[3, 600].h5') # auc: 0.746002
#model.load_weights('model_save/deep_fm_combined-ep001-loss0.165-val_loss0.146-bs15000-ee20-hz[3, 600].h5') # auc: 0.739382
# add pCTR and group and len of tokens
# model.load_weights('model_save/deep_fm_combined-ep003-loss0.142-val_loss0.142-bs15000-ee20-hz[3, 600].h5') # auc: 0.748758
# add pCTR and len of tokens
# model.load_weights('model_save/deep_fm_combined-ep012-loss0.134-val_loss0.134-bs15000-ee20-hz[3, 600]-t2018-12-17 17:09:17.h5') # auc: 0.717313
# model.load_weights('model_save/deep_fm_combined-ep001-loss0.151-val_loss0.141-bs15000-ee20-hz[3, 600]-t2018-12-17 17:09:17.h5') # auc: 0.739057
# add len of tokens
# model.load_weights('model_save/deep_fm_combined-ep013-loss0.153-val_loss0.153-bs15000-ee20-hz[3, 600]-t2018-12-18 00:59:06.h5') # auc: 0.72
# add pctr, group, len of token, age, depth, position, rposition
# model.load_weights('model_save/deep_fm_combined-ep003-loss0.142-val_loss0.142-bs15000-ee20-hz[3, 600]-t2018-12-18 12:48:01.h5') # auc: 0.749360
# add pctr, group, len of token, age, depth, position, rposition, num_imp
# model.load_weights('model_save/deep_fm_combined-ep004-loss0.141-val_loss0.141-bs15000-ee20-hz[3, 600]-t2018-12-18 18:35:59.h5') # auc: 0.739763
# add pctr, group, len of token, age, depth, position, rposition, num_occurs
#model.load_weights('model_save/deep_fm_combined-ep003-loss0.142-val_loss0.142-bs15000-ee20-hz[3, 600]-t2018-12-18 23:26:58.h5') # auc: 0.750127
# model.load_weights('model_save/deep_fm_combined-ep003-loss0.141-val_loss0.141-bs18000-ee20-hz[3, 600]-t2018-12-19 11:20:59.h5') # auc: 0.753482
# model.load_weights('model_save/deep_fm_combined-ep003-loss0.142-val_loss0.141-bs18000-ee20-hz[3, 500]-t2018-12-19 16:12:38.h5') # auc: 0.755629
# model.load_weights('model_save/deep_fm_combined-ep004-loss0.141-val_loss0.141-bs18000-ee20-hz[3, 300]-t2018-12-19 20:46:10.h5') # auc: 0.751999
# model.load_weights('model_save/deep_fm_combined-ep005-loss0.140-val_loss0.141-bs18000-ee20-hz[3, 400]-t2018-12-20 10:14:14.h5') # auc: 0.752596
# model.load_weights('model_save/deep_fm_combined-ep010-loss0.140-val_loss0.140-bs18000-ee25-hz[3, 500]-t2018-12-20 16:05:34.h5') # auc: 0.753794
# add pctr, group, len of token, age, depth, position, rposition, num_occurs, sum_idf
# model.load_weights('model_save/deep_fm_combined-ep006-loss0.140-val_loss0.140-bs18000-ee20-hz[3, 500]-t2018-12-21 17:17:20.h5') # auc: 0.753753
# add pctr, group, len of token, age, depth, position, rposition, num_occurs, tokens_vector
#model.load_weights('model_save/deep_fm_combined-ep003-loss0.142-val_loss0.141-bs18000-ee20-hz[3, 500]-t2018-12-22 23:35:54.h5') # auc: 0.761872
# model.load_weights('model_save/deep_fm_combined-ep007-loss0.140-val_loss0.140-bs18000-ee20-hz[3, 500]-t2018-12-25 18:44:01.h5') # auc: 0.762638
# add pctr, group, len of token, age, depth, position, rposition, num_occurs, tokens_vector, id_val
#model.load_weights('model_save/deep_fm_combined-ep004-loss0.139-val_loss0.139-bs18000-ee8-hz[512, 256, 256, 256, 128]-t2018-12-26 21:32:47.h5') # auc: 0.769375 778695
# model.load_weights('model_save/deep_fm_combined-ep005-loss0.139-val_loss0.139-bs18000-ee8-hz[512, 256, 128]-l2l1e-05-l2e1e-05-l2d0-t2018-12-27 13:41:15.h5') # auc: 0.768785 778826
#model.load_weights('model_save/deep_fm_combined-ep006-loss0.138-val_loss0.138-bs30000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e1e-05-l2d0.0-t2018-12-27 17:20:56.h5') # auc: 0.769178 778902
#model.load_weights('model_save/deep_fm_combined-ep005-loss0.139-val_loss0.139-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e1e-05-l2d1e-05-t2018-12-27 21:42:03.h5') # auc: 0.769245 780172
# model.load_weights('model_save/deep_fm_combined-ep005-loss0.139-val_loss0.139-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e1e-05-l2d0.0001-t2018-12-28 08:39:49.h5') # auc: 0.771686 780086
# model.load_weights('model_save/deep_fm_combined-ep004-loss0.139-val_loss0.139-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e1e-05-l2d0.001-t2018-12-28 13:39:25.h5') # auc: 0.769395 780504
#model.load_weights('model_save/deep_fm_combined-ep003-loss0.141-val_loss0.141-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e0.0001-l2d0.001-t2018-12-28 19:11:03.h5') # auc: 0.758987 0.770800
# model.load_weights('model_save/deep_fm_combined-ep002-loss0.134-val_loss0.138-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e0.0-l2d0.001-t2018-12-28 22:40:46.h5') # auc: 0.759393 0.770754
#model.load_weights('model_save/deep_fm_combined-ep006-loss0.137-val_loss0.137-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-05-l2e1e-06-l2d0.001-t2018-12-29 07:15:55.h5') # auc: 0.775155 0.781973
#model.load_weights('model_save/deep_fm_combined-ep013-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2018-12-29 12:56:43.h5') # auc: 0.774949 0.783688
# model.load_weights('model_save/deep_fm_combined-ep009-loss0.117-val_loss0.124-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l0.0-l2e1e-06-l2d0.001-t2018-12-30 00:45:54.h5') # auc: 0.763515 0.772781
# model.load_weights('model_save/deep_fm_combined-ep014-loss0.122-val_loss0.123-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-07-l2e1e-06-l2d0.001-t2018-12-30 09:43:53.h5') # auc: 0.769962 0.778210
# add group, len of token, age, depth, position, rposition, num_occurs, tokens_vector
#model.load_weights('model_save/deep_fm_combined-ep004-loss0.150-val_loss0.151-bs18000-ee20-hz[3, 500]-t2018-12-26 10:14:01.h5') # auc: 0.742817
# NEW DATA
# model_name = 'deep_fm_combined-ep013-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2018-12-31 00:27:56.h5' # auc: 0.773103 0.782441
# model_name = 'deep_fm_combined-ep009-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2018-12-31 11:40:59.h5' # auc: 0.774968 0.785623
# model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2018-12-31 11:40:59.h5' # auc: 0.774931 0.785993
# model_name = 'deep_fm_combined-ep005-loss0.129-val_loss0.128-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2018-12-31 11:40:59.h5' # auc: 0.774381 0.785257
# new data
# model_name = 'deep_fm_combined-ep009-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2018-12-31 22:51:48.h5' # auc: 0.773548 0.783240
# model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2018-12-31 22:51:48.h5' # auc: 0.774124 0.784411
# new data
# model_name = 'deep_fm_combined-ep009-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-01 11:52:39.h5' # auc: 0.772808 0.783461
# model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-01 11:52:39.h5' # auc: 0.773789 0.784705
# new data
# model_name = 'deep_fm_combined-ep011-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-01 20:34:54.h5' # auc: 0.772387 0.781246
# model_name = 'deep_fm_combined-ep008-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-01 20:34:54.h5' # auc: 0.772418 0.782084
# model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-01 20:34:54.h5' # auc: 0.772749 0.782517
# new data
# model_name = 'deep_fm_combined-ep010-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-02 10:21:56.h5' # auc: 0.774124 0.785176
# model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-02 10:21:56.h5' # auc: 0.775915 0.786094
# new data
#model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-02 20:18:42.h5' # auc: 0.773678 0.784004
# whole data, stop at epoch 6 base on exp
# model_name = '.h5' # auc: 0. 0.
# add pctr, group, len of token, age, depth, position, rposition, num_occurs, tokens_vector, id_val, sum_mean_idf
model_name = 'deep_fm_combined-ep010-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-03 03:45:50.h5' # auc: 0.774472 0.784256
#model_name = 'deep_fm_combined-ep006-loss0.127-val_loss0.127-bs18000-ee8-hz[512, 256, 256, 256, 128]-l2l1e-06-l2e1e-06-l2d0.001-t2019-01-03 03:45:50.h5' # auc: 0.773730 0.784448
model.load_weights('model_save/' + model_name)
ctr = []
click = []
imp = []
reader = pd.read_csv(PATH_SOLUTION, chunksize=chunk_size_test)
for df in reader:
print('df size: %d' % df.shape[0])
cnt = 0
while cnt < df.shape[0]:
end = cnt + batch_size
if end > df.shape[0]:
end = df.shape[0]
batch = df[cnt:end]
# Y = np.array(batch['clicks'].values, dtype=float) / np.array(batch['impressions'].values, dtype=float)
# ctr.extend(Y)
click.extend(batch['clicks'].values)
imp.extend(batch['impressions'].values)
cnt += batch_size
if cnt % (batch_size * 100):
print(click[0])
preds = []
# labels = ctr
reader = pd.read_csv(PATH_TEST, chunksize=chunk_size_test)
idf = 0
for df in reader:
len_df = df.shape[0]
print('\ndf size: %d' % len_df)
ctr_chunk = ctr[idf:idf+len_df]
# 这样打乱测试集没有提升效果
#random_state = random.randint(0, chunk_size)
#ctr_chunk = shuffle(ctr_chunk, random_state=random_state)
#data = shuffle(df, random_state=random_state)
data = df
#labels.extend(ctr_chunk)
idf += len_df
cnt = 0
if len(multi_val_features_dim) > 0:
print('combining multi_val features ......')
multi_vals, valid_lens = get_multi_val_features(df)
else:
multi_vals = []
valid_lens = []
if len(dense_need_combine_features) > 0:
print('combining get_dense_need_combine_features ......')
# sum of idf
dense_need_combine_val = get_dense_need_combine_features(df)
# mean of idf
if len(valid_lens) > 0:
dense_need_combine_val.extend([dense_need_combine_val[i] / valid_lens[i]
for i in range(len(valid_lens))])
else:
dense_need_combine_val = []
print('scaling features ...')
for key in features_min_max:
min_val = features_min_max[key][0]
max_val = features_min_max[key][1]
data[key] = np.clip((data[key] - min_val) / (max_val - min_val), 0, 1)
print(data[key].values[0])
while cnt < len_df:
end = cnt + batch_size
if end > len_df:
end = len_df
batch = data[cnt:end]
X = [batch[key].values for key in sparse_features] + \
[batch[key].values for key in dense_features] + \
[feat[cnt: end] for feat in dense_need_combine_val] + \
[vec[cnt: end] for vec in multi_vals] + [len[cnt: end] for len in valid_lens]
pred = model.predict_on_batch(X)
preds.extend(pred.flatten().tolist())
cnt += batch_size
if cnt % (batch_size * 100):
print(pred[0])
now_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print(now_time)
print('calculating auc ......')
print('labels； %d' % len(click))
print('preds: %d' % len(preds))
AUC = auc.auc(np.array(click, dtype=np.float) / np.array(imp, dtype=np.float), preds)
print('auc: %f' % AUC)
AUC = auc.scoreClickAUC(click, imp, preds)
print('scoreClickAUC: %f' % AUC)
now_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print(now_time)
# writing preds to csv
with open('data/' + model_name + '.csv', 'w') as fw:
for i in range(len(preds)):
if i % 1000000 == 0:
print('label: %f, pred: %f' % (click[i]/imp[i], preds[i]))
to_write = str(preds[i]) + '\n'
fw.write(to_write)
fw.close()
now_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print(now_time)
print("demo done")