-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_workstation.py
464 lines (424 loc) · 16.9 KB
/
main_workstation.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
# for some tf warnings
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
import argparse
import torch
import random
import numpy as np
import pandas as pd
from collections import defaultdict
# from deepctr_torch.inputs import (DenseFeat, SparseFeat, VarLenSparseFeat,
# get_feature_names)
# from deepctr_torch.models.din import DIN
from sklearn.metrics import roc_auc_score
from inputs import (DenseFeat, SparseFeat, VarLenSparseFeat,
get_feature_names)
from din import DIN
random.seed(10)
np.random.seed(10)
# process features into format for DIN
def process_features_din(
mode,
data_type,
feature_type,
sparse_feature_path,
hist_feature_path,
hist_feature_type,
split=0.2,
verbose=False,
):
# loading multiple files
if type(sparse_feature_path) == list:
for i in range(len(sparse_feature_path)):
cur_sparse_feature_path = sparse_feature_path[i]
cur_hist_feature_path = hist_feature_path[i]
# loaded features keys can be found in process_data.py
cur_sparse_features = pd.read_csv(
cur_sparse_feature_path,
engine='python',
header='infer',
)
# IC/UC features
cur_hist_features = np.load(cur_hist_feature_path, allow_pickle=True)
# if first path
if i == 0:
sparse_features = cur_sparse_features
hist_features = {}
for array_name in cur_hist_features.files:
hist_features[array_name] = cur_hist_features[array_name]
else:
# concatenate
sparse_features = pd.concat(
objs=[sparse_features, cur_sparse_features],
axis=0,
join="outer", # outer for union
ignore_index=False,
keys=None,
levels=None,
names=None,
verify_integrity=False,
copy=True,
)
for array_name in hist_features.keys():
hist_features[array_name] = np.concatenate(
(hist_features[array_name], cur_hist_features[array_name])
)
else:
# loaded features keys can be found in process_data.py
sparse_features = pd.read_csv(
sparse_feature_path,
engine='python',
header='infer',
)
# IC/UC features
hist_features = np.load(hist_feature_path, allow_pickle=True)
# users
user_id = sparse_features['user_id'].to_numpy()
if data_type == '1M':
gender = sparse_features['gender'].to_numpy()
age = sparse_features['age'].to_numpy()
occupation = sparse_features['occupation'].to_numpy()
# movies
movie_id = sparse_features['movie_id'].to_numpy() # 0 is mask value
score = sparse_features['rating'].to_numpy()
# movie_name = sparse_features['movie_name'].to_numpy()
# genre = sparse_features['genre'].to_numpy()
# ic/uc features
if hist_feature_type == 'IC':
positive_behavior_feature = hist_features['positive_ic_feature'].astype(int)
positive_behavior_length = hist_features['positive_ic_feature_length'].astype(int)
negative_behavior_feature = hist_features['negative_ic_feature'].astype(int)
negative_behavior_length = hist_features['negative_ic_feature_length'].astype(int)
elif hist_feature_type == 'UC':
positive_behavior_feature = hist_features['positive_uc_feature'].astype(int)
positive_behavior_length = hist_features['positive_uc_feature_length'].astype(int)
negative_behavior_feature = hist_features['negative_uc_feature'].astype(int)
negative_behavior_length = hist_features['negative_uc_feature_length'].astype(int)
else:
raise Exception(f'Unrecognized feature type {hist_feature_type}')
if len(positive_behavior_feature) != len(positive_behavior_feature):
raise Exception("History data length not matched")
# Make sure that the sparse and IC/UC features should have the same length
if len(sparse_features) != len(positive_behavior_feature):
raise Exception(
f"Sparse ({len(sparse_features)}) and IC/UC ({len(positive_behavior_feature)}) features should have the same length"
)
# labels
labels = sparse_features['labels'].to_numpy()
# DNN feature columns for the deep part of DIN
# duplicate user_id and movie_id for both positive and negative
if data_type == '1M':
feature_columns = [
SparseFeat('positive_user_id', len(user_id), embedding_dim=32),
SparseFeat('negative_user_id', len(user_id), embedding_dim=32),
SparseFeat('gender', 2, embedding_dim=8),
SparseFeat('age', 57, embedding_dim=8),
SparseFeat('occupation', 21, embedding_dim=8),
SparseFeat('positive_movie_id', len(movie_id)+1, embedding_dim=32), # 0 is mask value
SparseFeat('negative_movie_id', len(movie_id)+1, embedding_dim=32), # 0 is mask value
DenseFeat('score', 1),
# SparseFeat('movie_name', len(set(movie_name)), embedding_dim=8),
# SparseFeat('genre', len(set(genre)), embedding_dim=8),
]
else:
feature_columns = [
SparseFeat('positive_user_id', len(user_id), embedding_dim=32),
SparseFeat('negative_user_id', len(user_id), embedding_dim=32),
SparseFeat('positive_movie_id', len(movie_id)+1, embedding_dim=32), # 0 is mask value
SparseFeat('negative_movie_id', len(movie_id)+1, embedding_dim=32), # 0 is mask value
DenseFeat('score', 1),
# SparseFeat('movie_name', len(set(movie_name)), embedding_dim=8),
# SparseFeat('genre', len(set(genre)), embedding_dim=8),
]
# ic/uc feature
# list to indicate sequence sparse field
if feature_type == 'IC':
behavior_feature_list = [
'positive_movie_id',
'negative_movie_id',
]
elif feature_type == 'UC':
behavior_feature_list = [
'positive_user_id',
'negative_user_id'
]
feature_columns += [
VarLenSparseFeat(
SparseFeat(
f'hist_{behavior_feature_list[0]}',
len(positive_behavior_feature) + 1,
embedding_dim=32
),
maxlen=max(positive_behavior_length),
length_name='positive_seq_length',
),
VarLenSparseFeat(
SparseFeat(
f'hist_{behavior_feature_list[1]}',
len(negative_behavior_feature) + 1,
embedding_dim=32
),
maxlen=max(negative_behavior_length),
length_name='negative_seq_length'
),
]
# feature dictrionary
if data_type == '1M':
feature_dict = {
'positive_user_id': user_id,
'negative_user_id': user_id,
'gender': gender,
'age': age,
'occupation': occupation,
'positive_movie_id': movie_id,
'negative_movie_id': movie_id,
'score': score,
# 'movie_name': movie_name,
# 'genre': genre,
f'hist_{behavior_feature_list[0]}': positive_behavior_feature,
'positive_seq_length': positive_behavior_length,
f'hist_{behavior_feature_list[1]}': negative_behavior_feature,
'negative_seq_length': negative_behavior_length,
}
else:
feature_dict = {
'positive_user_id': user_id,
'negative_user_id': user_id,
'positive_movie_id': movie_id,
'negative_movie_id': movie_id,
'score': score,
# 'movie_name': movie_name,
# 'genre': genre,
f'hist_{behavior_feature_list[0]}': positive_behavior_feature,
'positive_seq_length': positive_behavior_length,
f'hist_{behavior_feature_list[1]}': negative_behavior_feature,
'negative_seq_length': negative_behavior_length,
}
if verbose:
print('Feature dict includes:')
for name in get_feature_names(feature_columns):
print(name, feature_dict[name].dtype)
# train/val or test split based on users
# get unique user ids and sample
unique_user_ids = np.array(list(set(user_id)))
num_train_users = int(len(unique_user_ids) * (1 - split))
train_user_ids = np.random.choice(unique_user_ids, size=num_train_users, replace=False)
val_user_ids = np.array(
[val_id for val_id in unique_user_ids if val_id not in train_user_ids]
)
temp_train_indices = [np.where(user_id == cur_id) for cur_id in train_user_ids]
train_indices = []
for cur_set in temp_train_indices:
for cur_id in cur_set[0]:
train_indices.append(cur_id)
temp_val_indices = [np.where(user_id == cur_id) for cur_id in val_user_ids]
val_indices = []
for cur_set in temp_val_indices:
for cur_id in cur_set[0]:
val_indices.append(cur_id)
# select features associated with selected user ids
if verbose:
print(
f'{len(unique_user_ids)} users splitted into {len(train_user_ids)} training users and {len(val_user_ids)} val/test users'
)
if mode == 'train':
# get all the data with associated users
train_input = {
name: feature_dict[name][train_indices]
for name in get_feature_names(feature_columns)
}
train_label = labels[train_indices]
val_input = {
name: feature_dict[name][val_indices]
for name in get_feature_names(feature_columns)
}
val_label = labels[val_indices]
return train_input, train_label, val_input, val_label, feature_columns, behavior_feature_list
elif mode == 'test':
test_input = {
name: feature_dict[name][val_indices]
for name in get_feature_names(feature_columns)
}
test_label = labels[val_indices]
return test_input, test_label, feature_columns, behavior_feature_list
if __name__ == "__main__":
# input arguments
parser = argparse.ArgumentParser()
# mode as either train or test
parser.add_argument(
'--mode', action='store', nargs=1, dest='mode', required=True
)
# sum or attenntion
parser.add_argument(
'--model_type', action='store', nargs=1, dest='model_type', required=True
)
# 1M, 10M, 20M or 25M
parser.add_argument(
'--data_type', action='store', nargs=1, dest='data_type', required=True
)
# IC or UC
parser.add_argument(
'--feature_type', action='store', nargs=1, dest='feature_type',required=True
)
# processed features path (.npz)
parser.add_argument(
'--feature_dir', action='store', nargs=1, dest='feature_dir', required=True
)
# output(train) model directory
parser.add_argument(
'--output_model_dir', action='store', nargs=1, dest='output_model_dir'
)
# input(test) model path
parser.add_argument(
'--input_model_path', action='store', nargs=1, dest='input_model_path'
)
# output(train) history directory
parser.add_argument(
'--output_hist_dir', action='store', nargs=1, dest='output_hist_dir', required=True
)
parser.add_argument(
'--num_epoch', action='store', nargs=1, dest='num_epoch'
)
parser.add_argument(
'--batch_size', action='store', nargs=1, dest='batch_size'
)
parser.add_argument(
'-v', '--verbose', action='store_true', dest='verbose', default=False
)
args = parser.parse_args()
mode = args.mode[0]
model_type = args.model_type[0]
data_type = args.data_type[0]
feature_type = args.feature_type[0]
feature_dir = args.feature_dir[0]
output_hist_dir = args.output_hist_dir[0]
if mode == 'train':
output_model_dir = args.output_model_dir[0]
if mode == 'test':
input_model_path = args.input_model_path[0]
if args.num_epoch:
num_epoch = int(args.num_epoch[0])
else:
num_epoch = 10
if args.batch_size:
batch_size = int(args.batch_size[0])
else:
batch_size = 256
verbose = args.verbose
if torch.cuda.is_available():
device = 'cuda:0'
else:
device = 'cpu'
# load features
if data_type == '1M':
sparse_feature_path = os.path.join(
feature_dir, f'movie_lens_{data_type}_sparse_features.csv'
)
hist_feature_path = os.path.join(feature_dir, f'movie_lens_{data_type}_IC_UC_features.npz')
else:
# truncate_indices = [0, 1, 5, 7, 10, 11, 12, 13, 14, 19]
truncate_indices = [0, 1, 5]
sparse_feature_path, hist_feature_path = [], []
for i in truncate_indices:
sparse_feature_path.append(
os.path.join(feature_dir, f'movie_lens_{data_type}_sparse_features_{i}.csv')
)
hist_feature_path.append(
os.path.join(feature_dir, f'movie_lens_{data_type}_IC_UC_features_{i}.npz')
)
# data for training DIN
if mode == 'train':
train_input, \
train_label, \
val_input, \
val_label, \
feature_columns, \
behavior_feature_list = process_features_din(
mode, data_type, feature_type, sparse_feature_path, hist_feature_path, feature_type
)
# model
model = DIN(
dnn_feature_columns=feature_columns,
history_feature_list=behavior_feature_list,
pooling_type=model_type,
device=device,
att_weight_normalization=True
)
model.compile(
optimizer='adagrad',
loss='binary_crossentropy',
# metrics=['accuracy'],
metrics=['auc'],
# metrics=['binary_crossentropy'],
)
# verbose 1: progress bar, verbose 2: one line per epoch
history = model.fit(
train_input,
train_label,
batch_size=batch_size,
epochs=num_epoch,
verbose=1,
validation_split=0.0,
validation_data=(val_input, val_label),
shuffle=True,
)
# save trained model
model_path = os.path.join(
output_model_dir,
f'DIN_{model_type}_{feature_type}_{data_type}_{num_epoch}_{batch_size}.pt'
)
if torch.cuda.device_count() > 1:
model_checkpoint = {
'epoch': num_epoch,
'state_dict': model.module.state_dict(),
}
else:
model_checkpoint = {
'epoch': num_epoch,
'state_dict': model.state_dict(),
}
torch.save(model_checkpoint, model_path)
print(f'\nTrained model checkpoint has been saved to {model_path}\n')
# save the history by pandas
history_df = pd.DataFrame(history.history)
hist_csv_path = os.path.join(
output_hist_dir, 'hist_{model_type}_{feature_type}_{data_type}_{num_epoch}_{batch_size}.csv'
)
history_df.to_csv(hist_csv_path)
print(f'\nAssociated model history has been saved to {hist_csv_path}\n')
elif mode == 'test':
test_input, \
test_label, \
feature_columns, \
behavior_feature_list = process_features_din(
mode, data_type, feature_type, sparse_feature_path, hist_feature_path, feature_type
)
# model
model = DIN(
dnn_feature_columns=feature_columns,
history_feature_list=behavior_feature_list,
pooling_type=model_type,
device=device,
att_weight_normalization=True
)
model.compile(
optimizer='adagrad',
loss='binary_crossentropy',
metrics=['auc'],
# metrics=['accuracy'],
# metrics=['binary_crossentropy'],
)
# load trained model
checkpoint = torch.load(input_model_path)
model.load_state_dict(checkpoint['state_dict'])
# run prediction
pred_ans = model.predict(
test_input,
batch_size=batch_size
)
print(
f'\nTest MovieLens{data_type} {feature_type} AUC',
round(roc_auc_score(test_label, pred_ans), 4)
)
else:
raise Exception(f"Unrecognized mode {mode}")