-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathmain.py
275 lines (241 loc) · 11.6 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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
# -*- coding: utf-8 -*-
import os
import gc
import time
import numpy as np
from collections import defaultdict
from keras import backend as K
from keras import optimizers
from utils import load_data, pickle_load, format_filename, write_log
from models import KGCN
from config import ModelConfig, PROCESSED_DATA_DIR, USER_VOCAB_TEMPLATE, ENTITY_VOCAB_TEMPLATE, \
RELATION_VOCAB_TEMPLATE, ADJ_ENTITY_TEMPLATE, ADJ_RELATION_TEMPLATE, LOG_DIR, PERFORMANCE_LOG, \
ITEM_VOCAB_TEMPLATE
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
def get_optimizer(op_type, learning_rate):
if op_type == 'sgd':
return optimizers.SGD(learning_rate)
elif op_type == 'rmsprop':
return optimizers.RMSprop(learning_rate)
elif op_type == 'adagrad':
return optimizers.Adagrad(learning_rate)
elif op_type == 'adadelta':
return optimizers.Adadelta(learning_rate)
elif op_type == 'adam':
return optimizers.Adam(learning_rate, clipnorm=5)
else:
raise ValueError('Optimizer Not Understood: {}'.format(op_type))
def train(dataset, neighbor_sample_size, embed_dim, n_depth, l2_weight, lr, optimizer_type,
batch_size, aggregator_type, n_epoch, callbacks_to_add=None, overwrite=False):
config = ModelConfig()
config.neighbor_sample_size = neighbor_sample_size
config.embed_dim = embed_dim
config.n_depth = n_depth
config.l2_weight = l2_weight
config.lr = lr
config.optimizer = get_optimizer(optimizer_type, lr)
config.batch_size = batch_size
config.aggregator_type = aggregator_type
config.n_epoch = n_epoch
config.callbacks_to_add = callbacks_to_add
config.user_vocab_size = len(pickle_load(format_filename(PROCESSED_DATA_DIR,
USER_VOCAB_TEMPLATE,
dataset=dataset)))
config.item_vocab_size = len(pickle_load(format_filename(PROCESSED_DATA_DIR,
ITEM_VOCAB_TEMPLATE,
dataset=dataset)))
config.entity_vocab_size = len(pickle_load(format_filename(PROCESSED_DATA_DIR,
ENTITY_VOCAB_TEMPLATE,
dataset=dataset)))
config.relation_vocab_size = len(pickle_load(format_filename(PROCESSED_DATA_DIR,
RELATION_VOCAB_TEMPLATE,
dataset=dataset)))
config.adj_entity = np.load(format_filename(PROCESSED_DATA_DIR, ADJ_ENTITY_TEMPLATE,
dataset=dataset))
config.adj_relation = np.load(format_filename(PROCESSED_DATA_DIR, ADJ_RELATION_TEMPLATE,
dataset=dataset))
config.exp_name = f'kgcn_{dataset}_neigh_{neighbor_sample_size}_embed_{embed_dim}_depth_' \
f'{n_depth}_agg_{aggregator_type}_optimizer_{optimizer_type}_lr_{lr}_' \
f'batch_size_{batch_size}_epoch_{n_epoch}'
callback_str = '_' + '_'.join(config.callbacks_to_add)
callback_str = callback_str.replace('_modelcheckpoint', '').replace('_earlystopping', '')
config.exp_name += callback_str
# logger to log output of training process
train_log = {'exp_name': config.exp_name, 'batch_size': batch_size, 'optimizer': optimizer_type,
'epoch': n_epoch, 'learning_rate': lr}
print('Logging Info - Experiment: %s' % config.exp_name)
model_save_path = os.path.join(config.checkpoint_dir, '{}.hdf5'.format(config.exp_name))
model = KGCN(config)
train_data = load_data(dataset, 'train')
valid_data = load_data(dataset, 'dev')
test_data = load_data(dataset, 'test')
if not os.path.exists(model_save_path) or overwrite:
start_time = time.time()
model.fit(x_train=[train_data[:, :1], train_data[:, 1:2]], y_train=train_data[:, 2:3],
x_valid=[valid_data[:, :1], valid_data[:, 1:2]], y_valid=valid_data[:, 2:3])
elapsed_time = time.time() - start_time
print('Logging Info - Training time: %s' % time.strftime("%H:%M:%S",
time.gmtime(elapsed_time)))
train_log['train_time'] = time.strftime("%H:%M:%S", time.gmtime(elapsed_time))
print('Logging Info - Evaluate over valid data:')
model.load_best_model()
auc, acc, f1 = model.score(x=[valid_data[:, :1], valid_data[:, 1:2]], y=valid_data[:, 2:3])
user_list, train_record, valid_record, item_set, k_list = topk_settings(train_data,
valid_data,
config.item_vocab_size)
topk_p, topk_r = topk_eval(model, user_list, train_record, valid_record, item_set, k_list)
print(f'Logging Info - dev_auc: {auc}, dev_acc: {acc}, dev_f1: {f1}, dev_topk_p: {topk_p}, '
f'dev_topk_r: {topk_r}')
train_log['dev_auc'] = auc
train_log['dev_acc'] = acc
train_log['dev_f1'] = f1
train_log['dev_topk_p'] = topk_p
train_log['dev_topk_r'] = topk_r
if 'swa' in config.callbacks_to_add:
model.load_swa_model()
print('Logging Info - Evaluate over valid data based on swa model:')
auc, acc, f1 = model.score(x=[valid_data[:, :1], valid_data[:, 1:2]], y=valid_data[:, 2:3])
topk_p, topk_r = topk_eval(model, user_list, train_record, valid_record, item_set, k_list)
train_log['swa_dev_auc'] = auc
train_log['swa_dev_acc'] = acc
train_log['swa_dev_f1'] = f1
train_log['swa_dev_topk_p'] = topk_p
train_log['swa_dev_topk_r'] = topk_r
print(f'Logging Info - swa_dev_auc: {auc}, swa_dev_acc: {acc}, swa_dev_f1: {f1}, '
f'swa_dev_topk_p: {topk_p}, swa_dev_topk_r: {topk_r}')
print('Logging Info - Evaluate over test data:')
model.load_best_model()
auc, acc, f1 = model.score(x=[test_data[:, :1], test_data[:, 1:2]], y=test_data[:, 2:3])
user_list, train_record, test_record, item_set, k_list = topk_settings(train_data,
test_data,
config.item_vocab_size)
topk_p, topk_r = topk_eval(model, user_list, train_record, test_record, item_set, k_list)
train_log['test_auc'] = auc
train_log['test_acc'] = acc
train_log['test_f1'] = f1
train_log['test_topk_p'] = topk_p
train_log['test_topk_r'] = topk_r
print(f'Logging Info - test_auc: {auc}, test_acc: {acc}, test_f1: {f1}, test_topk_p: {topk_p}, '
f'test_topk_r: {topk_r}')
if 'swa' in config.callbacks_to_add:
model.load_swa_model()
print('Logging Info - Evaluate over test data based on swa model:')
auc, acc, f1 = model.score(x=[test_data[:, :1], test_data[:, 1:2]], y=test_data[:, 2:3])
topk_p, topk_r = topk_eval(model, user_list, train_record, test_record, item_set, k_list)
train_log['swa_test_auc'] = auc
train_log['swa_test_acc'] = acc
train_log['swa_test_f1'] = f1
train_log['swa_test_topk_p'] = topk_p
train_log['swa_test_topk_r'] = topk_r
print(f'Logging Info - swa_test_auc: {auc}, swa_test_acc: {acc}, swa_test_f1: {f1}, '
f'swa_test_topk_p: {topk_p}, swa_test_topk_r: {topk_r}')
train_log['timestamp'] = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime())
write_log(format_filename(LOG_DIR, PERFORMANCE_LOG), log=train_log, mode='a')
del model
gc.collect()
K.clear_session()
def topk_settings(train_data, test_data, n_item, user_num=100):
k_list = [1, 2, 5, 10, 20, 50, 100]
train_record = get_user_record(train_data, True)
test_record = get_user_record(test_data, False)
user_list = list(set(train_record.keys()) & set(test_record.keys()))
if len(user_list) > user_num:
user_list = np.random.choice(user_list, size=user_num, replace=False)
item_set = set(list(range(n_item)))
return user_list, train_record, test_record, item_set, k_list
def get_user_record(data, is_train):
user_history_dict = defaultdict(set)
for interaction in data:
user = interaction[0]
item = interaction[1]
label = interaction[2]
if is_train or label == 1:
user_history_dict[user].add(item)
return user_history_dict
def topk_eval(model, user_list, train_record, test_record, item_set, k_list):
precision_list = {k: [] for k in k_list}
recall_list = {k: [] for k in k_list}
for user in user_list:
test_item_list = list(item_set - train_record[user])
item_score_map = dict()
input_user = np.expand_dims(np.array([user] * len(test_item_list)), axis=1)
input_item = np.expand_dims(np.array(test_item_list), axis=1)
item_scores = model.predict([input_user, input_item])
for item, score in zip(test_item_list, item_scores):
item_score_map[item] = score
item_score_pair_sorted = sorted(item_score_map.items(), key=lambda x: x[1], reverse=True)
item_sorted = [i[0] for i in item_score_pair_sorted]
for k in k_list:
hit_num = len(set(item_sorted[:k]) & test_record[user])
precision_list[k].append(hit_num / k)
recall_list[k].append(hit_num / len(test_record[user]))
precision = [np.mean(precision_list[k]) for k in k_list]
recall = [np.mean(recall_list[k]) for k in k_list]
return precision, recall
if __name__ == '__main__':
train(dataset='movie',
neighbor_sample_size=4,
embed_dim=32,
n_depth=2,
l2_weight=1e-7,
lr=2e-2,
optimizer_type='adam',
batch_size=65536,
aggregator_type='sum',
n_epoch=50,
callbacks_to_add=['modelcheckpoint', 'earlystopping', 'swa'])
train(dataset='movie',
neighbor_sample_size=4,
embed_dim=32,
n_depth=2,
l2_weight=1e-7,
lr=2e-2,
optimizer_type='adam',
batch_size=65536,
aggregator_type='concat',
n_epoch=50,
callbacks_to_add=['modelcheckpoint', 'earlystopping', 'swa'])
train(dataset='movie',
neighbor_sample_size=4,
embed_dim=32,
n_depth=2,
l2_weight=1e-7,
lr=2e-2,
optimizer_type='adam',
batch_size=65536,
aggregator_type='neigh',
n_epoch=50,
callbacks_to_add=['modelcheckpoint', 'earlystopping', 'swa'])
train(dataset='music',
neighbor_sample_size=8,
embed_dim=16,
n_depth=1,
l2_weight=1e-4,
lr=5e-4,
optimizer_type='adam',
batch_size=128,
aggregator_type='sum',
n_epoch=50,
callbacks_to_add=['modelcheckpoint', 'earlystopping', 'swa'])
train(dataset='music',
neighbor_sample_size=8,
embed_dim=16,
n_depth=1,
l2_weight=1e-4,
lr=5e-4,
optimizer_type='adam',
batch_size=128,
aggregator_type='concat',
n_epoch=50,
callbacks_to_add=['modelcheckpoint', 'earlystopping', 'swa'])
train(dataset='music',
neighbor_sample_size=8,
embed_dim=16,
n_depth=1,
l2_weight=1e-4,
lr=5e-4,
optimizer_type='adam',
batch_size=128,
aggregator_type='neigh',
n_epoch=50,
callbacks_to_add=['modelcheckpoint', 'earlystopping', 'swa'])