-
Notifications
You must be signed in to change notification settings - Fork 4
/
NFM.py
325 lines (269 loc) · 15.7 KB
/
NFM.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
#!/usr/bin/env python
#coding=utf-8
"""
Implementation of Neural Factorization Machines
"""
import tensorflow as tf
import glob
import os
import json
from datetime import datetime
import shutil
#flag parameters
FLAGS = tf.app.flags.FLAGS
#distributed config
tf.app.flags.DEFINE_boolean("dist_mode", False, "run use distribuion mode or not")
tf.app.flags.DEFINE_string("ps_hosts", '', "Comma-separated list of hostname:port pairs")
tf.app.flags.DEFINE_string("worker_hosts", '', "Comma-separated list of hostname:port pairs")
tf.app.flags.DEFINE_string("job_name", '', "One of 'ps', 'worker'")
tf.app.flags.DEFINE_integer("task_index", 0, "Index of task within the job")
#model parameters
tf.app.flags.DEFINE_string("keep_prob",'0.5,0.6,0.8',"keep probability for bilinear layer and hidden layer(1-dropout_ratio),last one is bilinear layer dropout")
tf.app.flags.DEFINE_string("hidden_layer",'256,64',"size of each hidden layers")
tf.app.flags.DEFINE_float("lr",0.001,"learning rate")
tf.app.flags.DEFINE_float("embedding_size",128,"Size of embedding")
tf.app.flags.DEFINE_float("l2_reg",0.0,"Regularizer for DNN")
tf.app.flags.DEFINE_float("l2_bilinear",0.0,"L2 regularizer for bilinear part")
tf.app.flags.DEFINE_float("l1_bilinear",0.0,"L1 regularizer for bilinear part")
tf.app.flags.DEFINE_string("optimizer","Adagrad","Optimizer type:{Adam,Adagrad,SGD,Momentum}")
tf.app.flags.DEFINE_boolean("batch_norm",True,"use batch_normailization or not")
tf.app.flags.DEFINE_float("batch_norm_decay",0.9,"decay for moving average")
tf.app.flags.DEFINE_string("loss_type","log_loss","loss function (square_loss or log_loss)")
#train config
tf.app.flags.DEFINE_integer("feat_size",0,"Number of features")
tf.app.flags.DEFINE_integer("field_size",0,"Number of fields")
tf.app.flags.DEFINE_integer("batch_size",128,"batch size")
tf.app.flags.DEFINE_integer("num_epochs",10,"Number of epochs")
tf.app.flags.DEFINE_integer("num_threads",0,"Number of reading data threads(suggest equal to the cpu cores)")
tf.app.flags.DEFINE_string("model_root","","Specify model root directory")
tf.app.flags.DEFINE_string("cpkt_dir","","Specify cpkt directory")
tf.app.flags.DEFINE_string("out_dir","","Specify predcition directory")
tf.app.flags.DEFINE_string("train_file","","Specify train data directory")
tf.app.flags.DEFINE_string("test_file","","Specify test data directory")
tf.app.flags.DEFINE_string("va_file","","Specify validation data directory")
tf.app.flags.DEFINE_integer("num_gpus",0,"Number of gpu cores")
tf.app.flags.DEFINE_boolean("clear_existing_model",False,"whether clear existing model file")
tf.app.flags.DEFINE_integer("summary_step",500,"Save summaries every this many steps")
tf.app.flags.DEFINE_integer("log_step_count_steps",500,"The frequency, in number of global steps, that the global step/sec and the loss will be logged during training")
tf.app.flags.DEFINE_integer("save_checkpoints_steps",10000,"Save checkpoints every this many steps")
tf.app.flags.DEFINE_integer("keep_checkpoint_max",0,"The maximum number of recent checkpoint files to keep (default save all the cpkt)")
tf.app.flags.DEFINE_string("task_type", 'train', "task type {train, train_eval, test, eval, export}")
tf.app.flags.DEFINE_integer("eval_step",100,"Positive number of steps for which to evaluate model")
tf.app.flags.DEFINE_integer("start_delay_secs",1000,"Start evaluating after waiting for this many seconds")
tf.app.flags.DEFINE_integer("throttle_secs",1200,"Do not re-evaluate unless the last evaluation was started at least this many seconds ago")
#input libsvm format data
#label feat_id:feat_val
#input file path support single widcard filename and list of files
def input_fn(filenames, batch_size=32, num_epochs=1, perform_shuffle=False):
print("Parsing ", filenames)
def decode_libsvm(line):
columns = tf.string_split([line]," ")
labels = tf.string_to_number(columns.values[0],out_type=tf.float32)
feat = tf.string_split(columns.values[1:],":")
feat = tf.reshape(feat.values,feat.dense_shape)
feat_id,feat_val = tf.split(feat,num_or_size_splits=2,axis=1)
feat_id = tf.string_to_number(feat_id,out_type=tf.int32)
feat_val = tf.string_to_number(feat_id,out_type=tf.float32)
return {"feat_id":feat_id,"feat_val":feat_val},labels
if type(filenames) != type([]):
filenames = glob.glob(filenames)
dataset = tf.data.TextLineDataset(filenames, buffer_size=FLAGS.buffer_size_gb*1024**3)
#dataset = dataset.shard()
dataset = dataset.repeat(num_epochs)
dataset = dataset.apply(tf.contrib.data.map_and_batch(decode_libsvm,batch_size,num_parallel_calls=FLAGS.num_threads))
dataset = dataset.prefetch(batch_size)
iterator = dataset.make_one_shot_iterator()
batch_features,batch_labels = iterator.get_next()
return batch_features,batch_labels
def nfm_fn(features,labels,mode,params):
embd_size = params["embedding_size"]
feat_size = params["feat_size"]
field_size = params["field_size"]
feat_id = features["feat_id"]
feat_id = tf.reshape(feat_id,[-1,field_size]) #None*field_size
feat_val = features["feat_val"]
feat_val = tf.reshape(feat_val,[-1,field_size,1]) #None*field_size
keep_prob = list(map(float,params["keep_prob"].split(",")))
assert len(keep_prob) >= 1
hidden_layer = list(map(int,params["hidden_layer"].split(",")))
assert len(keep_prob) == len(hidden_layer)+1
learning_rate = params["learning_rate"]
l2_reg= params["l2_reg"]
l2_bilinear = params["l2_bilinear"]
l1_bilinear = params["l1_bilinear"]
l1_l2_reg = tf.contrib.layers.l1_l2_regularizer(scale_l1=l1_bilinear,scale_l2=l2_bilinear)
#featur embedding
feat_embedding = tf.get_variable("feature_embedding",shape=[feat_size,embd_size],initializer=tf.glorot_normal_initializer(),regularizer=l1_l2_reg)
feat_bias = tf.get_variable("feature_bias",shape=[feat_size,1],initializer=tf.glorot_normal_initializer(),regularizer=l1_l2_reg)
bias = tf.get_variable("bias",shape=[1],initializer=tf.constant_initializer(0.0))
if mode == tf.estimator.ModeKeys.TRAIN:
train_phase = True
else:
train_phase = False
#bilinear part
with tf.variable_scope("bilinear"):
emb = tf.nn.embedding_lookup(feat_embedding,feat_id) #None*field_size*embd_size
sum_emb = tf.reduce_sum(tf.multiply(emb,feat_val),1) #None*embd_size
sum_emb_square = tf.square(sum_emb)
square_emb = tf.square(tf.multiply(emb,feat_val))
square_emb_sum = tf.reduce_sum(square_emb,1) #None*embd_size
bilinear = 0.5*tf.subtract(sum_emb_square,square_emb_sum) #None*embd_size
if params["batch_norm"]:
bilinear = batch_norm_layer(bilinear,train_phase,"bilinear_bn")
if train_phase:
bilinear = tf.nn.dropout(bilinear,keep_prob[-1])
with tf.variable_scope("deep"):
deep_out = bilinear
for i in range(len(hidden_layer)):
deep_out = tf.contrib.layers.fully_connected(inputs=deep_out, num_outputs=hidden_layer[i],weights_regularizer=tf.contrib.layers.l2_regularizer(l2_reg), scope='mlp_%d' % i)
if params["batch_norm"]:
deep_out = batch_norm_layer(deep_out,train_phase,scope_bn = "bn_%d"%i)
if train_phase:
deep_out = tf.nn.dropout(deep_out,keep_prob[i])
deep_out = tf.contrib.layers.fully_connected(inputs=deep_out, num_outputs = 1, activation_fn=tf.identity,
weights_regularizer = tf.contrib.layers.l2_regularizer(l2_reg), scope="prediction")
deep_out = tf.reshape(deep_out,shape=[-1]) #None*1
with tf.variable_scope("linear"):
wgts = tf.nn.embedding_lookup(feat_bias,feat_id) #None*field_size*1
linear_out = tf.reduce_sum(tf.multiply(wgts,feat_val),1) #None*1
with tf.variable_scope("nfm_out"):
bias = bias * tf.ones_like(linear_out, dtype=tf.float32)
logits = tf.add_n([linear_out,bias,deep_out])
pred = tf.sigmoid(logits)
predicted_labels = tf.cast(logits + 0.5, tf.int32)
predictions={"prob": pred}
eval_metric_ops = {"auc":tf.metrics.auc(labels, pred),
"auc_precision_recall":tf.metrics.auc(labels, pred, curve='PR'),
"accuracy":tf.metrics.accuracy(labels, predicted_labels),
"recall":tf.metrics.recall(labels, predicted_labels),
"precision":tf.metrics.precision(labels, predicted_labels),
"label/mean":tf.metrics.mean(labels)
}
export_outputs = {tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: tf.estimator.export.PredictOutput(predictions)}
if FLAGS.loss_type == "square_loss":
obj_loss = tf.reduce_mean(tf.nn.l2_loss(tf.subtract(labels,logits)))
if FLAGS.loss_type == "log_loss":
obj_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=labels))
loss = obj_loss + tf.losses.get_regularization_loss()
if FLAGS.optimizer == "Adam":
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
elif FLAGS.optimizer == "Adagrad":
optimizer = tf.train.AdagradOptimizer(learning_rate=learning_rate)
elif FLAGS.optimizer == "Momentum":
optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=0.95)
elif FLAGS.optimizer == "SGD":
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())
if mode == tf.estimator.ModeKeys.TRAIN:
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=loss,
train_op=train_op)
if mode == tf.estimator.ModeKeys.EVAL:
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=loss,
eval_metric_ops=eval_metric_ops)
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
export_outputs=export_outputs)
def batch_norm_layer(x, train_phase, scope_bn):
bn_train = tf.contrib.layers.batch_norm(x, decay=FLAGS.batch_norm_decay, center=True, scale=True, updates_collections=None, is_training=True, reuse=None, scope=scope_bn)
bn_infer = tf.contrib.layers.batch_norm(x, decay=FLAGS.batch_norm_decay, center=True, scale=True, updates_collections=None, is_training=False, reuse=True, scope=scope_bn)
z = tf.cond(tf.cast(train_phase, tf.bool), lambda: bn_train, lambda: bn_infer)
return z
def set_dist_env():
if FLAGS.dist_mode:
ps_hosts = FLAGS.ps_hosts.split(",")
worker_hosts = FLAGS.worker_hosts.split(",")
chief_hosts = worker_hosts[0:1] #first worker is chief
worker_hosts = worker_hosts[1:] #rest workers are worker
task_index = FLAGS.task_index
job_name = FLAGS.job_name
if job_name == "worker" and task_index == 0:
job_name = "chief"
if job_name == "worker" and task_index > 0:
task_index -= 1
tf_config = {'cluster': {'chief': chief_hosts, 'worker': worker_hosts, 'ps': ps_hosts}, 'task': {'type': job_name, 'index': task_index }}
print(json.dumps(tf_config))
os.environ["TF_CONFIG"]=json.dumps(tf_config)
def main(_):
model_dir = FLAGS.model_root
model_dir += FLAGS.cpkt_dir if FLAGS.cpkt_dir != "" else datetime.now().strftime("%Y-%m-%d-%H")
tr_file = FLAGS.train_file
va_file = FLAGS.va_file
te_file = FLAGS.test_file
print("task type: ",FLAGS.task_type)
print("model dir: ", model_dir)
print("train file: ",tr_file)
print("validation file: ",va_file)
print("test file: ",te_file)
print("embedding_size: ",FLAGS.embedding_size)
print('num epochs: ', FLAGS.num_epochs)
print("learning rate: ",FLAGS.lr)
print("batch normalization: ", FLAGS.batch_norm)
print("batch normalization decay: ", FLAGS.batch_norm_decay)
print("hidden layers: ",FLAGS.hidden_layer)
print("keep probability: ",FLAGS.keep_prob)
print("optimizer: ",FLAGS.optimizer)
print("l2 reg: ",FLAGS.l2_reg)
print("loss type: ",FLAGS.loss_type)
print("l2_bilinear: ",FLAGS.l2_bilinear)
print("l1_bilinear: ",FLAGS.l1_bilinear)
#remove if system forbide
if FLAGS.clear_existing_model:
try:
shutil.rmtree(model_dir)
except Exception as e:
print(e, "at clear_existing_model")
else:
print("existing model cleaned at %s" % model_dir)
set_dist_env()
model_params = {
"field_size": FLAGS.field_size,
"feat_size": FLAGS.feat_size,
"embedding_size": FLAGS.embedding_size,
"learning_rate": FLAGS.lr,
"batch_norm": FLAGS.batch_norm,
"l2_reg": FLAGS.l2_reg,
"hidden_layer": FLAGS.hidden_layer,
"keep_prob": FLAGS.keep_prob,
"l2_bilinear":FLAGS.l2_bilinear,
"l1_bilinear":FLAGS.l1_bilinear
}
if FLAGS.num_gpus != 0:
session_config = tf.ConfigProto(per_process_gpu_memory_fraction=1.0, allow_growth=True,allow_soft_placement = True,log_device_placement=False,
intra_op_parallelism_threads=0, inter_op_parallelism_threads=0,device_count={'GPU': FLAGS.num_gpus,"CPU":FLAGS.num_threads})
else:
session_config = tf.ConfigProto(device_count={'GPU':0, 'CPU':FLAGS.num_threads})
config = tf.estimator.RunConfig().replace(save_summary_steps=FLAGS.summary_step,save_checkpoints_steps=FLAGS.save_checkpoints_steps,
session_config= session_config,keep_checkpoint_max=FLAGS.keep_checkpoint_max,log_step_count_steps=FLAGS.log_step_count_steps)
NFM = tf.estimator.Estimator(model_fn=nfm_fn, model_dir=model_dir, params=model_params, config=config)
if FLAGS.task_type == 'train':
NFM.train(input_fn=lambda: input_fn(tr_file, num_epochs=FLAGS.num_epochs, batch_size=FLAGS.batch_size))
elif FLAGS.task_type == "train_eval":
train_spec = tf.estimator.TrainSpec(input_fn=lambda: input_fn(tr_file, num_epochs=FLAGS.num_epochs, batch_size=FLAGS.batch_size))
eval_spec = tf.estimator.EvalSpec(input_fn=lambda: input_fn(te_file, num_epochs=1, batch_size=FLAGS.batch_size),
steps=FLAGS.eval_step,start_delay_secs=FLAGS.start_delay_secs,throttle_secs=FLAGS.throttle_secs)
tf.estimator.train_and_evaluate(NFM,train_spec,eval_spec)
elif FLAGS.task_type == "test":
NFM.evaluate(input_fn=lambda: input_fn(te_file, num_epochs=1, batch_size=FLAGS.batch_size))
elif FLAGS.task_type == "eval":
preds = NFM.predict(input_fn=lambda: input_fn(te_file, num_epochs=1, batch_size=FLAGS.batch_size), predict_keys="prob")
with open(FLAGS.out_dir+"/pred.txt", "w") as fo:
for prob in preds:
fo.write("%f\n" % (prob['prob']))
elif FLAGS.task_type == 'export':
feature_spec = {
'feat_id': tf.placeholder(dtype=tf.int64, shape=[None, FLAGS.field_size], name='feat_id'),
'feat_val': tf.placeholder(dtype=tf.float32, shape=[None, FLAGS.field_size], name='feat_val')
}
serving_input_receiver_fn = tf.estimator.export.build_raw_serving_input_receiver_fn(feature_spec)
NFM.export_savedmodel(FLAGS.out_dir, serving_input_receiver_fn)
if __name__ == "__main__":
tf.logging.set_verbosity(tf.logging.INFO)
tf.app.run()