-
Notifications
You must be signed in to change notification settings - Fork 0
/
SeqDCNmodel.py
161 lines (151 loc) · 8.11 KB
/
SeqDCNmodel.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
#!/bin/python
#coding:utf-8
import tensorflow as tf
import numpy as np
from Word2Vec import Word2Vec
from DeepModel import DeepModel
# from tensorflow.python.tools import freeze_graph
import pyarrow.hdfs as hdfs
import os,re
class ConcatModel(DeepModel):
def __init__(self, file_name, init_lr=0.001, batch_size=256, process='train', rm_dir=False, use_checkpoint=False):
super().__init__(file_name, init_lr=init_lr, batch_size=batch_size,
process=process, rm_dir=rm_dir, use_checkpoint=use_checkpoint, use_normal=True)
self.e_arr = 2*np.ones(25, dtype=np.int)
self.e_arr[:3] = [300, 7, 24*12]
self.e_sum = sum(self.e_arr)
self.e_add_arr = [sum(self.e_arr[:i]) for i in range(len(self.e_arr))]
self.normal_means = []
self.normal_stds = []
self.normal_file_path = './normal_params.txt'
def load_normal_params(self):
f = open(self.normal_file_path, 'r')
for line in f.readlines():
if line[:6] == "column":
continue
else:
lines = line.strip().split(",")
if len(lines) < 6:
continue
else:
self.normal_means.append(float(lines[4]))
self.normal_stds.append(float(lines[5]))
f.close()
self.normal_means = self.normal_means[:36]
self.normal_stds = self.normal_stds[:36]
for x in range(36):
if self.normal_stds[x] > -0.001 and self.normal_stds[x] < 0.001:
self.normal_stds[x] = 1.0
else:
continue
def dense_layer(self, x, units=128, name='dense_layer'):
with tf.variable_scope(name):
initializer = tf.contrib.layers.variance_scaling_initializer()
kernel = tf.get_variable('kernel', shape=[x.shape[1], units], regularizer=self.regular(), initializer=initializer)
bias = tf.get_variable('bias', shape=[units,], initializer=tf.zeros_initializer())
output = tf.nn.relu(x @ kernel + bias)
return output
def deep_network(self, x, mode, name='deep_network'):
ulist = [512,256,128]
with tf.variable_scope(name):
for i in range(len(ulist)):
x = self.dense_layer(x, units = ulist[i],name="dense_layer_%s"%i)
if i == 0:
x = tf.layers.batch_normalization(x, training= mode == tf.estimator.ModeKeys.TRAIN)
return x
def DNN(self, x, mode, name='dnn'):
with tf.variable_scope(name):
cross_output = self.cross_network(x, mode=mode)
#Dense layers
deep_output = self.deep_network(x, mode=mode)
output = tf.concat([deep_output, cross_output], 1)
ulist = [256, 128]
for i in range(2):
output = self.dense_layer(output, units = ulist[i], name="dense_layer_%s"%i)
return output
def Word2Vec(self, x, mode, name='w2v'):
with tf.variable_scope(name):
x = tf.cast(x, tf.int32)
x = x + tf.constant(self.e_add_arr, dtype=tf.int32)
initializer = tf.truncated_normal_initializer(stddev=1.0/(self.e_sum ** 0.5))
kernel = tf.get_variable('kernel', (self.e_sum, 16), initializer=initializer)
output = tf.nn.embedding_lookup(kernel, x)
output = tf.reshape(output, (-1, 25 * 16))
ulist = [512, 256, 128]
for i in range(3):
output = self.dense_layer(output, units = ulist[i], name="dense_layer_%s"%i)
return output
def output_box(self, vec1, vec2, mode, name='output_box1'):
with tf.variable_scope(name):
vec = tf.concat([vec1, vec2], axis=1)
dense_vec = self.dense_layer(vec, 128, name='dense_1')
cross_vec = self.normal_cross_layer(vec1, vec2)
output = tf.concat([dense_vec, cross_vec], axis=1)
output = self.dense_layer(output, 128, name='dense_2')
return output
def bottle_neck(self, x, name='bottle_neck'):
with tf.variable_scope(name):
initializer = tf.glorot_uniform_initializer() # sigmoid activate func use Xvaier Initializer
kernel = tf.get_variable('kernel', shape=[x.shape[1],1], initializer=initializer)
bias = tf.get_variable('bias', shape=[1,], initializer=tf.zeros_initializer())
output = x @ kernel + bias
return output
def output_layers(self, dnn_vec, w2v_vec, mode, name="output_layers"):
with tf.variable_scope(name):
tmp_vec = dnn_vec
for i in range(6):
tmp_vec = self.output_box(tmp_vec, w2v_vec, mode=mode, name="output_box_%s"%i)
output = self.bottle_neck(tmp_vec, name="bottle_neck_%s"%i)
if i == 0:
bucket_vec = output
else:
bucket_vec = tf.concat([bucket_vec, output], axis=1)
return bucket_vec
def model_fn(self, features, labels, mode, params):
features, label = (features.get('features'), features.get('label'))
basic_feat, embed_feat, wait_time = tf.split(features, [36, 25, 6], axis=1)
basic_feat = (basic_feat - self.normal_means)/self.normal_stds/3
basic_feat = tf.clip_by_value(basic_feat, -1, 1)
dnn_vec = self.DNN(basic_feat, mode=mode)
w2v_vec = self.Word2Vec(embed_feat, mode=mode)
multi_output = self.output_layers(dnn_vec, w2v_vec, mode=mode)
#predicts = wait_time * multi_output @ tf.ones([6, 1])
product_id, _ = tf.split(basic_feat, [1, 35], axis=1)
predicts = tf.concat([wait_time * multi_output, product_id], axis=1)
predicts = self.dense_layer(predicts, 1, 'out_dense')
predicts = tf.nn.sigmoid(predicts)
train_op = metrics = loss = None
if mode == tf.estimator.ModeKeys.TRAIN or mode == tf.estimator.ModeKeys.EVAL:
with tf.variable_scope('loss'):
reg = tf.cast(tf.losses.get_regularization_loss(), tf.float32)
loss = tf.reduce_mean(tf.nn.weighted_cross_entropy_with_logits(label, predicts, self.params['loss_weight']))
if mode == tf.estimator.ModeKeys.TRAIN:
if self.use_ckp:
self.printLog('load checkpoint model from %s' % (params['checkpoint_path']))
tf.train.init_from_checkpoint(params['checkpoint_path'], {'deep_net/': 'deep_net/'})
with tf.variable_scope('optimizer'):
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
global_step = tf.train.get_or_create_global_step()
warmup_done = tf.cast(global_step, tf.float32) / (params['lr_warmup_msteps'] * 1e6)
warmup_lr = params['init_learning_rate'] * tf.minimum(warmup_done, 1.0)
is_warmup = tf.cast(global_step, tf.float32) < (params['lr_warmup_msteps'] * 1e6)
decay_lr = tf.train.polynomial_decay(
params['init_learning_rate'], global_step,
int(params['lr_decay_msteps'] * 1e6), params['end_learning_rate'])
learning_rate = tf.where(is_warmup, warmup_lr, decay_lr)
decay_rate = learning_rate * params['weight_decay_rate']
tf.summary.scalar('learning_rate',learning_rate)
optimizer = tf.contrib.opt.AdamGSOptimizer(learning_rate=learning_rate)
grads, variables = zip(*optimizer.compute_gradients(loss))
grads, global_norm = tf.clip_by_global_norm(grads, 5)
decay_vars = [v for v in variables if 'kernel:' in v.name]
train_op = optimizer.apply_gradients(zip(grads, variables), global_step)
train_op = tf.group([train_op, update_ops])
if mode == tf.estimator.ModeKeys.EVAL:
metrics = {
'auc': tf.metrics.auc(label, predicts),
'precision': tf.metrics.precision(label, tf.round(predicts)),
'recall': tf.metrics.recall(label, tf.round(predicts)),
'f1_score': tf.contrib.metrics.f1_score(label, tf.round(predicts))
}
return tf.estimator.EstimatorSpec(mode, predicts, loss, train_op, metrics)