-
Notifications
You must be signed in to change notification settings - Fork 0
/
opt_spark.py
193 lines (157 loc) · 6.82 KB
/
opt_spark.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
# -*- coding: utf-8 -*-
# Copyright (C) 2018, Tencent Inc.
# Author: bintan
from __future__ import absolute_import, division, print_function
import time
import numpy as np
import tensorflow as tf
from io import BytesIO
from numpy.lib import format
from operator import add
from pyspark.sql import SparkSession
def generator_ij(iterator):
buf = np.zeros(shape=(0, 3)).astype('int64')
for line in iterator:
biparite_line = line.split('\t')
orders = biparite_line[4].split(',')
tmp = np.zeros(shape=(len(orders), 3)).astype('int64')
# j,i,k
tmp[:, 0] = orders
tmp[:, 1] = biparite_line[1]
tmp[:, 2] = biparite_line[0]
buf = np.row_stack((buf, tmp))
yield buf
def init_sp(iterator):
tf.enable_eager_execution()
j_nums = bc_j.value
for e in iterator:
# print("e is {}".format(e))
j, i, k = 0, 1, 2
# es = e.shape[0]
mask_j = np.zeros(j_nums)
mask_id = np.random.randint(j_nums, size=500)
for x in mask_id:
mask_j[x] = 1
dense_shape = [j_nums, e[-1][i], e[-1][k]]
indices = tf.constant(e - [1, 1, 1], dtype=tf.int64)
indices = tf.boolean_mask(indices, tf.gather(mask_j, indices[:, j]))
sp_ji = tf.sparse.SparseTensor(indices[:, :2], tf.fill([indices.shape[0]], 1.0), dense_shape[:2])
sub_sj = tf.sparse.reduce_sum(sp_ji, axis=i).numpy()
yield [indices, dense_shape, sub_sj]
def get_tbl_key(id_j, id_u):
key_high = tf.bitwise.left_shift(id_j, 32)
return tf.bitwise.bitwise_or(key_high, id_u)
def opt_lambda(sum_i, f_j, lambda_0, lambda_1, lambda_):
lambda_0 = tf.where(sum_i < f_j, lambda_0, lambda_)
lambda_1 = tf.where(sum_i < f_j, lambda_, lambda_1)
lambda_ = (lambda_0 + lambda_1) / 2
return lambda_0, lambda_1, lambda_
def init_loop(iterator):
tf.enable_eager_execution()
theta_j = bc_theta.value
freq_j = bc_freq.value
max_beta_ = bc_max_beta.value
indices, dense_shape = 0, 1
j, i, u = 0, 1, 2
for x in iterator:
gather_theta = tf.gather(theta_j, x[indices][:, j])
xij_clip = tf.clip_by_value(gather_theta, 0, 1)
xij = tf.sparse.SparseTensor(indices=x[indices], values=xij_clip, dense_shape=x[dense_shape])
sum_iu = tf.sparse.reduce_sum(xij, axis=[i, u]).numpy()
sum_i = tf.sparse.reduce_sum_sparse(xij, axis=i)
tbl_key = get_tbl_key(sum_i.indices[:, 0], sum_i.indices[:, 1]) # id_j, id_u
tbl_init = tf.contrib.lookup.KeyValueTensorInitializer(keys=tbl_key, values=tf.range(0, tbl_key.shape[0]))
tbl = tf.contrib.lookup.HashTable(initializer=tbl_init, default_value=-1)
xij_key = get_tbl_key(xij.indices[:, j], xij.indices[:, u])
lambda_idx = tbl.lookup(xij_key)
lambda_ = tf.zeros(sum_i.values.shape)
lambda_0 = tf.zeros(lambda_.shape)
lambda_1 = tf.fill(lambda_.shape, -max_beta_)
gather_freq = tf.gather(freq_j, sum_i.indices[:, j]) # sum_i--->id_j
lambda_0, lambda_1, lambda_ = opt_lambda(sum_i.values, gather_freq, lambda_0, lambda_1, lambda_)
b = tf.zeros(x[dense_shape][i])
b0 = tf.zeros(b.shape)
b1 = tf.fill(b.shape, -max_beta_)
sum_ju = tf.sparse.reduce_sum(xij, axis=[j, u])
b0, b1, b = opt_lambda(sum_ju, 1, b0, b1, b)
yield [x[indices], x[dense_shape], lambda_idx, gather_theta, gather_freq, b0, b1, b, lambda_0, lambda_1,
lambda_, 0, sum_iu]
def main_loop(iterator):
tf.enable_eager_execution()
alpha_j = bc_alpha.value + 1
indices, dense_shape, lambda_idx, theta_j, freq_j, b0, b1, b, lambda_0, lambda_1, lambda_ = [i for i in range(11)]
j, i, u = 0, 1, 2
for x in iterator:
spv_alpha = tf.gather(alpha_j, x[indices][:, j])
spv_beta = tf.gather(x[b], x[indices][:, i])
spv_lambda = tf.gather(x[lambda_], lambda_idx)
sum_dual = spv_alpha + spv_beta + spv_lambda
xij_clip = tf.clip_by_value(sum_dual * x[theta_j], 0, 1)
xij = tf.SparseTensor(x[indices], xij_clip, x[dense_shape])
sum_iu = tf.sparse.reduce_sum(xij, axis=[i, u]).numpy() # 44.2%
sum_i = tf.sparse.reduce_sum_sparse(xij, axis=i) # 45.2%
x[lambda_0], x[lambda_1], x[lambda_] = opt_lambda(sum_i.values, x[freq_j], x[lambda_0], x[lambda_1], x[lambda_])
sum_ju = tf.sparse.reduce_sum(xij, axis=[j, u]) # 10.6%
x[b0], x[b1], x[b] = opt_lambda(sum_ju, 1, x[b0], x[b1], x[b])
x[-2] = xij.values
x[-1] = sum_iu
yield x
def opt_alpha(s_j, a0, a1, a):
a0 = np.where(s_j > d, a0, a)
a1 = np.where(s_j > d, a, a1)
a = (a0 + a1) / 2
return a0, a1, a
def flat(indices, xij):
for x, y in zip(indices[:, :2].numpy(), xij.numpy()):
yield '\t'.join(['\t'.join('%d' % i for i in x), str(y)])
if __name__ == "__main__":
import sys
spark = SparkSession.builder.getOrCreate()
sc = spark.sparkContext
biparitePath = sys.argv[0]
ordersPath = sys.argv[1]
partitions = int(sys.argv[2])
print('biparitePath:' + biparitePath)
print('ordersPath:' + ordersPath)
orderDF = spark.read.json(ordersPath).filter("order_play_type = 'out'") \
.select("vid", "should_play", "freq_times") \
.collect()
orderDF = pd.DataFrame(orderDF, columns=["vid", "should_play", "freq_times"])
orderDF['should_play'].fillna(0, inplace=True)
d = (orderDF['should_play'] * 1000).astype(np.float32)
f = orderDF['freq_times'].values.astype(np.float32)
j = d.shape[0]
bc_j = sc.broadcast(j)
bc_freq = sc.broadcast(f)
init = sc.textFile(biparitePath, partitions).mapPartitions(generator_ij).mapPartitions(init_sp).setName('init_sp').cache()
sj = init.map(lambda x: x[-1]).treeReduce(add)
theta = np.where(sj == 0, 0, d / sj).astype(np.float32)
bc_theta = sc.broadcast(theta)
max_alpha = 10000.0
max_beta = max_alpha + 10
bc_max_beta = sc.broadcast(max_beta)
alpha = np.zeros(j, dtype=np.float32)
alpha_0 = np.zeros(j, dtype=np.float32)
alpha_1 = np.full(j, max_alpha, dtype=np.float32)
loop = init.mapPartitions(init_loop).setName('loop_original').cache()
sum_j = loop.map(lambda x: x[-1]).treeReduce(add)
init.unpersist()
alpha_0, alpha_1, alpha = opt_alpha(sum_j, alpha_0, alpha_1, alpha)
ori_loss = 0
count = 1
tc_w = []
tc_i = []
while True:
bc_alpha = sc.broadcast(alpha)
start = time.time()
new_loop = loop.mapPartitions(main_loop).setName('loop_%d' % count).cache()
new_loop.count()
print('new_loop: ' + str(time.time() - start))
sum_j = new_loop.map(lambda x: x[-1]).treeReduce(add)
loop.unpersist()
loop = new_loop
alpha_0, alpha_1, alpha = opt_alpha(sum_j, alpha_0, alpha_1, alpha)
print('opt_alpha: ' + str(time.time() - start))
count += 1
if count > 20:
break