tensorflow 分布式 全局变量 数据并行 同步训练 in-graph 实例

import sys
task_number = int(sys.argv[1])

import tensorflow as tf

cluster = tf.train.ClusterSpec({"local": ["localhost:2222", "localhost:2223"]})
server = tf.train.Server(cluster, job_name="local", task_index=task_number)

print("Starting server #{}".format(task_number))

server.start()
server.join()

import numpy as np
import tensorflow as tf
import sys

cluster = tf.train.ClusterSpec({"local": ["localhost:2222", "localhost:2223"]})

point_num = int(sys.argv[1])

x = tf.placeholder(tf.float32, point_num)
y = tf.placeholder(tf.float32, point_num)

with tf.device("/job:local/task:1"):
batch_x1 = tf.slice(x, [0], [point_num/2])
batch_y1 = tf.slice(y, [0], [point_num/2])
result1 = tf.add(tf.square(batch_x1), tf.square(batch_y1))

with tf.device("/job:local/task:0"):
batch_x2 = tf.slice(x, [point_num/2], [-1])
batch_y2 = tf.slice(y, [point_num/2], [-1])
result2 = tf.add(tf.square(batch_x2), tf.square(batch_y2))
distance = tf.concat(0,[result1, result2])

with tf.Session("grpc://localhost:2222") as sess:
result = sess.run(distance, feed_dict={x: np.random.random(point_num), y: np.random.random(point_num)})
sum = 0;
# count the point in the quadrant
for i in range(point_num):
if result[i] < 1:
sum += 1;
print "pi = %f" % (float(sum) / point_num  * 4)

python compute-pi.py 1000
python create_worker 0
python create_worker 1