安装伪分布式的TensorFlow

安装伪分布式的TensorFlow

一．安装虚拟机Wmware:

二．在虚拟机上创建ubuntu，我在虚拟机上创建了3个Ubuntu，分别为ps,worker0,worker1

三．安装CPU版的TensorFlow

1.    安装Anaconda2

2.    安装TensorFlow：pip install tensorflow*.whl

四．关闭Ubuntu防火墙,每一个虚拟机的防火墙都要关闭,这步必须做，不然一会通信不了

sudo ufw disable

五．测试（参考https://github.com/thewintersun/distributeTensorflowExample）

首先在每一个虚拟机上创建一个python文件，并将一下代码贴入：

1. #coding=utf-8
2. import numpy as np
3. import tensorflow as tf
4.  
5. # Define parameters
6. FLAGS = tf.app.flags.FLAGS
7. tf.app.flags.DEFINE_float('learning_rate',0.00003, 'Initial learning rate.')
8. tf.app.flags.DEFINE_integer('steps_to_validate',1000,
9.                      'Steps to validate andprint loss')
10.  
11. # For distributed
12. tf.app.flags.DEFINE_string("ps_hosts","",
13.                            "Comma-separatedlist of hostname:port pairs")
14. tf.app.flags.DEFINE_string("worker_hosts","",
15.                           "Comma-separated list of hostname:port pairs")
16. tf.app.flags.DEFINE_string("job_name","", "One of 'ps', 'worker'")
17. tf.app.flags.DEFINE_integer("task_index",0, "Index of task within the job")
18. tf.app.flags.DEFINE_integer("issync",0, "是否采用分布式的同步模式，1表示同步模式，0表示异步模式")
19.  
20. # Hyperparameters
21. learning_rate = FLAGS.learning_rate
22. steps_to_validate =FLAGS.steps_to_validate
23.  
24. def main(_):
25.  ps_hosts = FLAGS.ps_hosts.split(",")
26.  worker_hosts = FLAGS.worker_hosts.split(",")
27.  cluster = tf.train.ClusterSpec({"ps": ps_hosts,"worker": worker_hosts})
28.  server =tf.train.Server(cluster,job_name=FLAGS.job_name,task_index=FLAGS.task_index)
29.  
30.  issync = FLAGS.issync
31.  if FLAGS.job_name == "ps":
32.    server.join()
33.  elif FLAGS.job_name == "worker":
34.    with tf.device(tf.train.replica_device_setter(
35.                    worker_device="/job:worker/task:%d" % FLAGS.task_index,
36.                     cluster=cluster)):
37.      global_step = tf.Variable(0, name='global_step', trainable=False)
38.  
39.      input = tf.placeholder("float")
40.      label = tf.placeholder("float")
41.  
42.      weight = tf.get_variable("weight", [1], tf.float32,initializer=tf.random_normal_initializer())
43.      biase  =tf.get_variable("biase", [1], tf.float32,initializer=tf.random_normal_initializer())
44.      pred = tf.multiply(input, weight) + biase
45.  
46.      loss_value = loss(label, pred)
47.      optimizer = tf.train.GradientDescentOptimizer(learning_rate)
48.  
49.      grads_and_vars = optimizer.compute_gradients(loss_value)
50.      if issync == 1:
51.        #同步模式计算更新梯度
52.        rep_op = tf.train.SyncReplicasOptimizer(optimizer,
53.                                                replicas_to_aggregate=len(
54.                                                   worker_hosts),
55.                                                replica_id=FLAGS.task_index,
56.                                                total_num_replicas=len(
57.                                                  worker_hosts),
58.                                                use_locking=True)
59.        train_op = rep_op.apply_gradients(grads_and_vars,
60.                                       global_step=global_step)
61.        init_token_op = rep_op.get_init_tokens_op()
62.        chief_queue_runner = rep_op.get_chief_queue_runner()
63.      else:
64.        #异步模式计算更新梯度
65.        train_op = optimizer.apply_gradients(grads_and_vars,
66.                                       global_step=global_step)
67.  
68.  
69.      init_op = tf.initialize_all_variables()
70.      
71.      saver = tf.train.Saver()
72.      tf.summary.scalar('cost', loss_value)
73.      summary_op = tf.summary.merge_all()
74.  
75.    sv = tf.train.Supervisor(is_chief=(FLAGS.task_index == 0),
76.                            logdir="./checkpoint/",
77.                             init_op=init_op,
78.                             summary_op=None,
79.                             saver=saver,
80.                            global_step=global_step,
81.                             save_model_secs=60)
82.  
83.    with sv.prepare_or_wait_for_session(server.target) as sess:
84.      # 如果是同步模式
85.      if FLAGS.task_index == 0 and issync == 1:
86.        sv.start_queue_runners(sess, [chief_queue_runner])
87.        sess.run(init_token_op)
88.      step = 0
89.      while  step < 1000000:
90.        train_x = np.random.randn(1)
91.        train_y = 2 * train_x + np.random.randn(1) * 0.33  + 10
92.        _, loss_v, step = sess.run([train_op, loss_value,global_step],feed_dict={input:train_x, label:train_y})
93.        if step % steps_to_validate == 0:
94.          w,b = sess.run([weight,biase])
95.          print("step: %d, weight: %f, biase: %f, loss: %f" %(step, w,b, loss_v))
96.  
97.    sv.stop()
98.  
99. def loss(label, pred):
100.  return tf.square(label - pred)
101.  
102.  
103.  
104. if __name__ == "__main__":
105.  tf.app.run()

 

用ifconfig查看的你的ip地址：ps_ip是PS节点上的·IP地址、worker0_ip和worker1_ip是worker节点的ip

在PS节点上用：

1. CUDA_VISIBLE_DEVICES=''python distribute.py --ps_hosts=ps_ip:2222 --worker_hosts=worker0_ip:2223, worker1_ip:2224--job_name=ps --task_index=0

在worker0上;

1. CUDA_VISIBLE_DEVICES=''python distribute.py --ps_hosts=ps_ip:2222 --worker_hosts=worker0_ip:2223, worker1_ip:2224--job_name=ps --task_index=0

在worker1上;

1. CUDA_VISIBLE_DEVICES=''python distribute.py --ps_hosts=ps_ip:2222 --worker_hosts=worker0_ip:2223, worker1_ip:2224--job_name=ps --task_index=1