Tensorflow1.x系列(6): TensorFlow计算加速实现

1. GPU实现

# # 通过tf.device将运算指定到特定的设备上。
with tf.device('/cpu:0'):
	a = tf.constant([1.0, 2.0, 3.0], shape=[3], name='a')
	b = tf.constant([1.0, 2.0, 3.0], shape=[3], name='b')
with tf.device('/gpu:0'):
    c = a + b

sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
print(sess.run(c))

如果把GPU换成/gpu:1 ,就会出错，这是什么原因？难道只能按照顺序来制定GPU？

InvalidArgumentError (see above for traceback): Cannot assign a device for opera
tion 'add': Operation was explicitly assigned to /device:GPU:1 but available dev
ices are [ /job:localhost/replica:0/task:0/device:CPU:0, /job:localhost/replica:
0/task:0/device:GPU:0 ]. Make sure the device specification refers to a valid de
vice.
         [[Node: add = Add[T=DT_FLOAT, _device="/device:GPU:1"](a, b)]]

还有一个问题，按道理，TensorFlow在device mapping的过程应该使用不同的GPU，该如何查看？

1.2GPU的允许操作

GPU只支持实数型（float16、float32和double类型的参数）

TensorFlow支持使用 allow_soft_placement来将不能放在GPU的操作，自动放置到CPU来进行运算。

2. 分布式TensorFlow

分布式TensorFlow 运行方式分成 ：同步模式，异步模式；
两者的区别在于，迭代一次之后，不同服务器的参数反馈是否同步更新。

因此，同步模式存在等待最慢服务器的缺点；异步模式存在无法得到较优训练结果的坏处。

2.1 分布式TensorFlow原理

创建单机器集群

import tensorflow as tf

# 创建一个本地集群。
c = tf.constant("Hello, distributed TensorFlow!")
server = tf.train.Server.create_local_server()
sess = tf.Session(server.target)
print(sess.run(c))

运行结果

创建两台机器集群

#coding=utf-8
#local_reload_test.py

import numpy as np
import tensorflow as tf

# Define parameters
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_float('learning_rate', 0.00003, 'Initial learning rate.')

# Hyperparameters
learning_rate = FLAGS.learning_rate

def main(_):

  global_step = tf.Variable(0, name='global_step', trainable=False)

  input = tf.placeholder("float")
  label = tf.placeholder("float")

  weight = tf.get_variable("weight", [1], tf.float32, initializer=tf.random_normal_initializer())
  biase  = tf.get_variable("biase", [1], tf.float32, initializer=tf.random_normal_initializer())
  pred = tf.multiply(input, weight) + biase

  loss_value = loss(label, pred)

  train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss_value, global_step=global_step)
  init_op = tf.global_variables_initializer()
  
  saver = tf.train.Saver()
  session     = tf.Session()
  ckpt = tf.train.get_checkpoint_state("./checkpoint/")
  if ckpt and tf.gfile.Exists(ckpt.model_checkpoint_path):
    saver.restore(session, ckpt.model_checkpoint_path)
    global_step = int(ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
    print("%s, global_step = %d" % (ckpt.model_checkpoint_path, global_step))
  else:
	  return

  w,b = session.run([weight,biase])
  print("weight: %f, biase: %f" %(w, b))


def loss(label, pred):
  return tf.square(label - pred)

if __name__ == "__main__":
  tf.app.run()

代码解释：

#coding=utf-8
#distributed

import numpy as np
import tensorflow as tf

# Define parameters
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_float('learning_rate', 0.00003, 'Initial learning rate.')
tf.app.flags.DEFINE_integer('steps_to_validate', 1000,
                     'Steps to validate and print loss')

# For distributed
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")
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")

# Hyperparameters
learning_rate = FLAGS.learning_rate
steps_to_validate = FLAGS.steps_to_validate

def main(_):
  ps_hosts = FLAGS.ps_hosts.split(",")
  worker_hosts = FLAGS.worker_hosts.split(",")
  cluster = tf.train.ClusterSpec({"ps": ps_hosts, "worker": worker_hosts})
  server = tf.train.Server(cluster,job_name=FLAGS.job_name,task_index=FLAGS.task_index)

  issync = FLAGS.issync
  if FLAGS.job_name == "ps":
    server.join()
  elif FLAGS.job_name == "worker":
    with tf.device(tf.train.replica_device_setter(
                    worker_device="/job:worker/task:%d" % FLAGS.task_index,
                    cluster=cluster)):
      global_step = tf.Variable(0, name='global_step', trainable=False)

      input = tf.placeholder("float")
      label = tf.placeholder("float")

      weight = tf.get_variable("weight", [1], tf.float32, initializer=tf.random_normal_initializer())
      biase  = tf.get_variable("biase", [1], tf.float32, initializer=tf.random_normal_initializer())
      pred = tf.multiply(input, weight) + biase

      loss_value = loss(label, pred)
      optimizer = tf.train.GradientDescentOptimizer(learning_rate)

      grads_and_vars = optimizer.compute_gradients(loss_value)
      if issync == 1:
        #同步模式计算更新梯度
        rep_op = tf.train.SyncReplicasOptimizer(optimizer,
                                                replicas_to_aggregate=len(
                                                  worker_hosts),
                                                replica_id=FLAGS.task_index,
                                                total_num_replicas=len(
                                                  worker_hosts),
                                                use_locking=True)
        train_op = rep_op.apply_gradients(grads_and_vars,
                                       global_step=global_step)
        init_token_op = rep_op.get_init_tokens_op()
        chief_queue_runner = rep_op.get_chief_queue_runner()
      else:
        #异步模式计算更新梯度
        train_op = optimizer.apply_gradients(grads_and_vars,
                                       global_step=global_step)


      init_op = tf.initialize_all_variables()
      
      saver = tf.train.Saver()
      tf.summary.scalar('cost', loss_value)
      summary_op = tf.summary.merge_all()
 
    sv = tf.train.Supervisor(is_chief=(FLAGS.task_index == 0),
                            logdir="./checkpoint/",
                            init_op=init_op,
                            summary_op=None,
                            saver=saver,
                            global_step=global_step,
                            save_model_secs=60)

    with sv.prepare_or_wait_for_session(server.target) as sess:
      # 如果是同步模式
      if FLAGS.task_index == 0 and issync == 1:
        sv.start_queue_runners(sess, [chief_queue_runner])
        sess.run(init_token_op)
      step = 0
      while  step < 1000000:
        train_x = np.random.randn(1)
        train_y = 2 * train_x + np.random.randn(1) * 0.33  + 10
        _, loss_v, step = sess.run([train_op, loss_value,global_step], feed_dict={input:train_x, label:train_y})
        if step % steps_to_validate == 0:
          w,b = sess.run([weight,biase])
          print("step: %d, weight: %f, biase: %f, loss: %f" %(step, w, b, loss_v))

    sv.stop()

def loss(label, pred):
  return tf.square(label - pred)

if __name__ == "__main__":
  tf.app.run()

设定使用的GPU

在GIT运行

RoFun@RoseFun MINGW64 ~
$ export CUDA_VISIBLE_DEVICES=0

然后运行代码distributed.py（确保代码在c:\username\user里，或者先在cmd上调出代码运行的路径）

$ CUDA_VISIBLE_DEVICES=0 python distributed.py --ps_hosts=122.225.220.136:2222

---

参考：

1. 分布式TensorFlow；
2. tensorflow入门教程之CIFAR-10源代码；
3. 分布式TensorFlow_简书;
4. Github案例；
5. tensorflow gpu使用_ CUDA_VISIBLE_DEVICES