TensorFlow分布式并行基于gRPC通信框架，一个master负责创建Session，多个worker负责执行计算图任务。
先创建TensorFlow Cluster对象，包含一组task(每个task一台独立机器)，分布式执行TensorFlow计算图。一个Cluster切分多个job，一个job是一类特定任务(parameter server ps,worker)，每个job可以包含多个task。每个task创建一个server，连接到Cluster，每个task执行在不同机器。也可以一台机器执行多个task(不同GPU)。tf.train.ClusterSpec初始化Cluster对象，初始化信息是Python dict，tf.train.ClusterSpec({"ps":["192.168.233.201:2222"],"worker":["192.168.233.202:2222","192.168.233.203:2222"]})，代表一个parameter server和两个worker，分别在三个不同机器上。每个task，定义自己身份，如server=tf.train.Server(cluster,job_name="ps",task_index=0)，机器job定义ps第0台机器。程序中with tf.device("/job:worker/task:7")，限定Variable存放在哪个task或机器。
TensorFlow分布式模式：In-graph replication模型并行，模型计算图不同部分放在不同机器执行；Between-graph replication数据并行，每台机器相同计算图，计算不同batch数据。异步并行，每台机器独立计算梯度，计算完更新到parameter server，不等其他机器。同步并行，等所有机器都完成梯度计算，多个梯度合成统一更新模型参数。同步并行训练，loss下降速度更快，可达到最大精度更高，同步并行速度取决最慢机器，设备速度一致，效率较高。
TensorFlow实现包含1个paramter server、2个worker分布式并行训练程序，MNIST手写数据识别任务示例。写一个完整Python文件，在不同机器不同task执行。载入依赖库。
tf.app.flags定义标记，命令行执行TensorFlow设置参数。命令行指定参数被TensorFlow读取，转flags。设定数据储存目录data_dir默认/tmp/mnist-data，隐藏节点数默认100,训练最大步数train_steps默认1000000,batch_size默认100,学习速率默认0.01。
设定是否使用同步并行标记sync_replicas默认False，命令行执行时可设True开户同步并行。设定需要累计梯度个数更新模型值默认None，代表同步并行积累多少个batch梯度再进行参数更新，设None 为worker数量，所有worker完成一个batch训练后再更新模型参数。
定义ps地址，默认192.168.233.201:2222,根据集群实际情况配置。worker地址设置192.168.233.202:2222和192.168.233.203:2222.设置job_name和task_index FLAG。
flags.FLAGS直接命名FLAGS，简化使用。设置图片尺寸IMAGE_PIXELS 28。
编写程序主函数main，input_data.read_data_sets下载读取MNIST数据集，设置one_hot编码格式。检测命令行输入参数，确保job_name和task_index两个必备参数。显示job_name和task_index，ps、worker所有地址解析成列表ps_spec、worker_spec。
计算总共worker数量，tf.train.ClusterSpec生成TensorFlow Cluster对象，传入参数ps地址信息和worker地址信息。tf.train.Server创建当前机器server，连接Cluster。如当前节点是parameter server，不进行后续操作，server.join等待worker工作。
判断当前机器是否主节点，task_index是否0。定义当前机器worker_device，格式"job:worker/task:0/gpu:0"。多台机器，每台机器1块GPU，总共需要机器数量worker。如一台机器多GPU，一个task管理多个GPU或多个task分别管理。tf.train.replica_device_setter()设置worker资源，worker_device计算资源，ps_device存储模型参数资源。replica_device_setter将模型参数部署在独立ps服务器"/job:ps/cpu:0"，训练操作部署在"/job:worker/task:0/gpu:0"，本机GPU。创建记录全局训练步数变量global_step。
定义神经网络模型，tf.truncated_normal初始化权重，tf.zeros初始化偏置，创建输入 placeholder，tf.nn.xw_plus_b输入矩阵乘法、偏置操作，ReLU激活函数处理，得到第一个隐层输出hid。tf.nn.xw_plus_b、tf.nn.softmax对第一层输出hid处理，得到网络最终输出y。最后计算损失cross_entropy，定义优化器Adam。
判断是否设置同步训练模式sync_replicas，如果同步模型，先获取同步更新模型参数需要副本数replicas_to_aggregate;如果没有单独设置，worker数作默认值。tf.train.SyncReplicasOptimizer创建同步训练优化器，对原有优化器扩展。传入原有优化器及其他参数(replicas_to_aggregate、total_num_replicas、replica_id)，原有优化器改造为同步分布式训练版本。用普通(异步)或同步优化器优化损失cross_entropy。
同步训练模式，主节点，opt.get_chief_queue_runner创建队列执行器，opt.get_init_tokens_op创建全局参数初始化器。
生成本地参数初始化操作init_op，创建临时训练目录，tf.train_Supervisor创建分布式训练监督器，传入参数is_chief、train_dir、init_op。Supervisor管理task参与到分布式训练。
设置Session参数，allow_soft_placement设True，代表操作在指定device不能执行时转到其他device执行。
如果主节点，显示初始化Session，其他节点显示等待主节点初始化操作。执行sv.prepate_or_wait_for_session()。
如果处于同步模型主节点，sv.start_queue_runners执行队列化执行器chief_queue_runner，执行全局参数初始化器init_tokens_op。
训练过程，记录worker执行训练启动时间，初始化本地训练步数local_step，进入训练循环。每步训练，从nnist.train.next_batch读取一个batch数据，生成feed_dict，调train_step执行训练。当全局训练步数达到预设最大值，停止训练。
训练结束，展示总训练时间，在验证数据上计算预测结果损失cross_entropy，展示。
在主程序执行tf.app.run()启动main函数，全部代码保存到distributed.py文件。3台不同机器分别执行distributed.py启动3个task，每次执行distributed.py，传入job_name、task_index指定worker身份。
分别在三台机器192.168.233.201､192.168.233.202、192.168.233.203执行python distributed.py。
同步模式，加上--sync_replicas=True。global_step，异步时，全局步数是所有worker训练步数和，同步时是多少轮并行训练。
#from __future__ import absolute_import #from __future__ import division #from __future__ import print_function import math #import sys import tempfile import time import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data flags = tf.app.flags flags.DEFINE_string("data_dir", "/tmp/mnist-data", "Directory for storing mnist data") #flags.DEFINE_boolean("download_only", False, # "Only perform downloading of data; Do not proceed to " # "session preparation, model definition or training") flags.DEFINE_integer("task_index", None, "Worker task index, should be >= 0. task_index=0 is " "the master worker task the performs the variable " "initialization ") #flags.DEFINE_integer("num_gpus", 2, # "Total number of gpus for each machine." # "If you don't use GPU, please set it to '0'") flags.DEFINE_integer("replicas_to_aggregate", None, "Number of replicas to aggregate before parameter update" "is applied (For sync_replicas mode only; default: " "num_workers)") flags.DEFINE_integer("hidden_units", 100, "Number of units in the hidden layer of the NN") flags.DEFINE_integer("train_steps", 1000000, "Number of (global) training steps to perform") flags.DEFINE_integer("batch_size", 100, "Training batch size") flags.DEFINE_float("learning_rate", 0.01, "Learning rate") flags.DEFINE_boolean("sync_replicas", False, "Use the sync_replicas (synchronized replicas) mode, " "wherein the parameter updates from workers are aggregated " "before applied to avoid stale gradients") #flags.DEFINE_boolean( # "existing_servers", False, "Whether servers already exists. If True, " # "will use the worker hosts via their GRPC URLs (one client process " # "per worker host). Otherwise, will create an in-process TensorFlow " # "server.") flags.DEFINE_string("ps_hosts","192.168.233.201:2222", "Comma-separated list of hostname:port pairs") flags.DEFINE_string("worker_hosts", "192.168.233.202:2223,192.168.233.203:2224", "Comma-separated list of hostname:port pairs") flags.DEFINE_string("job_name", None,"job name: worker or ps") FLAGS = flags.FLAGS IMAGE_PIXELS = 28 def main(unused_argv): mnist = input_data.read_data_sets(FLAGS.data_dir, one_hot=True) # if FLAGS.download_only: # sys.exit(0) if FLAGS.job_name is None or FLAGS.job_name == "": raise ValueError("Must specify an explicit `job_name`") if FLAGS.task_index is None or FLAGS.task_index =="": raise ValueError("Must specify an explicit `task_index`") print("job name = %s" % FLAGS.job_name) print("task index = %d" % FLAGS.task_index) #Construct the cluster and start the server ps_spec = FLAGS.ps_hosts.split(",") worker_spec = FLAGS.worker_hosts.split(",") # Get the number of workers. num_workers = len(worker_spec) cluster = tf.train.ClusterSpec({ "ps": ps_spec, "worker": worker_spec}) #if not FLAGS.existing_servers: # Not using existing servers. Create an in-process server. server = tf.train.Server( cluster, job_name=FLAGS.job_name, task_index=FLAGS.task_index) if FLAGS.job_name == "ps": server.join() is_chief = (FLAGS.task_index == 0) # if FLAGS.num_gpus > 0: # if FLAGS.num_gpus < num_workers: # raise ValueError("number of gpus is less than number of workers") # # Avoid gpu allocation conflict: now allocate task_num -> #gpu # # for each worker in the corresponding machine # gpu = (FLAGS.task_index % FLAGS.num_gpus) # worker_device = "/job:worker/task:%d/gpu:%d" % (FLAGS.task_index, gpu) # elif FLAGS.num_gpus == 0: # # Just allocate the CPU to worker server # cpu = 0 # worker_device = "/job:worker/task:%d/cpu:%d" % (FLAGS.task_index, cpu) # # The device setter will automatically place Variables ops on separate # # parameter servers (ps). The non-Variable ops will be placed on the workers. # # The ps use CPU and workers use corresponding GPU worker_device = "/job:worker/task:%d/gpu:0" % FLAGS.task_index with tf.device( tf.train.replica_device_setter( worker_device=worker_device, ps_device="/job:ps/cpu:0", cluster=cluster)): global_step = tf.Variable(0, name="global_step", trainable=False) # Variables of the hidden layer hid_w = tf.Variable( tf.truncated_normal( [IMAGE_PIXELS * IMAGE_PIXELS, FLAGS.hidden_units], stddev=1.0 / IMAGE_PIXELS), name="hid_w") hid_b = tf.Variable(tf.zeros([FLAGS.hidden_units]), name="hid_b") # Variables of the softmax layer sm_w = tf.Variable( tf.truncated_normal( [FLAGS.hidden_units, 10], stddev=1.0 / math.sqrt(FLAGS.hidden_units)), name="sm_w") sm_b = tf.Variable(tf.zeros([10]), name="sm_b") # Ops: located on the worker specified with FLAGS.task_index x = tf.placeholder(tf.float32, [None, IMAGE_PIXELS * IMAGE_PIXELS]) y_ = tf.placeholder(tf.float32, [None, 10]) hid_lin = tf.nn.xw_plus_b(x, hid_w, hid_b) hid = tf.nn.relu(hid_lin) y = tf.nn.softmax(tf.nn.xw_plus_b(hid, sm_w, sm_b)) cross_entropy = -tf.reduce_sum(y_ * tf.log(tf.clip_by_value(y, 1e-10, 1.0))) opt = tf.train.AdamOptimizer(FLAGS.learning_rate) if FLAGS.sync_replicas: if FLAGS.replicas_to_aggregate is None: replicas_to_aggregate = num_workers else: replicas_to_aggregate = FLAGS.replicas_to_aggregate opt = tf.train.SyncReplicasOptimizer( opt, replicas_to_aggregate=replicas_to_aggregate, total_num_replicas=num_workers, replica_id=FLAGS.task_index, name="mnist_sync_replicas") train_step = opt.minimize(cross_entropy, global_step=global_step) if FLAGS.sync_replicas and is_chief: # Initial token and chief queue runners required by the sync_replicas mode chief_queue_runner = opt.get_chief_queue_runner() init_tokens_op = opt.get_init_tokens_op() init_op = tf.global_variables_initializer() train_dir = tempfile.mkdtemp() sv = tf.train.Supervisor( is_chief=is_chief, logdir=train_dir, init_op=init_op, recovery_wait_secs=1, global_step=global_step) sess_config = tf.ConfigProto( allow_soft_placement=True, log_device_placement=False, device_filters=["/job:ps", "/job:worker/task:%d" % FLAGS.task_index]) # The chief worker (task_index==0) session will prepare the session, # while the remaining workers will wait for the preparation to complete. if is_chief: print("Worker %d: Initializing session..." % FLAGS.task_index) else: print("Worker %d: Waiting for session to be initialized..." % FLAGS.task_index) # if FLAGS.existing_servers: # server_grpc_url = "grpc://" + worker_spec[FLAGS.task_index] # print("Using existing server at: %s" % server_grpc_url) # # sess = sv.prepare_or_wait_for_session(server_grpc_url, config=sess_config) # else: sess = sv.prepare_or_wait_for_session(server.target, config=sess_config) print("Worker %d: Session initialization complete." % FLAGS.task_index) if FLAGS.sync_replicas and is_chief: # Chief worker will start the chief queue runner and call the init op print("Starting chief queue runner and running init_tokens_op") sv.start_queue_runners(sess, [chief_queue_runner]) sess.run(init_tokens_op) # Perform training time_begin = time.time() print("Training begins @ %f" % time_begin) local_step = 0 while True: # Training feed batch_xs, batch_ys = mnist.train.next_batch(FLAGS.batch_size) train_feed = {x: batch_xs, y_: batch_ys} _, step = sess.run([train_step, global_step], feed_dict=train_feed) local_step += 1 now = time.time() print("%f: Worker %d: training step %d done (global step: %d)" % (now, FLAGS.task_index, local_step, step)) if step >= FLAGS.train_steps: break time_end = time.time() print("Training ends @ %f" % time_end) training_time = time_end - time_begin print("Training elapsed time: %f s" % training_time) # Validation feed val_feed = {x: mnist.validation.images, y_: mnist.validation.labels} val_xent = sess.run(cross_entropy, feed_dict=val_feed) print("After %d training step(s), validation cross entropy = %g" % (FLAGS.train_steps, val_xent)) if __name__ == "__main__": tf.app.run()
参考资料：
《TensorFlow实战》
欢迎付费咨询(150元每小时)，我的微信：qingxingfengzi