''' Created on 2017年5月28日 @author: weizhen ''' import time import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import mnist_inference BATCH_SIZE = 100 LEARNING_RATE_BASE = 0.01 TRAINING_STEPS = 1000 LEARNING_RATE_DECAY = 0.99 REGULARAZTION_RATE = 0.0001 # 模型保存路径 MODEL_SAVE_PATH = "/path/to/model" # MNIST数据路径 DATA_PATH = "/path/to/data" # 通过flags指定运行的参数。对于不同的任务task给出了不同的程序 # 但这不是一种可扩展的方式，在这一小节中将使用运行程序是给出的参数来配置在不同任务中运行的程序 FLAGS = tf.app.flags.FLAGS # 指定当前运行的是参数服务器还是计算服务器。参数服务器只负责Tensorflow中变量的维护和管理 # 计算服务器则负责每一轮迭代时运行反向传播过程 tf.app.flags.DEFINE_string('job_name', 'worker', '"ps" or "worker" ') # 指定集群中的参数服务器地址 tf.app.flags.DEFINE_string( 'ps_hosts', 'tf-ps0:2222,tf-ps1:1111', 'Comma-separated list of hostname:port for the parameter server jobs. e.g. "tf-ps0:2222,tf-ps1:1111" ') # 指定集群中的计算服务器地址 tf.app.flags.DEFINE_string( 'worker_hosts', 'tf-worker0:2222,tf-worker1:1111', 'Comma-separated list of hostname:port for the worker jobs.' 'e.g. "tf-worker0:2222,tf-worker1:1111" ') # 指定当前程序的任务ID. Tensorflow 会自动根据参数服务器/计算服务器列表中的端口号 # 来启动服务。注意参数服务器和计算服务器的编号都是从0开始的 tf.app.flags.DEFINE_integer( 'task_id', 0, 'Task ID of the worker/replica running the training.' ) # 定义Tensorflow的计算图，并返回每一轮迭代时需要 运行的操作。 # 为了是处理分布式计算的部分更加突出，本校节将此过程整理为一个函数 def build_model(x, y_, is_chief): regularizer = tf.contrib.layers.l2_regularizer(REGULARAZTION_RATE) # 计算神经网络前向传播的结果 y = mnist_inference.inference(x, regularizer) global_step = tf.Variable(0, trainable=False) # 计算损失函数并定义反向传播过程 cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=tf.argmax(y_, 1)) cross_entropy_mean = tf.reduce_mean(cross_entropy) loss = cross_entropy_mean + tf.add_n(tf.get_collection('losses')) learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, global_step, 60000 / BATCH_SIZE, LEARNING_RATE_DECAY) # 定义每一轮迭代需要运行的操作 train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step) return global_step, loss, train_op # 训练分布式深度学习模型的主过程 def main(argv=None): # 解析flags并通过tf.train.ClusterSpec配置TensorFlow集群 ps_hosts = FLAGS.ps_hosts.split(',') worker_hosts = FLAGS.worker_hosts.split(',') cluster = tf.train.ClusterSpec({"ps":ps_hosts, "worker":worker_hosts}) # 通过ClusterSpec以及当前任务创建Server server = tf.train.Server(cluster, job_name=FLAGS.job_name, task_index=FLAGS.task_id) # 参数服务器只需要管理TensorFlow中的变量，不需要执行训练的过程。server.join()会一直停在这条语句上 if FLAGS.job_name == 'ps': server.join() # 定义计算服务器需要运行的操作。在所有的计算服务器中有一个是主计算服务器。它除了负责计算反向传播的结果，它还负责输出日志和保存模型 is_chief = (FLAGS.task_id == 0) mnist = input_data.read_data_sets(DATA_PATH, one_hot=True) # 通过tf.train.replica_device_setter函数来指定执行每一个运算的设备 # tf.train.replica_device_setter函数会自动将所有的参数分配到参数服务器上，而 # 计算分配到当前的计算服务器上 with tf.device(tf.train.replica_device_setter(worker_device="/job:worker/task:%d " % FLAGS.task_id, cluster=cluster)): x = tf.placeholder(tf.float32, [None, mnist_inference.INPUT_NODE], name='x-input') y_ = tf.placeholder(tf.float32, [None, mnist_inference.OUTPUT_NODE], name='y-input') # 定义训练模型需要运行的操作 global_step, loss, train_op = build_model(x, y_, is_chief) # 定义用于保存模型的saver saver = tf.train.Saver() # 定义日志输出操作 summary_op = tf.summary.merge_all() # 定义病了初始化操作 init_op = tf.global_variables_initializer() # 通过tf.train.Supervisor管理训练深度学习模型的通用功能 # tf.train.Supervisor能统一管理队列操作、模型保存、日志输出以及会话的生成 sv = tf.train.Supervisor( is_chief=is_chief, # 定义当前计算服务器是否为主计算服务器，只用主计算服务器会保存模型以及输出日志 logdir=MODEL_SAVE_PATH, # 指定保存模型和输出日志的地址 init_op=init_op, # 指定初始化操作 summary_op=summary_op, # 指定日志生成操作 saver=saver, # 指定用于保存模型的saver global_step=global_step, # 指定当前迭代的轮数，这个会用于生成保存模型文件的文件名 save_model_secs=60, # 指定保存模型的时间间隔 save_summaries_secs=60 # 指定日志输出的时间间隔 ) sess_config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False) # 通过tf.train.Supervisor生成会话 sess = sv.prepare_or_wait_for_session(server.target, config=sess_config) step = 0 start_time = time.time() # 执行迭代过程。在迭代过程中tf.train.Supervisor会帮助输出日志并保存模型 # 所以不需要直接调用这些过程 while not sv.should_stop(): xs, ys = mnist.train.next_batch(BATCH_SIZE) _, loss_value, global_step_value = sess.run( [train_op, loss, global_step], feed_dict={x:xs, y_:ys}) if global_step_value >= TRAINING_STEPS:break # 每隔一段时间输出训练信息 if step > 0 and step % 100 == 0: duration = time.time() - start_time # 不同的计算服务器都会更新全局的训练轮数，所以这里使用 # global_step_value可以直接得到在训练中使用过的batch的总数 sec_per_batch = duration / global_step_value format_str = ("After %d training steps (%d global steps), loss on training batch is %g. (%.3f sec/batch)") print(format_str % (step, global_step_value, loss_value, sec_per_batch)) step += 1 sv.stop() if __name__ == "__main__": tf.app.run()
下面是训练的结果，需要等到所有的机器都开起来之后才能进行训练
2017-05-28 22:38:45.122523: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:45.122960: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:45.123285: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:45.123847: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:45.124201: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:45.125153: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:45.125514: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:45.126016: W c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoreplatformcpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations. 2017-05-28 22:38:47.211250: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcorecommon_runtimegpugpu_device.cc:887] Found device 0 with properties: name: GeForce 940MX major: 5 minor: 0 memoryClockRate (GHz) 1.189 pciBusID 0000:01:00.0 Total memory: 2.00GiB Free memory: 1.66GiB 2017-05-28 22:38:47.211668: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcorecommon_runtimegpugpu_device.cc:908] DMA: 0 2017-05-28 22:38:47.211848: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcorecommon_runtimegpugpu_device.cc:918] 0: Y 2017-05-28 22:38:47.212045: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcorecommon_runtimegpugpu_device.cc:977] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce 940MX, pci bus id: 0000:01:00.0) 2017-05-28 22:38:47.375428: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtime pcgrpc_channel.cc:200] Initialize GrpcChannelCache for job ps -> {0 -> tf-ps0:2222, 1 -> tf-ps1:1111} 2017-05-28 22:38:47.376363: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtime pcgrpc_channel.cc:200] Initialize GrpcChannelCache for job worker -> {0 -> localhost:2222, 1 -> tf-worker1:1111} 2017-05-28 22:38:47.380830: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtime pcgrpc_server_lib.cc:240] Started server with target: grpc://localhost:2222 Extracting /path/to/data rain-images-idx3-ubyte.gz Extracting /path/to/data rain-labels-idx1-ubyte.gz Extracting /path/to/data 10k-images-idx3-ubyte.gz Extracting /path/to/data 10k-labels-idx1-ubyte.gz 2017-05-28 22:38:58.243494: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:0 2017-05-28 22:38:58.244680: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:1 2017-05-28 22:38:58.247390: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:worker/replica:0/task:1 2017-05-28 22:39:08.248725: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:0 2017-05-28 22:39:08.249804: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:1 2017-05-28 22:39:08.251307: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:worker/replica:0/task:1 2017-05-28 22:39:18.253692: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:0 2017-05-28 22:39:18.254576: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:1 2017-05-28 22:39:18.255448: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:worker/replica:0/task:1 2017-05-28 22:39:28.257660: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:0 2017-05-28 22:39:28.258782: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:ps/replica:0/task:1 2017-05-28 22:39:28.260428: I c: f_jenkinshomeworkspace elease-windevicegpuoswindows ensorflowcoredistributed_runtimemaster.cc:201] CreateSession still waiting for response from worker: /job:worker/replica:0/task:1