Tensorflow variables not initialized using cross-graph replication

I have Python code test.py

as shown below that uses "Replication between Graphs" for Distributed Tensorflow:

import argparse
import logging

import tensorflow as tf

log = logging.getLogger(__name__)

# Job Names
WORKER_SERVER = "worker"

# Cluster Details
    PARAMETER_SERVER: ["localhost:2222"],
    WORKER_SERVER: ["localhost:1111", "localhost:1112"]}

def parse_command_arguments():
    """ Set up and parse the command line arguments passed for experiment. """
    parser = argparse.ArgumentParser(
        description="Parameters and Arguments for the Test.")
        help="One of 'ps', 'worker'"
    # Flags for defining the tf.train.Server
        help="Index of task within the job"

    return parser.parse_args()

def start_server(job_name, task_index):
    """ Create a server based on a cluster spec. """
    cluster = tf.train.ClusterSpec(CLUSTER_SPEC)
    server = tf.train.Server(
        cluster, job_name=job_name, task_index=task_index)

    return server, cluster

def model():
    """ Build up a simple estimator model. """
    # Build a linear model and predict values
    W = tf.Variable([.3], tf.float32)
    b = tf.Variable([-.3], tf.float32)
    x = tf.placeholder(tf.float32)
    linear_model = W * x + b
    y = tf.placeholder(tf.float32)
    global_step = tf.get_variable('global_step', [],

    # Loss sub-graph
    loss = tf.reduce_sum(tf.square(linear_model - y))

    # optimizer
    optimizer = tf.train.GradientDescentOptimizer(0.01)
    train = optimizer.minimize(loss, global_step=global_step)

    init_op = tf.global_variables_initializer()
    log.info("Variables initialized ...")

    return W, b, loss, x, y, train, global_step, init_op

if __name__ == "__main__":
    # Initializing logging with level "INFO".

    # Parse arguments from command line.
    arguments = parse_command_arguments()
    job_name = arguments.job_name
    task_index = arguments.task_index

    # Start a server.
    server, cluster = start_server(job_name, task_index)

    if job_name == "ps":
        with tf.device(tf.train.replica_device_setter(
                worker_device="/job:worker/task:%d" % task_index,
            W, b, loss, x, y, train, global_step, init_op = model()
        with tf.train.MonitoredTrainingSession(
                is_chief=(arguments.task_index == 0 and (
                            arguments.job_name == 'worker'))) as sess:
            step = 0
            # training data
            x_train = [1, 2, 3, 4]
            y_train = [0, -1, -2, -3]
            while not sess.should_stop() and step < 1000:
                _, step = sess.run(
                    [train, global_step], {x: x_train, y: y_train})

            # evaluate training accuracy
            curr_W, curr_b, curr_loss = sess.run(
                [W, b, loss], {x: x_train, y: y_train})
            print("W: %s b: %s loss: %s" % (curr_W, curr_b, curr_loss))


I ran the code with three different processes on the same machine (MacPro with only processors) following the order below:

  • Parameter server: $ python test.py --task_index 0 --job_name ps

  • Worker 1: $ python test.py --task_index 0 --job_name worker

  • Worker 2: $ python test.py --task_index 1 --job_name worker

and I found that the process for "Worker 2" got into an error:

$ python test.py --task_index 1 --job_name worker
I tensorflow/core/distributed_runtime/rpc/grpc_channel.cc:197] Initialize GrpcChannelCache for job ps -> {0 -> localhost:2222}
I tensorflow/core/distributed_runtime/rpc/grpc_channel.cc:197] Initialize GrpcChannelCache for job worker -> {0 -> localhost:1111, 1 -> localhost:1112}
I tensorflow/core/distributed_runtime/rpc/grpc_server_lib.cc:211] Started server with target: grpc://localhost:1112
INFO:__main__:Variables initialized ...
I tensorflow/core/distributed_runtime/master_session.cc:993] Start master session 9912c75f2921fe13 with config: 

INFO:tensorflow:Waiting for model to be ready.  Ready_for_local_init_op:  None, ready: Variables not initialized: Variable, Variable_1, global_step
INFO:tensorflow:Waiting for model to be ready.  Ready_for_local_init_op:  None, ready: Variables not initialized: Variable, Variable_1, global_step


and this process for "Worker 2" was simply frozen. The error shows that the Tensorflow variables for "Worker 2" are badly initialized, so I am wondering if there is a bug for MonitoredTrainingSession

in terms of coordinating variable initializations in Tensorflow sessions or elsewhere, or I am missing something in my code.

NOTE: The code was running with Tensorflow 0.12


source to share

1 answer

I think this is the "intended behavior" for the coordination protocol tf.train.MonitoredTrainingSession

. In a recent answer, I explained how this protocol is geared towards long training tasks, so the worker will sleep for 30 seconds between checking to see if the variables have been initialized.

There is a race condition between Worker 1 initializing the op and Worker 2 checking the variables, and if Worker 2 wins the race, it will observe that some of the variables are uninitialized and go into a 30 second sleep before re-checking.

However, the total amount of computation in your program is very small, so in this 30-second period, employee 1 will be able to complete his work and shut down. When Worker 2 checks if the variables are initialized, it will create a new tf.Session

one that tries to connect to other tasks, but Worker 1 is no longer running, so you will see a log message like this (repeating every 10 seconds or so):

I tensorflow/core/distributed_runtime/master.cc:193] CreateSession still waiting for response from worker: /job:worker/replica:0/task:0


If the training task is significantly longer than 30 seconds, this will not be a problem.

One way to solve the problem is to eliminate interdependencies between workers by installing a "device filter". Since, in a typical configuration, no individual workers exchange data between graphs, you can tell TensorFlow to ignore the absence of another worker during session creation using tf. ConfigProto


# Each worker only needs to contact the PS task(s) and the local worker task.
config = tf.ConfigProto(device_filters=[
    '/job:ps', '/job:worker/task:%d' % arguments.task_index])

with tf.train.MonitoredTrainingSession(
    is_chief=(arguments.task_index == 0 and (
              arguments.job_name == 'worker'))) as sess:
  # ...




All Articles