def main(_):
config = get_config(FLAGS) or FLAGS
config.cnn_format = 'NHWC'
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)
if FLAGS.job_name == "ps":
server.join()
elif FLAGS.job_name == "worker":
env = GymEnvironment(config)
with tf.device(
tf.train.replica_device_setter(
worker_device="/job:worker/task:%d" % FLAGS.task_index,
cluster=cluster)):
lr_op = tf.placeholder('float', None, name='learning_rate')
optimizer = tf.train.RMSPropOptimizer(lr_op,
decay=0.99,
momentum=0,
epsilon=0.1)
agent = Agent(config, env, optimizer, lr_op)
agent.ep_end = random.sample([0.1, 0.01, 0.5], 1)[0]
print(agent.model_dir)
# Create a "supervisor", which oversees the training process.
is_chief = (FLAGS.task_index == 0)
sv = tf.train.Supervisor(is_chief=is_chief,
logdir="./logs/" + agent.model_dir,
init_op=agent.init_op,
summary_op=None,
saver=agent.saver,
global_step=agent.step_op,
save_model_secs=600)
if FLAGS.is_train:
if is_chief:
train_or_play = agent.train_with_summary
else:
train_or_play = agent.train
else:
train_or_play = agent.play
with sv.managed_session(server.target) as sess:
agent.sess = sess
agent.update_target_q_network()
train_or_play(sv, is_chief)
# Ask for all the services to stop.
sv.stop()
def main(_):
config = get_config(FLAGS) or FLAGS
config.cnn_format = 'NHWC'
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)
if FLAGS.job_name == "ps":
server.join()
elif FLAGS.job_name == "worker":
env = GymEnvironment(config)
with tf.device(tf.train.replica_device_setter(
worker_device="/job:worker/task:%d" % FLAGS.task_index,
cluster=cluster)):
lr_op = tf.placeholder('float', None, name='learning_rate')
optimizer = tf.train.RMSPropOptimizer(
lr_op, decay=0.99, momentum=0, epsilon=0.1)
agent = Agent(config, env, optimizer, lr_op)
agent.ep_end = random.sample([0.1, 0.01, 0.5], 1)[0]
print(agent.model_dir)
# Create a "supervisor", which oversees the training process.
is_chief = (FLAGS.task_index == 0)
sv = tf.train.Supervisor(is_chief=is_chief,
logdir="./logs/" + agent.model_dir,
init_op=agent.init_op,
summary_op=None,
saver=agent.saver,
global_step=agent.step_op,
save_model_secs=600)
if FLAGS.is_train:
if is_chief:
train_or_play = agent.train_with_summary
else:
train_or_play = agent.train
else:
train_or_play = agent.play
with sv.managed_session(server.target) as sess:
agent.sess = sess
agent.update_target_q_network()
train_or_play(sv, is_chief)
# Ask for all the services to stop.
sv.stop()
def main(_):
config = get_config(FLAGS) or FLAGS
config.cnn_format = 'NHWC'
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)
if FLAGS.job_name == "ps":
server.join()
elif FLAGS.job_name == "worker":
env = GymEnvironment(config)
with tf.device(
tf.train.replica_device_setter(
worker_device="/job:worker/task:%d" % FLAGS.task_index,
cluster=cluster)):
lr_op = tf.placeholder('float', None, name='learning_rate')
optimizer = tf.train.RMSPropOptimizer(lr_op,
decay=0.99,
momentum=0,
epsilon=0.1)
agent = Agent(config, env, optimizer, lr_op)
agent.ep_end = random.sample([0.1, 0.01, 0.5], 1)[0]
print(agent.model_dir)
logdir = "./logs/" + agent.model_dir
is_chief = (FLAGS.task_index == 0)
if is_chief:
agent.summary_writer = tf.summary.FileWriter(logdir)
if FLAGS.is_train:
if is_chief:
train_or_play = agent.train_with_summary
else:
train_or_play = agent.train
else:
train_or_play = agent.play
hooks = [
tf.train.StopAtStepHook(last_step=config.max_step),
]
scaffold = tf.train.Scaffold(saver=agent.saver)
with tf.train.MonitoredTrainingSession(
master=server.target,
is_chief=is_chief,
scaffold=scaffold,
checkpoint_dir=logdir,
save_summaries_steps=None, # Disable default SummarySaverHook.
hooks=hooks) as sess:
agent.sess = sess
agent.update_target_q_network()
train_or_play(is_chief)
