def main(_):
train_dir = setup_experiment.setup_experiment("worker%d_logs" % FLAGS.task)
run_fn = truncated_training.worker_run
while True:
logging.info("Starting a new graph reset iteration")
g = tf.Graph()
with g.as_default():
np_global_step = 0
tf.set_random_seed(FLAGS.tf_seed)
logging.info("building graph... with state from %d",
np_global_step)
stime = time.time()
graph_dict = truncated_training.build_graph(
np_global_step=np_global_step)
logging.info("done building graph... (took %f sec)",
time.time() - stime)
# perform a series of unrolls
run_fn(train_dir, graph_dict)
## reset the graph
if FLAGS.master:
logging.info("running tf.session.Reset")
config = tf.ConfigProto(
device_filters=["/job:worker/replica:%d" % FLAGS.task])
tf.Session.reset(FLAGS.master, config=config)
if FLAGS.master == "local":
logging.info("running tf.session.Reset")
config = tf.ConfigProto()
tf.Session.reset(FLAGS.master, config=config)
def main(_):
train_dir = setup_experiment.setup_experiment("chief_logs")
graph_dict = truncated_training.build_graph()
truncated_training.chief_run(train_dir, graph_dict)