momentum=0.0,
epsilon=RMSP_EPSILON,
clip_norm=GRAD_NORM_CLIP,
device=device)
# instantiate each training thread
# each thread is training for one target in one scene
training_threads = []
for i in range(PARALLEL_SIZE):
scene, task = branches[i % NUM_TASKS]
training_thread = A3CTrainingThread(i,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
MAX_TIME_STEP,
device=device,
network_scope="thread-%d" %
(i + 1),
scene_scope=scene,
task_scope=task)
training_threads.append(training_thread)
# prepare session
sess = tf.Session(config=tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# create tensorboard summaries
# instantiate each training thread
# each thread is training for one target in one scene
training_threads = []
for i in range(PARALLEL_SIZE
): #each local network has dedicated scene and a target
scene, task = branches[i % NUM_TASKS]
training_thread = A3CTrainingThread(
i,
global_network,
global_discriminator,
initial_learning_rate, #each thread trained of seperate secene and task
learning_rate_input,
grad_applier,
grad_applier_discriminator,
MAX_TIME_STEP,
device=device,
device2=device2,
network_scope="thread-%d" % (i + 1),
scene_scope=scene,
task_scope=task)
training_threads.append(training_thread)
# prepare session
sess = tf.Session(config=tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
training_threads = []
learning_rate_input = tf.placeholder("float")
grad_applier = RMSPropApplier(learning_rate=learning_rate_input,
decay=settings.rmsp_alpha,
momentum=0.0,
epsilon=settings.rmsp_epsilon,
clip_norm=settings.grad_norm_clip,
device=device)
for i in range(settings.parallel_agent_size):
training_thread = A3CTrainingThread(
i, global_network, initial_learning_rates[i], learning_rate_input,
grad_applier, settings.max_time_step, device, settings.action_size,
settings.gamma, settings.local_t_max, settings.entropy_beta,
settings.agent_type, settings.performance_log_interval,
settings.log_level, settings.random_seed)
training_threads.append(training_thread)
# prepare session
sess = tf.Session(config=tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# Statistics summary writer
summary_writer = tf.summary.FileWriter(LOG_FILE, sess.graph)
statistics = Statistics(sess, summary_writer, settings.average_summary)
grad_applier = tf.train.RMSPropOptimizer(
learning_rate=learning_rate_input,
decay=RMSP_ALPHA,
momentum=0.0,
epsilon=RMSP_EPSILON)
# instantiate each training thread
training_threads = []
for i in range(PARALLEL_SIZE):
training_thread = A3CTrainingThread(
config,
env,
i,
global_network_scope,
initial_learning_rate,
learning_rate_input,
grad_applier,
MAX_TIME_STEP,
device=device
)
training_threads.append(training_thread)
# prepare session
sess = tf.Session(config=tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True))
init = tf.global_variables_initializer()
sess.run(init)
# create tensorboard summaries