def create_sync_net_op(self):
actor_sync_net = sync_networks_op(self.master.shared_actor_net,
self.actor_net.vars, self.thread_id)
critic_sync_net = sync_networks_op(self.master.shared_critic_net,
self.critic_net.vars,
self.thread_id)
return tf.group(actor_sync_net, critic_sync_net)
def make_trainer(self):
actor_optimizer = tf.train.AdamOptimizer(
learning_rate=self.config["actor_learning_rate"])
critic_optimizer = tf.train.AdamOptimizer(
learning_rate=self.config["critic_learning_rate"])
self.actor_sync_net = sync_networks_op(self.master.shared_actor_net,
self.actor_net.vars,
self.thread_id)
actor_grads = tf.gradients(self.actor_net.loss, self.actor_net.vars)
self.critic_sync_net = sync_networks_op(self.master.shared_critic_net,
self.critic_net.vars,
self.thread_id)
critic_grads = tf.gradients(self.critic_net.loss, self.critic_net.vars)
if self.clip_gradients:
# Clipped gradients
gradient_clip_value = self.config["gradient_clip_value"]
processed_actor_grads = [
tf.clip_by_value(grad, -gradient_clip_value,
gradient_clip_value) for grad in actor_grads
]
processed_critic_grads = [
tf.clip_by_value(grad, -gradient_clip_value,
gradient_clip_value) for grad in critic_grads
]
else:
# Non-clipped gradients: don't do anything
processed_actor_grads = actor_grads
processed_critic_grads = critic_grads
# Apply gradients to the weights of the master network
# Only increase global_step counter once per update of the 2 networks
apply_actor_gradients = actor_optimizer.apply_gradients(
zip(processed_actor_grads, self.master.shared_actor_net.vars),
global_step=self.master.global_step)
apply_critic_gradients = critic_optimizer.apply_gradients(
zip(processed_critic_grads, self.master.shared_critic_net.vars))
return tf.group(apply_actor_gradients, apply_critic_gradients)
