def saveModel(self,
s_batch,
a_batch,
r_batch,
entropy_record,
end_of_video=False):
actor_gradient, critic_gradient, td_batch = \
a3c.compute_gradients(s_batch=np.stack(s_batch, axis=0), # ignore the first chuck
a_batch=np.vstack(a_batch), # since we don't have the
r_batch=np.vstack(r_batch), # control over it
terminal=end_of_video, actor=self.actor, critic=self.critic)
td_loss = np.mean(td_batch)
self.actor_gradient_batch.append(actor_gradient)
self.critic_gradient_batch.append(critic_gradient)
myprint("====")
myprint("Master: Quality: Epoch", self.epoch)
myprint("TD_loss", td_loss, "Avg_reward", np.mean(r_batch),
"Avg_entropy", np.mean(entropy_record))
myprint("====")
summary_str = self.sess.run(self.summary_ops,
feed_dict={
self.summary_vars[0]:
td_loss,
self.summary_vars[1]:
np.mean(r_batch),
self.summary_vars[2]:
np.mean(entropy_record)
})
self.writer.add_summary(summary_str, self.epoch)
self.writer.flush()
self.entropy_record = []
if len(self.actor_gradient_batch) >= GRADIENT_BATCH_SIZE:
assert len(self.actor_gradient_batch) == len(
self.critic_gradient_batch)
for i in range(len(self.actor_gradient_batch)):
self.actor.apply_gradients(self.actor_gradient_batch[i])
self.critic.apply_gradients(self.critic_gradient_batch[i])
self.actor_gradient_batch = []
self.critic_gradient_batch = []
self.epoch += 1
if self.epoch % MODEL_SAVE_INTERVAL == 0:
# Save the neural net parameters to disk.
save_path = self.saver.save(
self.sess, self.summary_dir + "/nn_model_ep_" +
str(self.epoch) + ".ckpt")
myprint("Model saved in file: %s" % save_path)
return self.getParams()
def saveModel(self, end_of_video=False):
if self._vReadOnly:
return
if self.ipcQueue:
self.ipcQueue[0].put({
"id":
self.ipcId,
"cmd":
IPC_CMD_UPDATE,
"pid":
self.pid,
"data": [
self.s_batch, self.a_batch, self.r_batch,
self.entropy_record, end_of_video
]
})
res = None
while True:
res = self.ipcQueue[1].get()
pid = res["pid"]
res = res["res"]
if pid == self.pid:
break
actor_net_params, critic_net_params = res
self.actor.set_network_params(actor_net_params)
self.critic.set_network_params(critic_net_params)
del self.s_batch[:]
del self.a_batch[:]
del self.r_batch[:]
del self.entropy_record[:]
return
actor_gradient, critic_gradient, td_batch = \
a3c.compute_gradients(s_batch=np.stack(self.s_batch, axis=0), # ignore the first chuck
a_batch=np.vstack(self.a_batch), # since we don't have the
r_batch=np.vstack(self.r_batch), # control over it
terminal=end_of_video, actor=self.actor, critic=self.critic)
td_loss = np.mean(td_batch)
self.actor_gradient_batch.append(actor_gradient)
self.critic_gradient_batch.append(critic_gradient)
myprint("====")
myprint("Quality: Epoch", self.epoch)
myprint("TD_loss", td_loss, "Avg_reward", np.mean(self.r_batch),
"Avg_entropy", np.mean(self.entropy_record))
myprint("====")
summary_str = self.sess.run(self.summary_ops,
feed_dict={
self.summary_vars[0]:
td_loss,
self.summary_vars[1]:
np.mean(self.r_batch),
self.summary_vars[2]:
np.mean(self.entropy_record)
})
self.writer.add_summary(summary_str, self.epoch)
self.writer.flush()
self.entropy_record = []
if len(self.actor_gradient_batch) >= GRADIENT_BATCH_SIZE:
assert len(self.actor_gradient_batch) == len(
self.critic_gradient_batch)
for i in range(len(self.actor_gradient_batch)):
self.actor.apply_gradients(self.actor_gradient_batch[i])
self.critic.apply_gradients(self.critic_gradient_batch[i])
self.actor_gradient_batch = []
self.critic_gradient_batch = []
self.epoch += 1
if self.epoch % MODEL_SAVE_INTERVAL == 0:
# Save the neural net parameters to disk.
save_path = self.saver.save(
self.sess, self.summary_dir + "/nn_model_ep_" +
str(self.epoch) + ".ckpt")
myprint("Model saved in file: %s" % save_path)
del self.s_batch[:]
del self.a_batch[:]
del self.r_batch[:]