The baseline PPO algorithm follows the original paper in https://arxiv.org/pdf/1707.06347.pdf
#1.train PPO python train_PPO.py GPU_id [pretrained_path] #The GPU_id value should be one of ['cpu','0','1','2','3']
model checkpoint path:baseline/PPO/results
----Communication setting see ../baseline_utils.py class Game
--reset method socket = context.socket(zmq.REQ) socket.connect('tcp://127.0.0.1:'+self.port_initial) #port_initial default 6667 #inform backend to initialize environment socket.send(b'initial')
#create new socket to receive initial environment self.socket = context.socket(zmq.REP) self.socket.bind('tcp://*:'+self.port_recv) #port_recv default 5556 message_raw = self.socket.recv()
--step method self.socket = context.socket(zmq.REP) self.socket.bind('tcp://*:'+ self.port_recv) #same as above self.socket
#send action to backend self.socket.send(message.SerializeToString()) #receive new environment information message_raw = self.socket.recv()
#default model checkpoint path is PPO/results
#2.test PPO #Method 1 python test_PPO.py GPU_id PORT [pretrained_path]
#method 2 #or you can modify inference.py and execute python inference.py > inference.sh bash inference.sh
#and you can execute bash start_reboot.sh to reboot inference server when it crash because of torch.cuda.OutOfMemoryError