I finished the train task on the cluster and try to restore the result on my local machine. However, error occurs due to the different address and the number of clients.
Is there any convenient way to make it?
@tiankaidong It’s hard to say without having an error trace. Could you post the error here?
Usually it should be unproblematic to evaluate a policy on a different machine.
Sorry for the blurry information.
I did the RL train on a cluster with server address of 10.246.224.114. Then copy the result of checkpoints to my local machine with a different IP address and config args.algo_checkpoint_path in server.py as the checkpoints folder. (the server.py script is really similar to the provided cartpole example). When I run the server.py on the local machine，I got the error message:
Creating a PolicyServer on 10.246.224.114:9900 failed!
[2023-09-21 09:38:43,384 E 2968 2968] core_worker.cc:1705: Pushed Error with JobID: 01000000 of type: task with message: ray::RolloutWorker.init() (pid=2968, ip=10.94.133.20, actor_id=a8f6c6634cbe20a73062567501000000, repr=<ray.rllib.evaluation.rollout_worker._modify_class..Class object at 0x7f8f7ae62190>)
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/site-packages/ray/rllib/evaluation/rollout_worker.py”, line 634, in init
self.input_reader: InputReader = self._get_input_creator_from_config()(
File “carfield_server_redis.py”, line 155, in _input
return PolicyServerInput(
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/site-packages/ray/rllib/env/policy_server_input.py”, line 166, in init
HTTPServer.init(self, (address, port), handler)
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/socketserver.py”, line 452, in init
self.server_bind()
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/http/server.py”, line 139, in server_bind
socketserver.TCPServer.server_bind(self)
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/socketserver.py”, line 466, in server_bind
self.socket.bind(self.server_address)
OSError: [Errno 99] Cannot assign requested address at time: 1.69526e+09
2023-09-21 09:38:43,387 ERROR actor.py:970 – Exception raised in creation task: The actor died because of an error raised in its creation task, ray::RolloutWorker.init() (pid=2968, ip=10.94.133.20, actor_id=a8f6c6634cbe20a73062577501000000, repr=<ray.rllib.evaluation.rollout_worker._modify_class..Class object at 0x7f8f7ae62190>)
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/site-packages/ray/rllib/evaluation/rollout_worker.py”, line 634, in init
self.input_reader: InputReader = self._get_input_creator_from_config()(
File “carfield_server_redis.py”, line 155, in _input
return PolicyServerInput(
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/site-packages/ray/rllib/env/policy_server_input.py”, line 166, in init
HTTPServer.init(self, (address, port), handler)
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/socketserver.py”, line 452, in init
self.server_bind()
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/http/server.py”, line 139, in server_bind
socketserver.TCPServer.server_bind(self)
File “/home/xxx/anaconda3/envs/rl_env38/lib/python3.8/socketserver.py”, line 466, in server_bind
self.socket.bind(self.server_address)
OSError: [Errno 99] Cannot assign requested address
I think it should be possible to modify the server IP address and num_workers based on the checkpoints but I didn’t see any documents about that and tried some ideas but failed.