search

bitsauce / Carla-ppo

This repository hosts a customized PPO based agent for Carla. The goal of this project is to make it easier to interact with and experiment in Carla with reinforcement learning based agents -- this, by wrapping Carla in a gym like environment that can handle custom reward functions, custom debug output, etc.
MIT License
229 stars 56 forks source link

Why the car looks very slow when running with " the python run_eval.py --model_name pretrained_agent -start_carla"? #13

Open zengsh-cqupt opened 4 years ago

zengsh-cqupt commented 4 years ago

Hi, wonderful and very interesting case. But ,why the car looks very slow when running with " the python run_eval.py --model_name pretrained_agent -start_carla"? Actually, the speed tells me about 15km/h. Is it normal?

My env is: Clould platform GPU T4 (16G) 4cpu Ubuntu18.04 MobaXterm(SSH log in) Internet clould bandwidth is 5Mbps

Thanks.

zengsh-cqupt commented 4 years ago

when I try to increase my Internet speed, it will be a little fast，but often still only about 3 fps . When I increase my server speed to 50Mbs, it still be slow, max s 4fps.

bitsauce commented 4 years ago

Hi!

Yeah, the evaluation phase is actually a bit slow (I think about 10 FPS in my case IIRC). I don't think I mentioned this anywhere, but the videos that are recorded by run_eval.py when --record_to_file is set are sped up to show how the agent would look if they ran in real-time.

However, if your FPS is 3, then that's probably too low given your specs. Are you running CARLA and the agent on different machines? I haven't tried running these on separate machines over a network before, so it is possible that our custom environment could be utilizing the CARLA API suboptimally in that case.

zengsh-cqupt commented 4 years ago

@bitsauce thanks. I only run the CARLA and the agent at clound server. Thank you for the info above.