search

AIcrowd / neurips2020-procgen-starter-kit

Starter Kit for NeurIPS 2020 - Procgen Competition on AIcrowd
Apache License 2.0
90 stars 43 forks source link

Is it possible to see how the agent runs graphically? #22

Open maraoz opened 4 years ago

maraoz commented 4 years ago

I've been able to get the repo up and running, and could train and rollout successfully, but I didn't understand too much what was going on. Is it possible to actually visually see the agent playing the games? Thanks!

maraoz commented 4 years ago

I found the --video-dir option in rollout.py, but it's only creating some .json files when I use it.

jyotishp commented 4 years ago

@maraoz If you are using procgen==0.10.x, you need to pass "render_mode": "rgb_array" as env config. There is a bug in the gym3 -> gym conversion. We added a temporary fix for that in https://github.com/AIcrowd/neurips2020-procgen-starter-kit/pull/24.

We generate videos for every few iterations during the evaluations and post the videos on the issues page and the dashboard.

CaesarQ commented 4 years ago

At present the fix suggested in #24 is not working for me.

I've only been able to get procgen render to work under gym3 via:

from gym3 import types_np
from gym3 import VideoRecorderWrapper
from procgen import ProcgenGym3Env
env = ProcgenGym3Env(num=1, env_name="coinrun", render_mode="rgb_array")
env = VideoRecorderWrapper(env=env, directory=".", info_key="rgb")
step = 0
while True:
    env.act(types_np.sample(env.ac_space, bshape=(env.num,)))
    rew, obs, first = env.observe()
    print(f"step {step} reward {rew} first {first}")
    if step > 0 and first:
        break
    step += 1

But the starter kit uses gym, so the above approach does not work