gym rendering

[christopherhesse]

render=True is confusing, should be that render_mode="human" triggers the window

[mtrazzi]

Hi @christopherhesse,

I'm probably missing something but it's not clear to me how to use render=True from the README.

When I do:
env = gym.make("procgen:procgen-coinrun-v0", start_level=0, num_levels=1, render=True)

I get:

Traceback (most recent call last):
...
  File "~/mtrazzi/procgen/lib/python3.7/site-packages/procgen/gym_registration.py", line 9, in make_env
    env = ProcgenGym3Env(num=1, num_threads=0, **kwargs)
  File "~/mtrazzi/procgen/lib/python3.7/site-packages/procgen/env.py", line 246, in __init__
    super().__init__(num, env_name, options, **kwargs)
TypeError: __init__() got an unexpected keyword argument 'render_human'

And when I do
env = ProcgenGym3Env(num=1, env_name="coinrun", start_level=0, num_levels=1, render=True)

I get:

Traceback (most recent call last):
...
  File "~/mtrazzi/procgen/lib/python3.7/site-packages/procgen/env.py", line 246, in __init__
  super().__init__(num, env_name, options, **kwargs)
  TypeError: __init__() got an unexpected keyword argument 'render'

Would be great to have minimal examples on how to render things in the README.

On a related note, when trying to render with mode=human or mode=rgb_array I don't get anything (maybe because of early reset, as mentioned in the README), and this is still true when using a wrapper from the NeurIPS procgen competion, in particular when running rollouts.py with --video-dir ./outputs (which calls the render() method multiple times inside a Monitor, but then doesn't seem to save any video). Any ideas on why this is hapenning?

Thanks!

[christopherhesse]

@mtrazzi what's the output of pip show procgen and pip show gym3?

[mtrazzi]

(Note: I changed computer but I get the same tracebacks.)

Version: 0.10.1
Summary: Procedurally Generated Game-Like RL Environments
Home-page: https://github.com/openai/procgen
Author: OpenAI
Author-email: None
License: UNKNOWN
Location: /home/a/procgen/lib/python3.6/site-packages
Requires: gym3, gym, filelock, numpy
Required-by: 

Name: gym3
Version: 0.3.2
Summary: Vectorized Reinforcement Learning Environment Interface
Home-page: https://github.com/openai/gym3
Author: OpenAI
Author-email: None
License: UNKNOWN
Location: /home/a/procgen/lib/python3.6/site-packages
Requires: imageio-ffmpeg, moderngl, glfw, numpy, imageio, cffi
Required-by: procgen

After updating procgen to 0.10.3 I still get the same Traceback for the ProcgenGym3Env command, but the gym.make command seems to run (and render) something!

Also, I got an answer for the NeurIPS competition, and apparently adding a render_mode="rgb_array" in the config should work.

[christopherhesse]

0.10.3 did contain a fix for that error. The docs are a bit ambiguous and partially refer to the previous argument, I'll fix them.

In the future to avoid this sort of stuff, the gym environment will take render_mode="rgb_array" or render_mode="human" and the gym3 will accept only render_mode="rgb_array"

[christopherhesse]

The docs are updated, 0.10.4 will likely contain the fix I mentioned, and I'll leave render=True there for backwards compatibility.

[mtrazzi]

Hi, thanks for updating the docs.

It seems that passing render_mode='rgb_array' works fine and sets configs correctly.
A slightly modified of the ViewerWrapper demo (cf. PR) render episodes correctly.

Now the question is how to get a video from rollouts, and it seems that gym3's VideoRecorderWrapper should do the trick.

When I replace the identity env in video_recorder_test.py by
env = gym.make("procgen:procgen-coinrun-v0", start_level=0, num_levels=1, render_mode='rgb_array')

I get

Traceback:
...
  super().__init__(ob_space or env.ob_space, ac_space or env.ac_space, env.num)
AttributeError: 'ToGymEnv' object has no attribute 'ob_space'

and when I replace it with env = procgen.ProcgenGym3Env(num=1, env_name="coinrun", render_mode='rgb_array')

I get

  File "script.py", line 54, in test_recorder
    env.act(types_np.zeros(env.ac_space, bshape=(env.num,)))
  File "~/mtrazzi/procgen/lib/python3.7/site-packages/gym3/video_recorder.py", line 101, in act
    self._append_observation()
  File "~/mtrazzi/procgen/lib/python3.7/site-packages/gym3/video_recorder.py", line 77, in _append_observation
    img = ob[self._env_index]
KeyError: 0

Simply using a Monitor doesn't seem to work for gym3/procgen environments so having a demo for video seems useful.
Any ideas on what I'm missing?

Thanks

[mtrazzi]

Update: problem is solved, code for a video of coinrun here.

[christopherhesse]

That code seems fine, VideoRecorderWrapper will also work, but you have to set info_key="rgb". The error is because by default it expects that the environment returns images in a numpy array. A better error message and/or fallback to "rgb" there would be an improvement.

[christopherhesse]

Here is an example script for future reference:

"""
Example random agent script using the gym3 API to demonstrate that procgen works
"""

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