How to run OpenAI Gym .render() over a server

问题:

I am running a python 2.7 script on a p2.xlarge AWS server through Jupyter (Ubuntu 14.04). I would like to be able to render my simulations.

Minimal working example

import gym
env = gym.make('CartPole-v0')
env.reset()
env.render()

env.render() makes (among other things) the following errors:

...
HINT: make sure you have OpenGL install. On Ubuntu, you can run 
'apt-get install python-opengl'. If you're running on a server, 
you may need a virtual frame buffer; something like this should work: 
'xvfb-run -s \"-screen 0 1400x900x24\" python <your_script.py>'")
...
NoSuchDisplayException: Cannot connect to "None"

I would like to some how be able to see the simulations. It would be ideal if I could get it inline, but any display method would be nice.

Edit: This is only an issue with some environments, like classic control.

---

Update I

Inspired by this I tried the following, instead of the xvfb-run -s \"-screen 0 1400x900x24\" python <your_script.py> (which I couldn't get to work).

xvfb-run -a jupyter notebook

Running the original script I now get instead

GLXInfoException: pyglet requires an X server with GLX

---

Update II

Issue #154 seems relevant. I tried disabling the pop-up, and directly creating the RGB colors

import gym
env = gym.make('CartPole-v0')
env.reset()

img = env.render(mode='rgb_array', close=True)  
print(type(img)) # <--- <type 'NoneType'>

img = env.render(mode='rgb_array', close=False) # <--- ERROR
print(type(img)) 

I get ImportError: cannot import name gl_info.

---

Update III

With inspiration from @Torxed I tried creating a video file, and then rendering it (a fully satisfying solution).

Using the code from 'Recording and uploading results'

import gym

env = gym.make('CartPole-v0')
env.monitor.start('/tmp/cartpole-experiment-1', force=True)
observation = env.reset()
for t in range(100):
#    env.render()
    print(observation)
    action = env.action_space.sample()
    observation, reward, done, info = env.step(action)
    if done:
        print("Episode finished after {} timesteps".format(t+1))
        break

env.monitor.close()

I tried following your suggestions, but got ImportError: cannot import name gl_info from when running env.monitor.start(....

From my understanding the problem is that OpenAI uses pyglet, and pyglet 'needs' a screen in order to compute the RGB colors of the image that is to be rendered. It is therefore necessary to trick python to think that there is a monitor connected

---

Update IV

FYI there are solutions online using bumblebee that seem to work. This should work if you have control over the server, but since AWS run in a VM I don't think you can use this.

---

Update V

Just if you have this problem, and don't know what to do (like me) the state of most environments are simple enough that you can create your own rendering mechanism. Not very satisfying, but.. you know.

回答1:

Got a simple solution working:

If on a linux server, open jupyter with

$ xvfb-run -s "-screen 0 1400x900x24" jupyter notebook

In Jupyter

import matplotlib.pyplot as plt
%matplotlib inline
from IPython import display

After each step

def show_state(env, step=0, info=""):
    plt.figure(3)
    plt.clf()
    plt.imshow(env.render(mode='rgb_array'))
    plt.title("%s | Step: %d %s" % (env._spec.id,step, info))
    plt.axis('off')

    display.clear_output(wait=True)
    display.display(plt.gcf())

Note: if your environment is not unwrapped, pass env.env to show_state.

回答2:

I managed to run and render openai/gym (even with mujoco) remotely on a headless server.

# Install and configure X window with virtual screen
sudo apt-get install xserver-xorg libglu1-mesa-dev freeglut3-dev mesa-common-dev libxmu-dev libxi-dev
# Configure the nvidia-x
sudo nvidia-xconfig -a --use-display-device=None --virtual=1280x1024
# Run the virtual screen in the background (:0)
sudo /usr/bin/X :0 &
# We only need to setup the virtual screen once

# Run the program with vitural screen
DISPLAY=:0 <program>

# If you dont want to type `DISPLAY=:0` everytime
export DISPLAY=:0

Usage:

DISPLAY=:0 ipython2

Example:

import gym
env = gym.make('Ant-v1')
arr = env.render(mode='rgb_array')
print(arr.shape)
# plot or save wherever you want
# plt.imshow(arr) or scipy.misc.imsave('sample.png', arr)

回答3:

This GitHub issue gave an answer that worked great for me. It's nice because it doesn't require any additional dependencies (I assume you already have matplotlib) or configuration of the server.

Just run, e.g.:

import gym
import matplotlib.pyplot as plt
%matplotlib inline

env = gym.make('Breakout-v0') # insert your favorite environment
render = lambda : plt.imshow(env.render(mode='rgb_array'))
env.reset()
render()

Using mode='rgb_array' gives you back a numpy.ndarray with the RGB values for each position, and matplotlib's imshow (or other methods) displays these nicely.

Note that if you're rendering multiple times in the same cell, this solution will plot a separate image each time. This is probably not what you want. I'll try to update this if I figure out a good workaround for that.

Update to render multiple times in one cell

Based on this StackOverflow answer, here's a working snippet (note that there may be more efficient ways to do this with an interactive plot; this way seems a little laggy on my machine):

import gym
from IPython import display
import matplotlib.pyplot as plt
%matplotlib inline

env = gym.make('Breakout-v0')
env.reset()
for _ in range(100):
    plt.imshow(env.render(mode='rgb_array'))
    display.display(plt.gcf())
    display.clear_output(wait=True)
    action = env.action_space.sample()
    env.step(action)

Update to increase efficiency

On my machine, this was about 3x faster. The difference is that instead of calling imshow each time we render, we just change the RGB data on the original plot.

import gym
from IPython import display
import matplotlib
import matplotlib.pyplot as plt
%matplotlib inline

env = gym.make('Breakout-v0')
env.reset()
img = plt.imshow(env.render(mode='rgb_array')) # only call this once
for _ in range(100):
    img.set_data(env.render(mode='rgb_array')) # just update the data
    display.display(plt.gcf())
    display.clear_output(wait=True)
    action = env.action_space.sample()
    env.step(action)

回答4:

I ran into this myself. Using xvfb as X-server somehow clashes with the Nvidia drivers. But finally this post pointed me into the right direction. Xvfb works without any problems if you install the Nvidia driver with the -no-opengl-files option and CUDA with --no-opengl-libs option. If you know this, it should work. But as it took me quite some time till I figured this out and it seems like I'm not the only one running into problems with xvfb and the nvidia drivers.

I wrote down all necessary steps to set everything up on an AWS EC2 instance with Ubuntu 16.04 LTS here.

回答5:

There's also this solution using pyvirtualdisplay (an Xvfb wrapper). One thing I like about this solution is you can launch it from inside your script, instead of having to wrap it at launch:

from pyvirtualdisplay import Display
display = Display(visible=0, size=(1400, 900))
display.start()

回答6:

I think we should just capture renders as video by using OpenAI Gym wrappers.Monitor and then display it within the Notebook.

Example:

Dependencies

!apt install python-opengl
!apt install ffmpeg
!apt install xvfb
!pip3 install pyvirtualdisplay

# Virtual display
from pyvirtualdisplay import Display

virtual_display = Display(visible=0, size=(1400, 900))
virtual_display.start()

Capture as video

import gym
from gym import wrappers

env = gym.make("SpaceInvaders-v0")
env = wrappers.Monitor(env, "/tmp/SpaceInvaders-v0")

for episode in range(2):
    observation = env.reset()
    step = 0
    total_reward = 0

    while True:
        step += 1
        env.render()
        action = env.action_space.sample()
        observation, reward, done, info = env.step(action)
        total_reward += reward
        if done:
            print("Episode: {0},\tSteps: {1},\tscore: {2}"
                  .format(episode, step, total_reward)
            )
            break
env.close()

Display within Notebook

import os
import io
import base64
from IPython.display import display, HTML

def ipython_show_video(path):
    """Show a video at `path` within IPython Notebook
    """
    if not os.path.isfile(path):
        raise NameError("Cannot access: {}".format(path))

    video = io.open(path, 'r+b').read()
    encoded = base64.b64encode(video)

    display(HTML(
        data="""
        <video alt="test" controls>
        <source src="data:video/mp4;base64,{0}" type="video/mp4" />
        </video>
        """.format(encoded.decode('ascii'))
    ))

ipython_show_video("/tmp/SpaceInvaders-v0/openaigym.video.4.10822.video000000.mp4")

I hope it helps. ;)

回答7:

I avoided the issues with using matplotlib by simply using PIL, Python Image Library:

import gym, PIL
env = gym.make('SpaceInvaders-v0')
array = env.reset()
PIL.Image.fromarray(env.render(mode='rgb_array'))

I found that I didn't need to set the XV frame buffer.

回答8:

I had the same problem and I_like_foxes solution to reinstall nvidia drivers with no opengl fixed things. Here are the commands I used for Ubuntu 16.04 and GTX 1080ti https://gist.github.com/8enmann/931ec2a9dc45fde871d2139a7d1f2d78

回答9:

I was looking for a solution that works in Colaboratory and ended up with this

from IPython import display
import numpy as np
import time

import gym
env = gym.make('SpaceInvaders-v0')
env.reset()

import PIL.Image
import io


def showarray(a, fmt='png'):
    a = np.uint8(a)
    f = io.BytesIO()
    ima = PIL.Image.fromarray(a).save(f, fmt)
    return f.getvalue()

imagehandle = display.display(display.Image(data=showarray(env.render(mode='rgb_array')), width=450), display_id='gymscr')

while True:
    time.sleep(0.01)
    env.step(env.action_space.sample()) # take a random action
    display.update_display(display.Image(data=showarray(env.render(mode='rgb_array')), width=450), display_id='gymscr')

EDIT 1:

You could use xvfbwrapper for the Cartpole environment.

from IPython import display
from xvfbwrapper import Xvfb
import numpy as np
import time
import pyglet
import gym
import PIL.Image
import io    

vdisplay = Xvfb(width=1280, height=740)
vdisplay.start()

env = gym.make('CartPole-v0')
env.reset()

def showarray(a, fmt='png'):
    a = np.uint8(a)
    f = io.BytesIO()
    ima = PIL.Image.fromarray(a).save(f, fmt)
    return f.getvalue()

imagehandle = display.display(display.Image(data=showarray(env.render(mode='rgb_array')), width=450), display_id='gymscr')


for _ in range(1000):
  time.sleep(0.01)
  observation, reward, done, info = env.step(env.action_space.sample()) # take a random action
  display.update_display(display.Image(data=showarray(env.render(mode='rgb_array')), width=450), display_id='gymscr')


vdisplay.stop()

If you're working with standard Jupyter, there's a better solution though. You can use the CommManager to send messages with updated Data URLs to your HTML output.

IPython Inline Screen Example

In Colab the CommManager is not available. The more restrictive output module has a method called eval_js() which seems to be kind of slow.

回答10:

In my IPython environment, Andrew Schreiber's solution can't plot image smoothly. The following is my solution:

If on a linux server, open jupyter with

$ xvfb-run -s "-screen 0 1400x900x24" jupyter notebook

In Jupyter

import matplotlib.pyplot as plt
%matplotlib inline
%matplotlib notebook
from IPython import display

Display iteration:

done = False
obs = env.reset()

fig = plt.figure()
ax = fig.add_subplot(111)
plt.ion()

fig.show()
fig.canvas.draw()

while not done:
    # action = pi.act(True, obs)[0] # pi means a policy which produces an action, if you have
    # obs, reward, done, info = env.step(action) # do action, if you have
    env_rnd = env.render(mode='rgb_array')
    ax.clear()
    ax.imshow(env_rnd)
    fig.canvas.draw()
    time.sleep(0.01)

回答11:

Referencing my other answer here: Display OpenAI gym in Jupyter notebook only

I made a quick working example here which you could fork: https://kyso.io/eoin/openai-gym-jupyter with two examples of rendering in Jupyter - one as an mp4, and another as a realtime gif.

The .mp4 example is quite simple.

import gym
from gym import wrappers

env = gym.make('SpaceInvaders-v0')
env = wrappers.Monitor(env, "./gym-results", force=True)
env.reset()
for _ in range(1000):
    action = env.action_space.sample()
    observation, reward, done, info = env.step(action)
    if done: break
env.close()

Then in a new cell Jupyter cell, or download it from the server onto some place where you can view the video.

import io
import base64
from IPython.display import HTML

video = io.open('./gym-results/openaigym.video.%s.video000000.mp4' % env.file_infix, 'r+b').read()
encoded = base64.b64encode(video)
HTML(data='''
    <video width="360" height="auto" alt="test" controls><source src="data:video/mp4;base64,{0}" type="video/mp4" /></video>'''
.format(encoded.decode('ascii')))

If your on a server with public access you could run python -m http.server in the gym-results folder and just watch the videos there.