Default docker image (and all autoscaler images) does not support rllib

[richardliaw]

What is the problem?

Got:

    from ray.rllib.models.preprocessors import get_preprocessor, \
  File "/root/anaconda3/lib/python3.7/site-packages/ray/rllib/models/preprocessors.py", line 2, in <module>
    import cv2
  File "/root/anaconda3/lib/python3.7/site-packages/cv2/__init__.py", line 5, in <module>
    from .cv2 import *
ImportError: libGL.so.1: cannot open shared object file: No such file or directory

when running:

rllib train --run=PPO --env=Pong-v0 --config='{"num_workers": 16, "num_sgd_iter": 1}' --num-samples=10000 --stop='{"timesteps_total": 1}' --ray-address=auto

workaround: conda-forge/pygridgen-feedstock#10

cc @ericl @sven1977 @ijrsvt

[ijrsvt]

@richardliaw Which images (ray, ray-ml, autoscaler) and which tags (nightly, latest)

[richardliaw]

ah, i see - i ran rayproject/ray:latest; perhaps instead, it would be good to add a reference to ray-ml this explicitly in all of the example yamls.

[ericl]

@ijrsvt can you change the yamls to ray-ml by default?

[ijrsvt]

@ericl I'm okay with doing this--my only concern is that it will slow down the pull time.

[ericl]

Can you add the original image as a second line with a comment as a hint to the user? e.g.

image: ray-ml:latest
# image: ray:latest   # use this one if you don't need ML dependencies, it's faster to pull

[ijrsvt]

@ericl Great idea!! :) Will do ASAP