kouroshHakha

The one that works on CI machine
# https://console.anyscale-staging.com/o/anyscale-internal/configurations/app-config-details/bld_izp5d4ptbfkla4q1prbuxydpit

[INFO] 1/24/2023, 2:18:26 AM: * Step 19/28 (commit,modifyfs) : RUN pip install --upgrade --force-reinstall --no-cache-dir --use-deprecated=legacy-resolver     "gymnasium[atari,mujoco]==0.26.3"     "ale-py==0.8.0"     "gym==0.26.2"     "mujoco-py<2.2,>=2.1"     "autorom[accept-rom-license]"     && pip freeze  (d6259223)
[INFO] 1/24/2023, 2:18:26 AM: Client TLS is disabled
[INFO] 1/24/2023, 2:18:26 AM: reg: localhost:5555 repo: anyscale/customer_image url: http://localhost:5555/v2/anyscale/customer_image/blobs/uploads/
[INFO] 1/24/2023, 2:18:26 AM: * Started pushing layer sha256:6d076eb12568fd23ca48a7a31640a2e2e40310f1981d2f1f026906ab687bc1f9
[INFO] 1/24/2023, 2:18:26 AM: Client TLS is disabled
[INFO] 1/24/2023, 2:18:27 AM: Collecting gymnasium[atari,mujoco]==0.26.3
[INFO] 1/24/2023, 2:18:27 AM:   Downloading Gymnasium-0.26.3-py3-none-any.whl (836 kB)

kouroshHakha

pip install —no-deps recsim
pip install gin-config
pip install git+https://github.com/google/dopamine --no-deps --user

kouroshHakha

import math
from black import main
import numpy as np
import pandas as pd

import ray
from ray.data.preprocessors import MinMaxScaler, MultiHotEncoder, OrdinalEncoder
from ray.rllib.policy.sample_batch import SampleBatch

ray.init(ignore_reinit_error=True)

kouroshHakha

from ray import tune
import ray
from ray.rllib.algorithms.ppo import PPO, PPOConfig
from ray.rllib.algorithms.callbacks import DefaultCallbacks
from ray.rllib.models.torch.torch_modelv2 import TorchModelV2
from ray.rllib.utils.typing import TensorType

import torch
import torch.nn as nn

kouroshHakha

import gym
import math
import torch
from gym import spaces, logger
import numpy as np

class TorchBox(spaces.Box):

    def __init__(self, *args, **kwargs):
        device = kwargs.pop('device')

kouroshHakha

"""For an example on how to share observations between agents in MARL look at examples/centralized_critic_2.py
This gist, gives an example of when we want to share neural network weights between agents.
"""
from typing import Tuple

import ray
from ray.rllib.env.multi_agent_env import MultiAgentEnv, MultiAgentDict
from ray.rllib.algorithms.ppo import PPOConfig
from ray.rllib.models.torch.torch_modelv2 import TorchModelV2
import torch.nn as nn

kouroshHakha

FROM nvidia/cuda:11.5.1-cudnn8-devel-ubuntu18.04

# install anaconda3
RUN apt-get update && apt-get install -y \
  curl \
  wget \
  git \
  vim \
  virtualenv

kouroshHakha

To sucessfully render mujoco on headless machines make sure the following dependencies are installed via apt-get:

sudo apt-get --assume-yes install libxrender1 libgl1-mesa-dev libgl1-mesa-glx libglew-dev libosmesa6-dev xpra patchelf libglfw3-dev libglfw3 libglew2.0

Install mujoco-py with headless gpu rendering

You may perform additional steps to build mujoco-py with gpu rendering support on a remote server without any monitor attached.

kouroshHakha

import doodad
from doodad.wrappers.sweeper import launcher


local_mnt = doodad.mount.MountLocal(
    local_dir='data_in',
    mount_point='./data_in',
    output=False
)