Skip to content

Latest commit

 

History

History
105 lines (82 loc) · 5.01 KB

surreal_subproc.md

File metadata and controls

105 lines (82 loc) · 5.01 KB
Raw

Install and Run SURREAL Locally

This guide will allow you to setup Surreal on your local machine. Multiple processes are launched by the surreal-subproc commandline interface.

Requirements
Install Surreal
Create and run an experiment
Develop Algorithms Locally

Requirements

  • This guide is written and tested primarily on Mac OS X and Ubuntu 16.04. If you run into issues when installing, you can check our docker file (adapted from the image for mujoco_py). For linux users, the dependencies that we needed on top of a nvidia image is listed as follows.
sudo apt-get update
sudo apt-get install \
    curl \
    git \
    cmake \
    unzip \
    bzip2 \
    wget \
    libgl1-mesa-dev \
    libgl1-mesa-glx \
    libglew-dev \
    libosmesa6-dev \
    software-properties-common \
    net-tools \
    unzip \
    vim \
    virtualenv \
    wget \
    xpra \
    xserver-xorg-dev \
  • Mujoco is the physics simulator for our environments. First download MuJoCo 1.5.0 (Linux and Mac OS X) and place the mjpro150 folder and your license key mjkey.txt in ~/.mujoco. You can obtain a license key from here.
    • For Linux, you will need to install some packages to build mujoco-py (sourced from here, with a couple missing packages added). If using apt, the required installation command is: 
      $ sudo apt install libgl1-mesa-dev libgl1-mesa-glx libglew-dev \
        libosmesa6-dev software-properties-common net-tools wget \
        xpra xserver-xorg-dev libglfw3-dev patchelf

Install Surreal

  1. (Optional) Setup python environment. First, create a python environment. Surreal is developed on python 3.5+. We recommend using conda or a virtual environment for the dependencies of Surreal. For example, we usually use conda.
conda create -n surreal python>=3.5
source activate surreal

  1. Install pytorch. Install pytorch following the official guide. Optionally, setup cuda if you have GPUs enabled.

  2. Install Surreal. Installing surreal through pip would install the surreal library and all its dependencies.

pip install surreal
  1. Configure Surreal. Setup .surreal.yml. Run the following command to setup the surreal config file at ~/.surreal.yml.
surreal-default-config

This will generate a config file at ~/.surreal.yml. Optionally, you can put it in another location and specify SURREAL_CONFIG_PATH.

Experiments are automatically prepended with your username. Specify it in the config. You can turn this behavior off by setting prefix_experiment_with_username = False.

username: <your_username>

Every time an experiment is created using surreal-subproc, result data will be written to <subproc_results_folder>/<experiment_name>. subproc_results_folder is specified in the config. (e.g. You can put all your experiment results in ~/surreal/subproc/)

subproc_results_folder: <put path here> # e.g. ~/surreal/subproc/

If you want to know more about the config and other fields, refer to this guide. For now, we have what we need to setup local experiments.

Launch an Experiment

You are now ready to launch an experiment. Run

surreal-subproc <experiment_name> --num-agents 4

If you have your own virtualenv, you can activate it before running subproc:

source activate mypythonenv
surreal-subproc <experiment_name> --num-agents 4

If you setup your .surreal.yml as default, you need to make sure the subproc_results_folder field is properly set. You will see experiment outputs in ~/<your_subproc_folder>/experiment_name and see tensorboard output at localhost:6006. You can choose one of the two pre-installed surreal algorithms by using the --algorithm flag.

surreal-subproc --algorithm ppo <experiment_name> # Runs Surreal-PPO
surreal-subproc --algorithm ddpg <experiment_name> # Runs Surreal-DDPG

Consuming GPU. surreal-subproc respects CUDA_VISIBLE_DEVICES for using GPU acceleration for rendering and neural network acceleration. You can control what GPUs are used with

CUDA_VISIBLE_DEVICES=0,1 surreal-subproc <experiment_name> --num-agents 4

When there is only one GPU present, the launcher will assign agents, evals and learner to this GPU. When there is more than one. The launcher will assign learner to one GPU and evenly distribute agents and evals to the remaining GPUs.

Use Ctrl-C to stop a running distributed experiment. surreal-subproc has mechanism to capture the interruption signal and terminate all child processes gracefully.