Project Overview

The code is based on the code available from the paper "Teacher-Student Curriculum Learning" (https://arxiv.org/abs/1707.00183).

This is a fork made to run in Google Colab

Environment

Google Colab on Ubuntu 18.04 Python 3.6

Limitations on Google Colab

• Docker is not available on Google Colab at this moment
• Minecraft Java requires an OpenGL context to run
• Only one cell active at a time

Workarounds

Docker

There is no workaround for docker on Google Colab at the moment. Will have to run natively.

OpenGL context

We need to start a virtual display. Thankfully, there are a few methods to do this. The python native approach does not work, because for one, it needs a display manager active, but having no display means we don't have a display manager. Even with a display manager active, it would not work as it would only create a virtual display, but not attached to a valid OpenGL context (The GPU).

The easiest approach is to run a virtual display using xvfb (Or X, but xvfb is somewhat simpler). You can create a virtual display with the following command:

Xvfb :0 -screen 0 1920x1080x16

This means create screen 0 with 16bit colour and 1920x1080 pixels and export it to display nr. :0.

One active cell at a time

This is not a limitation of Google Colab, but of Python Notebook in general. It's not really a limitation, but more of a design choice.

Full tutorial

Optional - Mount Google Drive

from google.colab import drive
drive.mount('/content/gdrive')

You will need to click the link and paste the code in the box that appears in the output terminal

Install dependencies

Here, we do not need to install cuda + libcudnn since google colab comes with them pre-configured. (For other versions of Malmo, you may need to change the CUDA+libcudnn version. Check your GPU, CUDA version and libcudnn version.)

!sudo apt-get install libboost-all-dev -y
!sudo apt-get install libxerces-c3.2 -y
!sudo apt-get install python3-pip -y
!pip3 install git+https://github.com/tambetm/minecraft-py.git
!pip3 install opencv-contrib-python   
!pip3 install gym
!pip3 install pygame
!pip3 install git+https://github.com/tambetm/gym-minecraft
!pip3 install tensorflow-gpu===1.15.3
!pip3 install keras===2.2.5
!pip3 install h5py
!sudo apt-get install libboost-all-dev libpython3.6 openjdk-8-jdk ffmpeg python-tk python-imaging-tk
!sudo update-ca-certificates -f
!git clone https://github.com/tambetm/minecraft-py.git
!cd minecraft-py && pip3 install -e .
!pip3 install future
!pip3 install pillow

Install malmo

!wget --no-check-certificate --content-disposition https://github.com/microsoft/malmo/releases/download/0.37.0/Malmo-0.37.0-Linux-Ubuntu-18.04-64bit_withBoost_Python3.6.zip
# --no-check-cerftificate was necessary for me to have wget not puke about https
!curl -LJO https://github.com/microsoft/malmo/releases/download/0.37.0/Malmo-0.37.0-Linux-Ubuntu-18.04-64bit_withBoost_Python3.6.zip
!unzip ./Malmo-0.37.0-Linux-Ubuntu-18.04-64bit_withBoost_Python3.6.zip -d ./Malmo-install
!sudo update-ca-certificates -f
!export MALMO_XSD_PATH=./Malmo-install/Malmo-0.37.0-Linux-Ubuntu-18.04-64bit_withBoost_Python3.6/Schemas
!source ~/.bashrc

Change to jdk8. Current version of Malmo does not work with JDK11 due to the version number being written differently. eg. OpenJDK 8.0 vs JDK version-11.0 so Malmo can not find a valid JDK.

!sudo apt install openjdk-8-jdk-headless openjdk-8-jdk
%env JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64
!java -version

Remove any previous branch of this project & clone it (If it's already present it will crash)

!sudo rm -rf ./TSCL/
!git clone https://github.com/ChrisAB/TSCL.git

Install dependencies for running the virtual display

!sudo apt-get install -y xvfb x11-xserver-utils 

Start up virtual display. Since starting a virtual display stops if the terminal/cell is 'closed', we need to run it as a subprocess within python.

import subprocess
subprocess.Popen(["Xvfb",":0","-screen", "0", "1920x1080x16"])

Set display 0 as the default display and allow anyone within the machine to access it.

%env DISPLAY=:0
!xhost +

Run the minecraft clients. We need to run multiple clients. Also, the instances stop when cell stops, just like Xvfb, need to run it with subprocess

import subprocess
subprocess.Popen(["./TSCL/minecraft/create_minecraft_envs.sh","2"])

The java clients take some time to boot up and there is a sleep for 5 minutes for each client so they don't boot up at the same time. This would cause them to read from the same file/stream causing inconsistent errors. (Some clients start, sometimes its the first, sometimes second, sometimes both but crashes later on, sometimes neither starts) THis could maybe be improved.

!sleep 10m

Start training.

!cd ./TSCL/minecraft && python ./run_minecraft.py --num_runners 2 train basic_0

Project in Action

Directories addition and minecraft contain code for the respective experiments. See readme.txt in each directory for instructions.