Unity ML Agents Installation for Windows

Comments

Source

Here are some notes on how to set up Unity ML Agents under Windows 10. The instructions are consolidated from multiple locations. They've been cleaned up or updated to make sense for the current Unity 2019.30f3 version, the current ml-agents and Windows 10.

These instructions do not use Anaconda but just the "standard" Python from www.python.org. I prefer to use the standard Python together with venv based virtual environments where possible. Anaconda installations together with conda virtual environments could be used just as well but require a lot more disk space and installation time.

These notes follow the instructions here and here but are customized for my conventions:

  • Use folder D:\repos for all git repositories.
  • Use folder D:\repos\zz-python-envs to store venv virtual environments across many projects.
  • Use batch files to activate environments for convenience (no need to remember the weird paths for activation)
  • Extra attention to using a 64-bit Python installer. (I've inadvertently used the 32-bit Python installer before and it resulted in unexplainable errors.)
  • Use "python -m pip install" rather than just "pip install" as the former is safer and is the officially recommended approach.

If you don't have a D: drive, you can perform the same setup under the C: drive instead, or in any other folder.

Python and ML-Agents Toolkit Setup

At the time of writing this, the ML-Agents Toolkit did not support Python 3.8. So we'll use Python 3.7.6 which is the latest at this time.

  • Download the Python 3.7.6 (not 3.8) installer from here. Ensure that it is 64-bit version. This is the direct path for the 64-bit version of Python 3.7.6.

  • Verify the installer's MD5 checksum. For more info on certutil, refer to this article.

D:\Downloads>certutil -hashfile python-3.7.6-amd64.exe MD5
MD5 hash of python-3.7.6-amd64.exe:
cc31a9a497a4ec8a5190edecc5cdd303
CertUtil: -hashfile command completed successfully.

  • Install Python 3.7.6 by running the downloaded executable. Use the default install option but check the box for "Add Python 3.7 to PATH"

  • Verify the installation by running python from a command prompt and making sure the version is right and AMD64 is indicated as shown below.

D:\repos>python
Python 3.7.6 (tags/v3.7.6:43364a7ae0, Dec 19 2019, 00:42:30) [MSC v.1916 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
  • Create a batch file in D:\repos\activatemlagents.bat with lines:
cd D:\repos
zz-python-envs\mlagents\Scripts\activate
  • Set up virtual environment and upgrade pip. Note that we are not installing mlagents from PyPi.
cd D:\repos
md zz-python-envs
python -m venv zz-python-envs\mlagents
activatemlagents
python -m pip install --upgrade pip
  • Clone the ML-Agents Toolkit repository, and switch to the latest stable release.
cd D:\repos
git clone https://github.com/Unity-Technologies/ml-agents.git
cd ml-agents
git checkout latest_release
  • Install mlagents from the cloned repo rather than from PyPi.
cd D:\repos
activatemlagents
cd D:\repos\ml-agents\ml-agents-envs
python -m pip install -e ./
cd ..
cd ml-agents
python -m pip install -e ./
  • Verify installation by opening a new command prompt and executing the following:
cd D:\repos
activatemlagents
mlagents-learn --help

Using ML-Agents Toolkit With Unity

For Unity 2019.30f3 or later, here are the steps to add ML-Agents to a project. These are adapted from here.

  • In Unity 2019.30f3 or later, create a new 3D project or open an existing project.
  • In the Unity project, install the Barracuda package: Use the Window menu, select Package Manager, click Advanced dropdown menu to the left of the search bar and ensure "Show preview packages" is checked. Search for the Barracuda package and install it.
  • From the Windows file system window, navigate to the folder containing your cloned ML-Agents repository, i.e. D:\repos\ml-agents\UnitySDK\Assets
  • Drag the ML-Agents folder from D:\repos\ml-agents\UnitySDK\Assets to the Unity Editor Project window.

Comments

Comments powered by Disqus