For best results, BabyVW currently uses Unity 2020.2.1.
This repository depends on VoxSim: https://www.github.com/VoxML/VoxSim. Clone BabyVW, then:
$ git submodule init
$ git submodule update --remote --merge
Finally, follow the depedency installation instructions at the VoxSim link.
The learning environment is based on Unity ML-Agents. The Unity ML-Agents package is included in manifest.json and should be available when this repository is cloned. To complete the setup of the learning environment, you should:
$ conda create --name ml-agents python=3.8
$ conda activate ml-agents
(ml-agents) $ conda install grpcio
(ml-agents) $ pip install numpy
(ml-agents) $ pip install stable_baselines3
(ml-agents) $ python -m pip install mlagents
(ml-agents) $ pip install notebook
(ml-agents) $ pip install sklearn
The BabyVW learning environment uses Stable-Baselines3, which is written using PyTorch, but some minor functionality, such as converting SB3 TensorBoard plots to PyPlot plots, requires TensorFlow. To use this functionality on:
(ml-agents) $ conda install -c apple tensorflow-deps
(ml-agents) $ pip install tensorflow-macos
(ml-agents) $ pip install tensorflow-metal
(ml-agents) $ pip install tensorflow
The main training and testing pipeline should work fine without these dependencies.
You can create the entire environment (including optional dependencies) in one command by running:
- (On M1 Macs)
conda create --name ml-agents --file ml-agents-conda-mac-m1.txt. - (On Intel Macs)
conda create --name ml-agents --file ml-agents-conda-mac-intel.txt. - (On Windows)
conda create --name ml-agents --file ml-agents-conda-win.txt.
Next, check out an environment and agent, e.g., the stacker