High Performance Parallel Reinforcement Learning Implementation in C++

Dependencies

We manage all the dependencies using Anaconda for the most simplicity

• nlohmann_json
• spdlog
• fmt
• curl
• pybind11
• flask

conda install nlohmann_json spdlog fmt curl pybind11 flask -c conda-forge

• Install Pytorch from here

If your Pytorch is installed with GPU support, you also need to install the following packages

• cudatoolkit-dev
• CuDNN

conda install cudnn cudatoolkit-dev==${CUDA_VERSION} -c conda-forge

• gym-remote

cd envs
pip install -e . 

Make sure the CUDA_VERSION matches the one you installed your Pytorch

Build

Before building, make sure your Python is pointing to the Anaconda environment using

which python

Then, execute

mkdir build;
cd build;
cmake ..
make -j 8

To compile with FPGA support, use

cmake .. -DFPGA=ON

Running