XRLFLOW is a GNN based neural network optimisation extension of TASO. The basic idea was to substitute TASO's cost-based backtracking search and replace it with a RL-based optimisation. Potentially, this should be able to generalise to large unseen graphs and find better performing solutions than the backtracking search.
XRLFLOW consists of three main parts:
- An extension of the TASO library exposing a low-level RL environment
- A gym-style high-level environment written in Python, interacting with the XRLFLOW environment
- A PPO agent interacting with this environment, iteratively applying graph substitutions to a target graph.
XRLFLOW interacts with TASO and thus depends on the TASO library. TASO on the other hand depends on a working CUDA and CuDNN installation.
The experiments used Python 3.9, though later versions might also work. It is strongly advised to use a virtual environment such as pyenv or venv for installation.
At the time of the initial experiments, we used Cuda 10.2 and CuDNN 7.6.5.32-1.
It is recommanded to use a nvidia docker
Build taso in the "taso-build" directory, follow its instructions
Most importantly, set the TASO_HOME variable:
export TASO_HOME=/path/to/tasoAlso, for some builds to succeed I had to set the LD_LIBRARY_PATH variable:
export LD_LIRBARY_PATH=/usr/local/libThen follow TASO's installation page.
Note that, check installation by 'import taso' after built.
The installation of XRLFLOW is very similar to that of TASO. First, make sure the environment variables are set:
export TASO_HOME=/path/to/taso
export LD_LIRBARY_PATH=/usr/local/libThen, go to taso_ext, create a build folder, and run cmake:
cd taso_ext
mkdir build
cd build
cmake ..
make -j
sudo make installAt this path, you should now have two .so files:
ls -alp /usr/local/lib/
# ...
# -rw-r--r-- 1 root root 448920 Apr 7 13:16 libtaso_rl.so
# -rw-r--r-- 1 root root 1173704 Apr 7 13:39 libtaso_runtime.so
# ...
The extension has been built. Now the python package can be installed:
# In the project root:
pip install -e python/