NOTE. A lightweight version of the code that removes the GPU dependency is available at this repository, which makes it easier to run the code in case you don't have a GPU. Also useful if you just want to try out the algorithm.
This is the code for the article Goal-directed graph construction using reinforcement learning by Victor-Alexandru Darvariu, Stephen Hailes and Mirco Musolesi, Proc. R. Soc. A. 477:20210168. If you use this code, please consider citing:
@article{darvariu2021goal,
author = {Darvariu, Victor-Alexandru and Hailes, Stephen and Musolesi, Mirco},
title = {Goal-directed graph construction using reinforcement learning},
journal = {Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences},
year = 2021,
month = {oct},
publisher = {The Royal Society Publishing},
volume = {477},
number = {2254},
pages={20210168},
doi = {10.1098/rspa.2021.0168},
url = {https://doi.org/10.1098/rspa.2021.0168},
}
MIT.
Please ensure that you clone this repository under the relnet root directory, e.g.
git clone git@github.com:VictorDarvariu/graph-construction-rl.git relnet
Currently tested on Linux and MacOS (specifically, CentOS 7.4.1708 and Mac OS Big Sur 11.2.3), can also be adapted to Windows through WSL. The host machine requires NVIDIA CUDA toolkit version 9.0 or above (tested with NVIDIA driver version 384.81). Makes heavy use of Docker, see e.g. here for how to install on CentOS. Tested with Docker 19.03. The use of Docker largely does away with dependency and setup headaches, making it significantly easier to reproduce the reported results.
The Docker setup uses Unix groups to control permissions. You can reuse an existing group that you are a member of, or create a new group groupadd -g GID GNAME and add your user to it usermod -a -G GNAME MYUSERNAME.
Create a file relnet.env at the root of the project (see relnet_example.env) and adjust the paths within: this is where some data generated by the container will be stored. Also specify the group ID and name created / selected above.
Add the following lines to your .bashrc, replacing /home/john/git/relnet with the path where the repository is cloned.
export RN_SOURCE_DIR='/home/john/git/relnet'
set -a
. $RN_SOURCE_DIR/relnet.env
set +a
export PATH=$PATH:$RN_SOURCE_DIR/scriptsMake the scripts executable (e.g. chmod u+x scripts/*) the first time after cloning the repository, and run apply_permissions.sh in order to create and permission the necessary directories.
Some scripts are provided for convenience. To build the containers (note, this will take a significant amount of time e.g. 2 hours, as some packages are built from source):
update_container.shTo start them:
manage_container_gpu.sh up
manage_container.sh upTo stop them:
manage_container_gpu.sh stop
manage_container.sh stopTo purge the queue and restart the containers (useful for killing tasks that were launched):
purge_and_restart.shTo take maximum advantage of your machine's capacity, you may want to tweak the number of threads for the GPU and CPU workers. This configuration is provided in projectconfig.py.
Additionally, you may want to enforce certain memory limits for your workers to avoid OOM errors. This can be tweaked in docker-compose.yml and manage_container_gpu.sh.
It is also relatively straightforward to add more workers from different machines you control. For this, you will need to mount the volumes on networked-attached storage (i.e., make sure paths provided in relnet.env are network-accessible) and adjust the location of backend and queue in projectconfig.py to a network location instead of localhost. On the other machines, only start the worker container (see e.g. manage_container.sh).
The real-world network data can be downloaded from Figshare at this URL.
- Copy the
rw_network_data.zipfile provided to$RN_EXPERIMENT_DATA_DIR/real_world_graphs/raw_data. - Then, unzip it
unzip rw_network_data.zip - Re-grant group permissions so they can be read/modified by container users:
chgrp -R $RN_GNAME $RN_EXPERIMENT_DATA_DIR/real_world_graphs/raw_data. - Delete zip file
rm rw_network_data.zip - Run the following commands to wrangle the data into the expected formats:
docker exec -it relnet-worker-cpu /bin/bash -c "python relnet/data_wrangling/process_networks.py --dataset scigrid --task clean"
docker exec -it relnet-worker-cpu /bin/bash -c "python relnet/data_wrangling/process_networks.py --dataset scigrid --task process"
docker exec -it relnet-worker-cpu /bin/bash -c "python relnet/data_wrangling/process_networks.py --dataset euroroad --task clean"
docker exec -it relnet-worker-cpu /bin/bash -c "python relnet/data_wrangling/process_networks.py --dataset euroroad --task process"
Synthetic data will be automatically generated when the experiments are ran and stored to $RN_EXPERIMENT_DIR/stored_graphs.
There are several services running on the manager node.
- Jupyter notebook server:
http://localhost:8888 - Flower for queue statistics:
http://localhost:5555 - Tensorboard (currently disabled due to its large memory footprint):
http://localhost:6006 - RabbitMQ management:
http://localhost:15672
The first time Jupyter is accessed it will prompt for a token to enable password configuration, it can be grabbed by running docker exec -it relnet-manager /bin/bash -c "jupyter notebook list".
Experiment data and results are stored in part as files (under your configured $RN_EXPERIMENT_DATA_DIR) as well as in a MongoDB database.
To access the MongoDB database with a GUI, you can use a MongoDB client such as Robo3T and point it to http://localhost:27017.
Some functionality is provided in relnet/evaluation/storage.py to insert and retrieve data, you can use it in e.g. analysis notebooks.
Experiments are launched from the manager container and processed (in a parallel way) by the workers.
The file relnet/evaluation/experiment_conditions.py contains the configuration for the experiments reported in the paper, but you may modify e.g. agents, objective functions, hyperparameters etc. to suit your needs.
Then, you can launch all the experiments as follows:
# synthetic graph experiments
docker exec -d relnet-manager /bin/bash -c "source activate ucfadar-relnet; python run_experiments.py --which synth --experiment_part both --edge_percentage 1 --experiment_id synth_1 --force_insert_details"
docker exec -d relnet-manager /bin/bash -c "source activate ucfadar-relnet; python run_experiments.py --which synth --experiment_part both --edge_percentage 2.5 --experiment_id synth_2_5 --force_insert_details"
docker exec -d relnet-manager /bin/bash -c "source activate ucfadar-relnet; python run_experiments.py --which synth --experiment_part both --edge_percentage 5 --experiment_id synth_5 --force_insert_details"
# real-world graph experiment
docker exec -d relnet-manager /bin/bash -c "source activate ucfadar-relnet; python run_experiments.py --which real_world --experiment_part both --edge_percentage 2.5 --train_individually --experiment_id real_world --force_insert_details"
To check the progress of running experiments, you can use the peek_eval_losses.py tool as follows:
docker exec -it relnet-manager /bin/bash -c "source activate ucfadar-relnet; python relnet/experiment_launchers/peek_val_losses.py --experiment_id synth_1"It will print out the validation losses and step numbers of current training runs.
Jupyter notebooks are used to perform the data analysis and produce tables and figures. Navigate to http://localhost:8888, then notebooks folder.
The relationships between notebooks and tables/figures are as follows:
Evaluation_Models.ipynb: Table 1, Figure 2, Figure 3Evaluation_Models_RealWorld_Individual.ipynb: Table 2
Open the notebook selecting the Python (RelNET) kernel and run all cells. Resulting .pdf figures and .tex tables can be found at $RN_EXPERIMENT_DIR/aggregate.
In case the kernel is not found, try reinstalling the kernel by running docker exec -it -u 0 relnet-manager /bin/bash -c "source activate ucfadar-relnet; python -m ipykernel install --user --name relnet --display-name "Python (RelNET)"
If you face any issues or have any queries feel free to contact v.darvariu@ucl.ac.uk and I will be happy to assist.