Edge Rewiring Goes Neural: Boosting Network Resilience via Policy Gradient
This repository is the official PyTorch implementation of "Edge Rewiring Goes Neural: Boosting Network Resilience via Policy Gradient".
Shanchao Yang, Kaili Ma, Baoxiang Wang, Hongyuan Zha, Edge Rewiring Goes Neural: Boosting Network Resilience via Policy Gradient
Installation
-
CUDA 11.+
-
Create Python environment (3.+), using
anaconda
is recommended:conda create -n my-resinet-env python=3.8 conda activate my-resinet-env
-
Install Pytorch using
anaconda
conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia
or using
Pip
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html
-
Install networkx, tensorflow, tensorboardX, numpy, numba, dm-tree, gym, dgl, pyg
pip install networkx==2.5 pip install tensorflow-gpu==2.3.0 pip install numpy==1.20.3 pip install numba==0.52.0 pip install gym==0.18.0 pip install tabulate pip install dm-tree pip install lz4 pip install opencv-python pip install tensorboardX pip install dgl-cu111 -f https://data.dgl.ai/wheels/repo.html pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.9.0+cu111.html pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.9.0+cu111.html pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.9.0+cu111.html pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.9.0+cu111.html pip install torch-geometric
-
Install ray
-
Use the specific commit version of ray
8a066474d44110f6fddd16618351fe6317dd7e03
For Linux:
pip install -U https://s3-us-west-2.amazonaws.com/ray-wheels/master/8a066474d44110f6fddd16618351fe6317dd7e03/ray-2.0.0.dev0-cp38-cp38-manylinux2014_x86_64.whl
For Windows:
pip install -U https://s3-us-west-2.amazonaws.com/ray-wheels/master/8a066474d44110f6fddd16618351fe6317dd7e03/ray-2.0.0.dev0-cp38-cp38-win_amd64.whl
-
Download our repository, which includes the source codes of ray and ResiNet.
git clone https://github.com/yangysc/ResiNet.git
-
Set the symlink of rllib to use our custom rllib (remeber to remove these symlinks before uninstalling ray!)
python ResiNet/ray-master/python/ray/setup-dev.py -y
-
Code description
There are 4 important file folders.
-
Environment:
ResiNet/ray-master/rllib/examples/env/
-
graphenv.py
is the edge rewiring environment based on OpenAI gym. -
parametric_actions_graph.py
is the env wrapper that accesses the graph fromgraphenv.py
and returns thedict
observation. -
utils_.py
defines the reward calculation strategy. -
get_mask.py
defines the action mask calculation for selecting the first edge and the second edge. -
datasets
is the folder for providing training and test datasets. The following table (Table 2, Page 17 in the paper) records the statistics of graphs used in the paper.Dataset Node Edge Action Space Size BA-15 15 54 5832 BA-50 50 192 73728 BA-100 100 392 307328 EU 217 640 819200 BA-10-30 () 10-30 112 25088 BA-20-200 () 20-200 792 1254528
-
-
Model:
ResiNet/ray-master/rllib/examples/models/
autoregressive_action_model.py
is the network architecture of ResiNet.gnnmodel.py
defines theGIN
model based ondgl
.
-
Distribution:
ResiNet/ray-master/rllib/examples/models/
autoregressive_action_dist.py
is the action distribution module of ResiNet.
-
Loss:
ResiNet/ray-master/rllib/agents/ppo/
ppo_torch_policy.py
defines theDDPPO
loss function.
Run
Platform
We tested the following experiments (see Command
) with
- GPU: GEFORCE RTX 3090 * 2 (24 G memory * 2 = 48G in total)
- CPU: AMD 3990X
Adjust the corresponding hyperparameters according to your GPU hardware. Our code supports the multiple gpus training thanks to ray
. The GPU memory capacity and the number of gpu are the main bottlenecks for DDPPO. The usage of more gpus means a faster training.
num-gpus
: the number of GPU available in total (increase it if more gpus are available)bs
: batch sizemini-bs
: minibatch sizetasks-per-gpu
:the number of paralleled workergpus_per_instance
: the number of GPU used for this train instance (ray can support tune multiple instances simultaneously) (increase it if more gpus are available)
Command
First go to the following folder.
cd ResiNet/ray-master/rllib/examples
Train
-
Transductive setting (
dataset
is in[example_15, example_50, example_100, EU]
)-
Run the experiment on optimizing the
BA-15
dataset withalpha
=0, risilience metricR
, node degree-based attack:CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=example_15 --tasks-per-gpu=2 --gpus_per_instance=2 --bs=4096 --mini-bs=256 --filtration_order=-1 --alpha=0 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=ge --seed=0
-
Optimize the
BA-15
dataset with a grid search of thefiltration order
(set to-3
):CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=example_15 --tasks-per-gpu=2 --gpus_per_instance=2 --bs=4096 --mini-bs=256 --filtration_order=-3 --alpha=0 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=ge --seed=0
-
Optimize the
BA-15
dataset with a grid search ofalpha
(the coefficient of weighted sum of resilience and utility) (set to-1
):CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=example_15 --tasks-per-gpu=2 --gpus_per_instance=2 --bs=4096 --mini-bs=256 --filtration_order=-1 --alpha=-1 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=ge --seed=0
-
Optimize the
BA-15
dataset with a grid search ofrobust-measure
(resilience metric, choice is[R, sr, ac]
) (set to-1
):CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=example_15 --tasks-per-gpu=2 --gpus_per_instance=2 --bs=4096 --mini-bs=256 --filtration_order=-1 --alpha=0 --robust-measure=-1 --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=ge --seed=0
-
Optimize the
BA-15
dataset with a grid search ofsecond-obj-func
(utility metric, choice is[ge, le]
) (set to-1
):CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=example_15 --tasks-per-gpu=2 --gpus_per_instance=2 --bs=4096 --mini-bs=256 --filtration_order=-1 --alpha=0 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=-1 --seed=-1
-
Optimize the
BA-15
dataset with a grid search ofseed
(set to-1
):CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=example_15 --tasks-per-gpu=2 --gpus_per_instance=2 --bs=4096 --mini-bs=256 --filtration_order=-1 --alpha=0 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=ge --seed=-1
-
Optimize the
EU
dataset (increasebs
andhidden_dim
if more gpus are available. Four gpus would be better forhidden_dim=64
):CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=EU --tasks-per-gpu=1 --gpus_per_instance=2 --bs=1024 --mini-bs=256 --filtration_order=1 --alpha=0 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=32 --attack_strategy=degree --second-obj-func=ge --seed=0
-
-
Inductive setting (
dataset
is in[ba_small_30, ba_mixed]
)-
for the
ba_small_30
dataset (use full filtration)CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=ba_small_30 --tasks-per-gpu=1 --gpus_per_instance=2 --bs=2048 --mini-bs=256 --filtration_order=-1 --alpha=0 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=ge --seed=0
-
for the
ba_mixed
dataset (setfiltratio_order
to 1,tasks-per-gpu
to 1 andbs
to 2048)CUDA_VISIBLE_DEVICES=0,1 python autoregressivegraph_decouple_action_dist_dppo.py --num-gpus=2 --cwd-path=./ --stop-iters=2000 --stop-timesteps=800000 --dataset=ba_mixed --tasks-per-gpu=1 --gpus_per_instance=2 --bs=2048 --mini-bs=256 --filtration_order=1 --alpha=0 --robust-measure=R --reward_scale=10 --dual_clip_param=10 --lr=7e-4 --vf_lr=7e-4 --ppo_alg=dcppo --hidden_dim=64 --attack_strategy=degree --second-obj-func=ge --seed=0
-
We highly recommend using tensorboard
to monitor the training process. To do this, you may run
tensorboard --logdir log/DDPPO
Set checkpoint_freq
to be non-zero (zero by default) if you want to save the trained models during the training process. And the final trained model will be saved by default when the training is done. All trained models and tensorboard logs are saved in the folder log/DDPPO/
.
Test
- BA-15 (
dataset
is in[example_15, example_50, example_100, EU, ba_small_30, ba_mixed]
) (The problem setting related hyperparameters need to be consistent with the values used in training.)CUDA_VISIBLE_DEVICES=0,1 python evaluate_trained_agent_dppo.py --num-gpus=2 --tasks-per-gpu=1 --bs=400 --mini-bs=16 --gpus_per_instance=1 --ppo_alg=dcppo --attack_strategy=degree --second-obj-func=le --seed=0 --reward_scale=1 --test_num=-1 --cwd-path=./test --alpha=0.5 --dataset=example_15 --filtration_order=-1 --robust-measure=ac --hidden_dim=64
restore_path
inevaluate_trained_agent_dppo.py
(Line 26) to the trained model folder.