GFlowNet-CombOpt

Pytorch implementation for our paper

Let the Flows Tell: Solving Graph Combinatorial Optimization Problems with GFlowNets.

Dinghuai Zhang, Hanjun Dai, Nikolay Malkin, Aaron Courville, Yoshua Bengio, Ling Pan.

We formulate a set of graph combinatorial optimization problems as sequential decision-making sampling problems, and design efficient GFlowNet algorithms to tackle them.

For ME

GPU

dgl                       1.1.2+cu116              pypi_0    pypi
torch                     1.13.0+cu116             pypi_0    pypi
NVIDIA-SMI 511.79       Driver Version: 511.79       CUDA Version: 11.6

OR

ubuntu20.04-cuda11.3.0-py38-torch1.11.0-tf1.15.5-1.8.1

https://pypi.tuna.tsinghua.edu.cn/simple/dgl/

pip3 install networkx==2.5.0 pydot hydra-core==1.1.0 omegaconf fairseq submitit hydra-submitit-launcher
pip3 uninstall dgl -y
pip3 install dgl -f https://pypi.tuna.tsinghua.edu.cn/packages/e4/9c/ed51e0f42c3910b4af7f75ddd24d52668314e8826ec4286d48562e29503a/dgl-1.1.1-cp311-cp311-manylinux1_x86_64.whl
pip3 install ipdb einops

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
fairseq 0.12.2 requires hydra-core<1.1,>=1.0.7, but you have hydra-core 1.1.0 which is incompatible.
fairseq 0.12.2 requires omegaconf<2.1, but you have omegaconf 2.1.2 which is incompatible.

cd GFlowNet-CombOpt/
cd data/
python3 rbgraph_generator.py --num_graph 4000 --save_dir rb200-300/train
python3 rbgraph_generator.py --num_graph 500 --save_dir rb200-300/test
cd ..
cd gflownet/
python3 main.py input=rb200-300 alg=fl bsit=8 > log

Dependency

pip install hydra-core==1.1.0 omegaconf submitit hydra-submitit-launcher
pip install dgl==0.6.1

Data generation

cd data/
python rbgraph_generator.py --num_graph 4000 --save_dir rb200-300/train
python rbgraph_generator.py --num_graph 500 --save_dir rb200-300/test  

Training

cd gflownet/
python main.py input=rb200-300 alg=fl bsit=8