This is the source code for paper: Disentangled Graph Prompting for Out-Of-Distribution Detection.

Requirements:

• torch-geometric==2.0.4
• torch-scatter==2.0.93
• torch-sparse==0.6.15
• numpy==1.21.2
• pandas==1.3.0
• python==3.9.15
• scikit-learn=1.0.2
• scipy==1.9.3
• torch==1.11.0
• torchvision==0.12.0

Training:

Run DGP-GCL:

python DGP_GCL.py --DS xxxx --model_type dgp-gcl --lr xxxx --aug xxxx --DS_pair xxxx --lambda_ xxxx --gamma xxxx --alpha_1 xxxx --alpha_2 xxxx --dgp_lr xxxx

Run DGP-Sim:

python DGP_Sim.py --DS xxxx --model_type dgp-sim  --lr xxxx --eta xxxx --DS_pair xxxx --lambda_ xxxx --gamma xxxx --alpha_1 xxxx --alpha_2 xxxx --dgp_lr xxxx

To run pre-trained GNNs or their fine-tuned versions, simply modify the --model_type parameter.

We also provide the code to search hyper-parameters, you can use the following command (for TOX21-SIDER dataset) to run it:

bash run_grid_search.sh