Steps for training/running the models

Setting up the environment

• Modify the prefix in environment.yaml to point to the location where the environment will be stored.

• Create a conda environment using

  conda env create -f environment.yaml

• Activate the environment

  conda activate subg_matching

Training a model

• Navigate to scripts/. You are supposed to run the run_large_dataset.sh or run_new_dataset.sh scripts.
• Set the gpus variable to indicate a tuple of all the GPU indices available for the experiment. If just GPU 2 is available, set it to (2). Any non-zero length for the list works.
• Run bash run_large_dataset.sh on the command line. This will start training the model. The model evaluates on the test dataset at the end of training by default.
• Note that in our original codebase, we have used wandb to manage and monitor runs. However, we have set WANDB_MODE=disabled in the bash script since we don't expect every user to be familiar with wandb. In case the user has experience using it, the WANDB_MODE=disabled part of the command can be deleted, so that it starts as such - CUDA_VISIBLE_DEVICES=...
• Results will be stored in the <experiment_dir>/<experiment_id> directory, which in this case is experiments/rqX_custom_models. This includes trained models, partial configs and logs. The train/validation scores are printed at every epoch in the corresponding log file, and the test score is evaluated at the end of training.

Additional files

• We provide additional files here - https://rebrand.ly/dessub.

Model names

• Some models have a different naming convention in the codebase than in the paper.

• Models discussed in the main text

  1. Our-Early-Best - configs/edge_models/scoring=sinkhorn_pp=hinge___tp=sinkhorn_pp=hinge_when=post___unify=true.yaml

  2. Our-Late-Best - configs/edge_models/scoring=sinkhorn_pp=hinge___tp=none.yaml

  3. GMN adaptation - configs/node_models/scoring=attention_pp=identity___tp=attention_pp=identity_when=post.yaml

  4. IsoNet adapation - configs/edge_models/scoring=sinkhorn_pp=lrl___tp=none.yaml

Other details

• AIDS, MUTAG, PTC-FM, PTC-FR, PTC-MM and PTC-MR are under the directory large_dataset while NCI-H23H, MOLT-4H, MCF-7H and MSRC-21 come under new_dataset.

• Training GraphSim for new datasets requires the conv_pool_size: [4,4,3,3] line to be active while for large datasets, conv_pool_size: [3,3,2,2] should be chosen.