ISONET++

Iteratively Refined Early Interaction Alignment for Subgraph Matching based Graph Retrieval (NeruIPS24)

This directory contains code necessary for running all the experiments.

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 einsmatch

Training a model

• Navigate to scripts/. You are supposed to run the run_custom.sh script.
• 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_custom.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.
• Note that the experiments directory in this folder is intentionally kept empty. It will be the home for logs and configs of any new training runs that are started by the user.

Additional files

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

Model names

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

• Key models

  1. Multi-round EinsMatch (Node) - configs\rq4_iterative\iterative___scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=pre___unify=true.yaml

  2. Multi-layer EinsMatch (Node) - configs\rq4_baselines\scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=pre___unify=true.yaml

  3. Multi-round EinsMatch (Edge) - configs\edge_early_variants\edge_early_interaction.yaml

  4. Multi-layer EinsMatch (Edge) - configs\edge_early_variants\edge_early_interaction_baseline.yaml

  5. GMN-Match - configs\rq4_baselines\scoring=agg___tp=attention_pp=identity_when=post.yaml

• Ablations

  1. Multi-round Node pair partner (msg-only) - configs\rq8\iterative___scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=msg_passing_only___unify=true.yaml

  2. Multi-layer Node pair partner (msg-only) - configs\rq8\scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=msg_passing_only___unify=true.yaml

  3. Multi-round Node partner (with additional MLP) - configs\rq8\iterative___scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=update_only___unify=true.yaml

  4. Multi-layer Node partner (with additional MLP) - configs\rq8\scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=update_only___unify=true.yaml

  5. Multi-round Node partner - configs\rq4_iterative\iterative___scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=post___unify=true.yaml

  6. Multi-layer Node partner - configs\rq4_baselines\scoring=sinkhorn_pp=lrl___tp=sinkhorn_pp=lrl_when=post___unify=true.yaml

• Configs to perform ablations on the (T,K) parameters for multi-round and (K) parameter for multi-layer can be found in configs\rq7_efficiency.

Reference

If you find the code useful, please cite our paper:

@inproceedings{ramachandraniteratively,
  title={Iteratively Refined Early Interaction Alignment for Subgraph Matching based Graph Retrieval},
  author={Ramachandran, Ashwin and Raj, Vaibhav and Roy, Indradyumna and Chakrabarti, Soumen and De, Abir},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}
}

Indradyumna Roy, Indian Institute of Technology - Bombay
indraroy15@cse.iitb.ac.in