SkipPool: Improved Sparse Hierarchical Graph Pooling with Differentiable Exploration

Folder Descriptions

• classification - this folder contains code required to reproduce experiments in the section 4.1
  • Full - this subfolder contains code to reproduce experiments on SkipPool-Full
  • Fixed - this subfolder contains code to reproduce experiments on standard SkipPool
• autoencoder - this folder contains code to reproduce the experiment in the section 4.2

Training models from scratch

Standard SkipPool

• Go to classification/Fixed/model.
• For example run the following to train on DD:

python main.py --dataset --nhid 128 --batch_size 64 --lr 1e-3 --weight_decay 1e-4 --dropout_ratio 0.4

• Data required for this experiment should be available in classification/data directory.
• The trained models are inside [dataset_name]/rep_0/fold_[fold no.], accuracy for each fold is logged in file folds_res_log.txt.
• Average accuracy for the rep 0 is logged in rep_res_log.txt.

SkipPool-Full

• Go to classification/Full/model.
• For example run the following to train on DD:

python main.py --dataset DD --fixStride 5 --batch_size 128 --nhid 64 --lr 1e-4 --conv gcn

• Data required for this experiment should be available in classification/data dir.
• The trained models are inside [dataset_name]/fold_[fold no.], accuracy for each fold is logged in file folds_res_log.txt

For both SkipPool & SkipPool-Full you can specify the number process for each pooling layer by a list. use --p-list command.

• For example run the following to have 2,2 number of processes in respective pooling layers:

python main.py --dataset DD --fixStride 5 --batch_size 128 --nhid 64 --lr 1e-4 --conv gcn --p-list 2,2

Training GAE

• Go to autoencoder directory.
• If you want to train with SkipPool-Full then use --full command. Otherwise standard SkipPool is used.

Example:

python run_ae.py --dataset [graph name] --full --fixStride 5
python run_ae.py --dataset Ring --full --fixStride 5

• MODELNET40 from which Airplane, Car, Guitar, Person are obtained, is not readily available. If one of these Graphs are needed they it will be downloaded.
• Data used to train GAE are at autoencoder/data
• Training models will be at [method]/[graph_name]/rep_0, method is either Full or Fixed.

Dependencies

• Pytorch (2.0.1)
• Pytorch_Scatter (2.1.1)
• torch_sparse (0.6.17)
• Pytorch_Geometric (2.3.1)
• PyGSP (0.5.1)

Hyperparameters to reproduce reported scores in the paper

SkipPool

Dataset	Hidden Dimension	Batch Size	Learning Rate	Weight Decay	Dropout
DD	128	64	1e-3	1e-4	0.4
PROTEINS	64	64	1e-3	1e-4	0.5
NCI1	128	32	5e-4	1e-4	0.0
NCI109	128	32	5e-4	1e-4	0.2
FRANKENSTEIN	32	32	1e-3	1e-4	0.0

SkipPool-Full

Dataset	Hidden Dimension	Batch Size	Learning Rate
DD	64	128	1e-4
REDDIT-B	128	32	1e-3