The datasets are hypergraphs with categorical hyper edges. They are provided here by Arthur R. Benson.
The following shows the command to perform training on any of the datasets present in ./datasets.
python train.py -e [experiment_name] --dataset [Datasetname] --num_classes [Number of classes] --batch_size [ ] --iter [ number of samples ] --max_dim [ maximum dimension of simplex ] --dim [ particular simplex dimension to infere on] --gpu [GPU device number] --disable_cuda [Boolean] --reset_model [Boolean]
For example :
python train.py -e default_experiment --dataset cat_edge_cooking --num_classes 20 --iter 10000 --batch_size 32
The above will:
- generate 10000 positive and negative samples
- initialize Graph Attention based and Simplicial Message Passing based (our approach) models
- train them for 1 epoch i.e. 10000 iterations
- with batch size of 32
- save models in
./datasets/cat_edge_cooking/models/* - save log file as
./datasets/cat_edge_cooking/log_train.txt - generate tensorboad log files in
./datasets/cat_edge_cooking/logs/*(Loss curves) - generate
./datasets/cat_edge_cooking/default_experiment_scores.csv
The following with test pretrained models generated from train.py.
python test.py -e [experiment_name] --dataset [Datasetname] --num_classes [Number of classes] --batch_size [ ] --iter [ number of samples ] --max_dim [ maximum dimension of simplex ] --dim [ particular simplex dimension to infere on] --gpu [GPU device number] --disable_cuda [Boolean]
For example:
python test.py -e default_experiment --dataset cat_edge_cooking --num_classes 20 --iter 10000 --batch_size 32
The above will:
- generate 10000 positive and negative samples
- load pretrained models from
./datasets/cat_edge_cooking/models/* - generate AUC and AUC_PR for 4 settings (Union, Intersection, GAT, Simplicial MPSN)
- save log file as
./datasets/cat_edge_cooking/log_test.txtwith the results. - store results in
./datasets/cat_edge_cooking/default_experiment_scores.csv
The scripts used for experiments are available in ./scripts/*.