Datasets - SOLAR, ENERGY. This code provides a running example with all components on MTGNN model (we acknowledge the authors of the work).
python train_multi_step.py --data ./data/{0} --model_name {1} --device cuda:0 --expid {2} --epochs 100 --batch_size 64 --runs 10 --random_node_idx_split_runs 100 --lower_limit_random_node_selections 15 --upper_limit_random_node_selections 15 --step_size1 {3} --mask_remaining {4}
Here,
{0} - refers to the dataset directory: ./data/{ENERGY/SOLAR}
{1} - refers to the model name
{2} - refers to the manually assigned "ID" of the experiment
{3} - step_size1 is 2500 for SOLAR.
{4} - inference post training in the partial setting, set to true or false. Note - mask_remaining is the alias for "Partial" setting in the original paper
- random_node_idx_split_runs - the number of randomly sampled subsets per trained model run
- lower_limit_random_node_selections and upper_limit_random_node_selections - the percentage of variables in the subset S.
python train_multi_step.py --data ./data/{0} --model_name {1} --device cuda:0 --expid {2} --epochs 100 --batch_size 64 --runs 50 --predefined_S --random_node_idx_split_runs 1 --lower_limit_random_node_selections 100 --upper_limit_random_node_selections 100 --step_size1 {3}
python train_multi_step.py --data ./data/{0} --model_name {1} --device cuda:0 --expid {2} --epochs 100 --batch_size 64 --runs 10 --adj_identity_train_test --random_node_idx_split_runs 100 --lower_limit_random_node_selections 100 --upper_limit_random_node_selections 100 --step_size1 {3}
Download Solar dataset from https://github.com/laiguokun/multivariate-time-series-data. Uncompress them and move them to the data folder.
Download Energy dataset from
# Create data directories
mkdir -p data/{SOLAR,ENERGY}
# for any dataset, run the following command
python generate_training_data.py --ds_name {0} --output_dir data/{1} --dataset_filename data/{2}
Here
{0} is for the dataset: solar, energy
{1} is the directory where to save the train, valid, test splits. These are created from the first command
{2} the raw data filename (the downloaded file), such as - solar.txt.