This is a STID implementation to train over a large sparse dataset.
- Basic "train" script
- Data Processing/Adapter
- Model Saving
- Result Plot
- LossFunction API
- --model, model name, used to save the temporary results/models
- -n, --nodes, Number of nodes in the dataset
- -s, --step, Steps to predict (1 step = 5 mintues)
- -b, --batch_size, number of batches to train in each epoch
- -d --dataset, dataset name, used to save the temporary results/models
- --data_path, File path to the dataset, only support ".h5" file and ".csv" file.
- --null, null value filter (float).
- --gpu, GPU ID to run, default is -1 which means the GPU with most free memory.
- --verbose, NOT IMPLEMENTED YET.
- -l, --loss, "MAE, MSE, RMSE, MAPE and HYBRID" (MAE+MAPE) are supported.
- -c, --coefficient, hybrid coeffcient, for example, c*MAE+MAPE.
| Step3 |
MAE |
RMSE |
MAPE |
| STID |
2.566 |
4.906 |
6.67% |
| AGCRN |
2.567 |
4.838 |
6.51% |
| GWNET |
2.467 |
4.567 |
6.12% |
| Step6 |
MAE |
RMSE |
MAPE |
| STID |
2.810 |
5.622 |
7.69% |
| AGCRN |
2.795 |
5.483 |
7.48% |
| GWNET |
2.740 |
5.305 |
7.13% |
| Step12 |
MAE |
RMSE |
MAPE |
| STID |
3.213 |
6.489 |
9.23% |
| AGCRN |
3.116 |
6.381 |
8.84% |
| GWNET |
3.045 |
6.154 |
8.39% |