Table 11 Hyperpamareter settings for the selected initial prediction and prediction refinement models
Model | Type | Parameter | |
|---|---|---|---|
Initial prediction (TCN ) | Model architecture | #Hidden features | 32 |
#TCN layers | 5 | ||
Dilation for each TCN layer | 1, 3, 9, 27, 81 | ||
Training procedure | #Epochs | 50 | |
Batch size | 1024 | ||
Optimizer | Amsgrad (beta = (0.9, 0.999)) | ||
Learning rate | 0.0005 | ||
Learning rate decay (per epoch) | 0.95 | ||
Loss function | Weighted cross-entropy loss | ||
Prediction refinement (multi-stage TCN ) | Model architecture | #Hidden features | 32 |
#Stages | 4 | ||
#TCN layers | 8 | ||
Dilation for each TCN layer | 1, 2, 4, 8, 16, 32, 64, 128 | ||
Training procedure | #Epochs | 50 | |
Batch size | 1 | ||
Optimizer | Amsgrad (beta = (0.9, 0.999)) | ||
Learning rate | 0.0005 | ||
Learning rate decay (per epoch) | 0.95 | ||
Loss functions | Weighted cross-entropy loss + smoothing loss (\(\tau\) = 4, \(\lambda\) = 0.15) | ||