Table 2 Hyperparameter details.
Hyperparameter | Description | Search Range / Value |
|---|---|---|
Number of convolutional blocks | Number of hierarchical LFTBs with CLSTM | 4 |
Filters per conv layer | Number of filters in each block | [32, 64, 128, 256] |
Kernel size | Size of 1D convolutional kernel | 3 |
Pooling type | Downsampling method | Max Pooling |
Batch normalization | After each convolutional block | Applied |
LSTM units | Number of hidden units in the LSTM layer | [64, 128, 256] |
GRU units (GFTB) | Number of units in global GRU stack | [64, 128, 256] |
Dense layers | Number and size of dense layers after LSTM | 2 layers: [128, 64] |
Dropout rate | Dropout is applied to prevent overfitting | [0.2, 0.3, 0.4, 0.5] |
Learning rate | Learning rate for optimizer | [0.0001, 0.001, 0.01] |
Batch size | Number of samples per training batch | [32, 64, 128] |
Activation Function | Activation used in dense layers | ReLU |
Output activation | Final classification activation | Sigmoid |
Loss function | A combination of classification and distance loss | Softmax + Center Loss (λ tuned) |
Epochs | Maximum training iterations | 50 |
