TABLE 4.
Ranges and chosen values for hyperparameters and augmentation parameters
Parameter | Type | Range | Value |
---|---|---|---|
Batch size | Hyperparameter | [16, 32, 64, 128] | 64 |
Dropout | Hyperparameter | [0.2, 0.5] | 0.5 |
Epochs | Hyperparameter | 20 | 20 |
L2 amount | Hyperparameter | [1e−6, 1e−2] | 1e−6 |
Learning rate | Hyperparameter | [1e−9, 1e−2] | 0.0075 |
Network size scale | Hyperparameter | [1, 2, 3, 4] | 2 |
Gaussian noise intensity | Augmentation | [0, 20] | 0.8 |
Gaussian blur | Augmentation | [0, 3] | 0 |
Horizontal shift | Augmentation | [0, 50 px] | [0, 20 px] |
Random dB offset | Augmentation | [0, −40 dB] | [0, −40 dB] |
Vertical shift | Augmentation | [0, 5 px] | [0, 2 px] |
Single value ranges indicate that the parameter was held constant. Zero‐values for continuous parameters indicate that the parameter was not used. Values that are ranges indicate that values were randomly chosen from this range on a per‐spectrogram basis (augmentation only).