Pure numpy MLP for multi-classification.
| Modules | Baseline | Best Model |
|---|---|---|
| Batch size | 128 | 1024 |
| Learning rate | 0.1 | 0.01 |
| Scheduler | CosineAnnealingLR | None |
| Epoch | 100 | 200 |
| Pre-processing | Standardization | None |
| Number of Hidden layer | 2 | 2 |
| Hidden units | [64, 32] | [256, 128] |
| Activations | [Relu, Relu] | [Relu, Relu] |
| Weight initialisation | Kaiming | Kaiming |
| Weight decay | 5e-4 | 5e-4 |
| Optimizer | SGD with Momentum | SGD with Momentum |
| Momentum | 0.9 | 0.9 |
| Batch Normalisation | Yes | Yes |
| Dropout rate | 0.1 | 0.3 |
| Accuracy | 53.03% | 58.71% |