Table 6.
Ablation study on changing kernel size, kernel size of the pooling layer, loss function, and batch size.
| Study 7: Changing the Kernel Size | |||||
| Configuration No. | No. of Kernel Size | No. of Parameter | Epoch × Training Time | Test Accuracy (%) | Finding |
| 1 | 4 | 284,166 | 200 × 19 s | 84.33 | Previous accuracy |
| 2 | 3 | 225,478 | 200 × 16 s | 84.62 | Highest Accuracy |
| 3 | 2 | 183,558 | 200 × 13 s | 80.12 | Accuracy dropped |
| 4 | 1 | 158,406 | 200 × 11 s | 76.31 | Accuracy dropped |
| Study 8: Changing the kernel size of the pooling layer | |||||
| Configuration No. | No. of pooling kernel size | No. of Parameter | Epoch × training time | Test accuracy (%) | Finding |
| 1 | 5 | 225,478 | 200 × 16 s | 84.57 | Accuracy dropped |
| 2 | 4 | 225,478 | 200 × 16 s | 85.12 | Accuracy improved |
| 3 | 3 | 225,478 | 200 × 16 s | 85.68 | Highest Accuracy |
| 4 | 2 | 225,478 | 200 × 16 s | 84.62 | Previous accuracy |
| 5 | 1 | 225,478 | 200 × 16 s | 83.9 | Accuracy dropped |
| Study 9: Changing the loss function | |||||
| Configuration No. | Loss Function | No. of Parameter | Epoch × training time | Test accuracy (%) | Finding |
| 1 | Binary Cross-entropy | 225,478 | 200 × 16 s | 87.12 | Accuracy improved |
| 2 | Categorical Cross-entropy | 225,478 | 200 × 16 s | 87.35 | Highest Accuracy |
| 3 | Mean Squared Error | 225,478 | 200 × 16 s | 85.68 | Previous accuracy |
| 4 | Mean absolute error | 225,478 | 200 × 16 s | 84.93 | Accuracy dropped |
| 5 | Mean squared logarithmic error | 225,478 | 200 × 16 s | 85.76 | Accuracy dropped |
| Study 10: Changing the batch size | |||||
| Configuration No. | Batch size | No. of Parameter | Epoch × training time | Test accuracy (%) | Finding |
| 1 | 256 | 225,478 | 200 × 12 s | 86.88 | Accuracy dropped |
| 2 | 128 | 225,478 | 200 × 16 s | 87.35 | Previous accuracy |
| 3 | 64 | 225,478 | 200 × 22 s | 87.53 | Accuracy improved |
| 4 | 32 | 225,478 | 200 × 31 s | 87.78 | Accuracy improved |
| Study 11: Changing the optimizer | |||||
| Configuration No. | Optimizer | No. of Parameter | Epoch × training time | Test accuracy (%) | Finding |
| 1 | Adam | 225,478 | 200 × 16 s | 87.42 | Highest Accuracy |
| 2 | Nadam | 225,478 | 200 × 16 s | 86.78 | Accuracy dropped |
| 3 | SGD | 225,478 | 200 × 16 s | 87.35 | Previous accuracy |
| 4 | Adamax | 225,478 | 200 × 16 s | 84.18 | Accuracy dropped |
| 5 | RMSprop | 225,478 | 200 × 16 s | 86.8 | Accuracy dropped |
| Study 12: Changing the learning rate | |||||
| Configuration No. | Learning rate | No. of Parameter | Epoch × training time | Test accuracy (%) | Finding |
| 1 | 0.01 | 225,478 | 200 × 16 s | 86.12 | Accuracy dropped |
| 2 | 0.006 | 225,478 | 200 × 16 s | 87.42 | Previous accuracy |
| 3 | 0.001 | 225,478 | 200 × 16 s | 90.17 | Highest Accuracy |
| 4 | 0.0008 | 225,478 | 200 × 16 s | 89.8 | Accuracy improved |