Table 3.
Hyperparameter optimization of the Messidor dataset trained using transfer learning on a pretrained GoogLeNet model from ImageNet. 2-ary dataset classes were group C0:R0, R1 and C1:R2, R3. 3-ary dataset classes were C0: R0, C1:R1, and C2:R2, R3. 4-ary dataset classes were C0:R0, C1:R1, C2:R2, C3:R3. C represents the label within the CNN architecture, and R represents the label from the dataset.
| GoogLeNet Rapid Prototyping Results-Raw Images | |||||
|---|---|---|---|---|---|
| Model | Solver | Learning Rate | Policy | Validation Accuracy% | Test Set Accuracy% |
| 2-ary | SGD | 1e-3 | Step Down | 83.82 | 72.75 |
| 2-ary | NAG | 1e-3 | Step Down | 82.36 | 72.75 |
| 2-ary | Adam | 1e-4 | Step Down | 86.40 | 71.75 |
| 2-ary | AdaGrad | 1e-3 | Exponential Decay | 84.55 | 64.25 |
| 2-ary | RMSProp | 1e-4 | Sigmoid Decay | 79.04 | 64.25 |
| 3-ary | RMSProp | 1e-4 | Exponential Decay | 63.97 | 66.25 |
| 3-ary | SGD | 1e-3 | Step Down | 71.69 | 64.25 |
| 3-ary | Adam | 1e-4 | Step Down | 72.40 | 61.50 |
| 3-ary | NAG | 1e-3 | Step Down | 69.85 | 58.75 |
| 3-ary | AdaGrad | 1e-3 | Exponential Decay | 72.43 | 58.25 |
| 4-ary | Adam | 1e-4 | Step Down | 67.65 | 57.25 |
| 4-ary | SGD | 1e-3 | Step Down | 65.07 | 55.25 |
| 4-ary | AdaGrad | 1e-3 | Exponential Decay | 66.54 | 53.25 |
| 4-ary | NAG | 1e-3 | Step Down | 66.18 | 52.75 |
| 4-ary | RMSProp | 1e-4 | Step Down | 62.50 | 49.75 |