Table 2.
Models of ResNET-50, multi-layer CNN, and Xception along with their hyperparameters. The ResNET-50 and multi-layered CNN models were trained using 113 hpi embryo images. Accuracy and loss represent the validation accuracy and loss.
| Model | Layers between bottleneck and classification layer | Learning rate | Optimizer | Decay | Loss | Accuracy |
|---|---|---|---|---|---|---|
| ResNET-50 | ||||||
|
113 hpi | ||||||
| 1 | Base architecture | 0.001 | Adam | ND | 0.8825 | 0.6011 |
| 2 | Base architecture + dropout (0.5) | 0.01 | SGD | ND | 0.9665 | 0.6137 |
| 3 | Base architecture + dropout (0.5) | 0.005 | Adam | ND | 0.9112 | 0.6039 |
| 4 | Base architecture + dropout (0.5) | 0.01 | SGD | DwS | 0.9269 | 0.6000 |
| 5 | Base architecture + dropout (0.5), additional layers 3 (1024, 1024, 512) + 20 trainable layers | 0.001 | Adam | ND | 1.6235 | 0.2058 |
| 6 |
Base architecture + dropout (0.5) + additional layers (1024) |
0.001 |
SGD |
DwS |
0.9323 |
0.5847 |
| CNN | ||||||
|
113 hpi | ||||||
| 1 | 2 layer3-3 + flatten + 1000 + 5 | 0.0005 | SGD | Decay | 1.4641 | 0.3549 |
| 2 | 2 layer5-5 + flatten + 1000 + 5 | 0.0005 | SGD | Decay | 1.4183 | 0.3843 |
| 3 | 2 layer3-5 + flatten + 1000 + 5 | 0.0005 | SGD | Decay | 1.4198 | 0.3803 |
| 4 | 2 layer5-5 + BN-BN + flatten + 1000 + 5 | 0.0005 | Adam | Decay | 1.3510 | 0.4058 |
| 5 | 2 layer5-5 + BN-BN + global average pooling + 32 + 5 | 0.0005 | Adam | Decay | 1.5231 | 0.3333 |
| 6 | 2 layer5-5 + global average pooling + 32 + 5 | 0.0005 | SGD | Decay | 1.5956 | 0.2588 |
| 7 | 2 layer5-5 + BN-BN + flatten + 1000 + 5 | 0.0005 | SGD | ND | 1.4178 | 0.3647 |
| 8 | 2 layer5-5 + BN-BN + flatten + 1000 + 5 | 0.0005 | Adam | Decay | 1.8889 | 0.3200 |
| 9 | 5 layer5-5 + BN + flatten + 64 + 5 | 0.005 | Adam | Decay | 1.2799 | 0.4313 |
| 10 | 5 layer5-5 + BN + same padding + flatten + 512 + 5 | 0.0005 | SGD | Decay | 1.2255 | 0.4294 |
| 11 | 5 layer5-5 + BN + global average pooling + 16 + 5 | 0.0005 | SGD | Decay | 1.2730 | 0.4215 |
| 12 | 40 layer3-2 + global average pooling + flatten + dense | 0.0005 | SGD | Decay | 1.1581 | 0.4830 |
| 13 |
40 layer3-3 + global average pooling + flatten + dense |
0.0005 |
Adam |
ND |
1.1689 |
0.5304 |
| Xception | ||||||
|
113 hpi | ||||||
| 1 | Base architecture | 0.0005 | SGD | DwS | 0.8601 | 0.6373 |
| 2 | Base architecture + dropout (0.5) | 0.001 | SGD | DwS | 0.8866 | 0.6333 |
| 3 | Base architecture | 0.0001 | SGD | DwS | 0.9087 | 0.6235 |
| 4 | Base architecture + additional layers 3 (1024,1024,512) + dropout (0.5) | 0.0005 | SGD | DwS | 0.8732 | 0.6549 |
| 5 | Base architecture + dropout (0.5) | 0.0007 | SGD | DwS | 0.8704 | 0.6078 |
| 6 | Base architecture + additional layers (1024) + dropout (0.5) | 0.001 | Adam | ND | 0.8668 | 0.6372 |
| 7 | Base architecture | 0.006 | SGD | ND | 0.8651 | 0.6588 |
| 8 | Base architecture | 0.0008 | SGD | ND | 0.8850 | 0.6196 |
ND: No decay; DwS: Decay with scheduler; BN: Batch Normalization; SGD: Stochastic descent gradient.