Table 1.
Performance table of training models.
| Transferred models | Accuracy |
GPU time (seconds) | Parameters (millions) | Number of layers | |||
|---|---|---|---|---|---|---|---|
| Full | Full-H25 | Quarter | Half | ||||
| SqueezeNet | 85.55 | 85.5 | 73.5 | 82.8 | 4137 | 1.24 | 68 |
| Alexnet | 87.2 | 83.6 | 73.7 | 82.6 | 3805 | 61 | 25 |
| ResNet18 | 90.65 | 90.2 | 83.4 | 86 | 4256 | 11.7 | 72 |
| MobileNet-v2 | 90.75 | 89.8 | 79.9 | 84.9 | 7032 | 3.5 | 155 |
| GoogLeNet | 90.9 | 88.7 | 68.2 | 85.5 | 5104 | 7 | 144 |
| Resnet50 | 91.2 | 91.4 | 81.3 | 86.3 | 7302 | 25.6 | 177 |
| Resnet101 | 91.55 | 91.7 | 83.6 | 86.1 | 12,215 | 44.6 | 347 |
| Inception-v3 | 92 | 92.1 | 84.1 | 89.5 | 11,938 | 23.9 | 316 |
| InceptionResnet-v2 | 92.1 | 91.9 | 82.2 | 86.9 | 33,283 | 55.9 | 825 |
“Quarter” set used about 2000 images for training and validation.
“Half” set used about 5000 images for training and validation.
“Full” represents an average accuracy of twice evaluation of full dataset (80% training and 20% validation).
“Full-H25” represents adding additional 25 fully connected hidden layer to “Full” model.
GPU time represents the processing power needed for training the model.