Link to access the models weights: https://drive.google.com/drive/folders/1-G_RMiavFqebwgmJxSrFXQu-GdzEQFFa?usp=sharing
Standard ResNet freezed with the following added:
- Dense of 128 nodes, using relu
- Dropout of 0.2
- Dense of 64 nodes, using relu
- Dropout of 0.2
- Dense of 3 nodes, using softmax
Learning Rate of 0.0001 using Adam, with decaying callback on keras.
This model was retrained with the imagenet weights and the three last layers equal to ResNet50 1.
The frozen ResNet architecture had a horrible performance for this case. For the ResNet50 1 we tried many iterations, but the results were the same: training from scratch resulted in a bad training. On the other hand, retraining using the already trained weights handled better results, but still overall worse than the other models explored.