The dataset is Oxford Flowers but it has splitted into two datasets: A and B. A has 4617 samples for train and 1538 samples for test. B has 100 samples for train and 20 samples for test and test_all is 2056 samples from both A and B. Train a CNN on dataset A. then train on dataset B which is much smaller than A with three different ways:
- Copying all layers from the previous model that trained on A and just add 20 neurons which are associated with 20 classes of dataset B then train this new network.
- Copying all layers from the previous model that trained on A and Freezing all layers except the last layer which is fully connected and just add 20 neurons to the last fc layer and train only the last layer.
- Copying all layers from the previous model that trained on A and Freezing all layers except the last layer and also freezing 80 neurons of the last fc layer which are associated with 80 classes of dataset A and just add 20 neurons to the last fc layer and train only these 20 neurons.
You can check the results in document that we prepared. I'm trying to translate it to English.