I understand Transfer Learning as follows.
? Adapt a model trained in an area where data can be widely applied to another area where data collection is difficult
? Lower levels of CNN seem to learn general-purpose features
? Apply to the new domain by learning only the part close to the output
So, the steps in this case are written in the section Transfer Learning of the example as follows.
? Remove the linear layer from the pre-trained net
? Create a new net composed of the pretrained net followed by a linear and softmax layer(LinearLayer[6] in this case)
? Train on the dataset, freezing all the weights except for new layer