Figure 9.

There are three transfer learning scenarios: (a) train whole model, (b) apply a new classifier on top of a pre-trained convolutional base, (c) fine-tune the base, by re-training one or more convolution layers.