A standard VGG16 network has been pretrained on the whole ImageNet with precomputed weights downloaded from the Internet. This network has been then juxtaposed with a custom network that has been also trained separately. Then, the juxtaposed network has been retrained as a whole, keeping 15 lower layers of the VGG16 frozen.

This combination is very effective. It allows us to save significant computational power and reuse the work already performed for VGG16 by performing a transfer learning of what the network has learned on ImageNet, apply this learning to our new specific domain, and therefore perform a fine-tuning classification task.