We just take a pre-made network from top (which recognizes many different things), cut off several top layers, then putting own classifier there. And then quickly learn it on a small dataset keras-team/keras#4465 ImageNet itself has categories like food and person

First, we run the vgg model with a cut-off top on the data, and save the output to a file We will get generated data bottleneck_features_train.npy and bottleneck_classes_train.npy (wich is done by vgg_train_save_outputs.py). From these do the training "data" for our little model bottleneck_vgg_model.h5 Model learned with vgg_train_on_saved.py Then lets try to evaluate it using vgg_evaluate_on_saved.py with small dataset

Useful links

• keras-team using pre trained VGG16 for another classification task

• Francois Chollet Building powerful image classification models using very little data

• Greg Chu How to use transfer learning and fine-tuning in Keras and Tensorflow to build an image recognition system and classify (almost) any object

Especially thanks to hcl14