Using TensorFlow Hub
Of these three approaches (TensorFlow Hub, the Estimators API, and the Keras API), TensorFlow Hub stands out from the other two. It is a library for open source machine learning models. The main purpose of TensorFlow Hub is to enable model reusability through transfer learning. Transfer learning is a very practical and convenient technique in deep learning modeling development. The hypothesis is that as a well-designed model (peer reviewed and made famous by publications) learned patterns in features during the training process, the model learned to generalize these patterns, and such generalization can be applied to new data. Therefore, we do not need to retrain the model again when we have new training data.
Let's take human vision as an example. The content of what we see can be decomposed from simple to sophisticated patterns in the order of lines, edges, shapes, layers, and finally a pattern. As it turns out, this is how a computer vision model recognizes...