TL;DR version:

• TensorFlow if you want to control things inside the models precisely (e.g. improving the deep learning algorithm).
• Keras if you want high level control (e.g. applying deep learning to a new problem).