For myself, I was surprised by how low-level Tensorflow was. I found that basic operations (things like saving and reloading a trained model) were much more difficult than in Keras, scikit-learn, or other machine learning libraries. Using Tensorflow feels sort of like computing derivatives with limits.