Although Keras is good, and interactive between TF with Keras is possible, it’s not so convenient to jump between Keras and TF cause the backend of Keras is not the same with TF. And, we have to admit that TF is more flexible than Keras, sometimes our need is verbose under Keras for we have to define a new layer. Plus, I have found that nearly every model can be found in TF code, not Keras. So, I think working mainly under TF is a good idea. For these reason, TF+TF-Slim may be the best choice.