If you check my codes, I followed the structure used in Google GitHub. We have a dense layer at the end with linear activation because we are using the logits. You can change this to softmax activation and then the keras precision and recall metrics would work. Hopefully this helps!