softmax_cross_entropy_with_logits issue with keras

I'm using keras with tensorflow backend and using cross entropy as loss function. I saw that tensorflow requires pre softmax as an input for the loss, but I put output of the softmax in keras. What is the right way to use it?