Why does model.evalute() return incorrect accuracy in Keras?

Question

I trained a model with tensorflow.keras.applications.Xception. The model after training got loss less than 0.2 and accuracy more than 0.9.