One must use the same loss function for model training as one uses for model validation code example

Example: loss funfction suited for softmax

model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])