Using optimizer = keras.optimizers.SGD(clipvalue=1.0), means that all the partial derivatives of the loss (with regard to each and every trainable parameter) will be clipped between 0 and 1.0.