noncudnn_model.compile( loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True), optimizer="sgd", metrics=["accuracy"], )