Can I directly obtain the output value of DNN reasoning on GPU？which is convenient for the use of cascaded network