A Power Efficient Neural Network Implementation on Heterogeneous FPGA and GPU Devices