Towards Efficient Hardware Acceleration Of Deep Neural Networks On Fpga