Efficient Inference of Large-Scale and Lightweight Convolutional Neural Networks on FPGA