Created by: wojtuss

@luotao1 , using the model from @hshen14 (https://github.com/PaddlePaddle/Paddle/issues/14074) we get top1 accuracy 70.43 % (on CPU, batchsize 50, without optimizations as well as with MKL-DNN and optimizations) on the full validation ImageNet dataset. Is it acceptable for Baidu?