Created by: luotao1

conv_transpose_mkldnn_op has accuracy diff on our internal face model. commit: 3fe2def1 is OK. commit: 1eb15175 is wrong. If I remove conv_transpose in op_lists, the accuracy is OK.

if (use_mkldnn) {
     cfg.EnableMKLDNN();
     std::unordered_set<std::string> op_list={"conv", "batch_norm", "relu", "leaky_relu", "elementwise_add", "tanh"};
     cfg.SetMKLDNNOp(op_list);
   }