That's not really related to the network. When you build a NN, you shouldn't expect different accuracy using different frameworks (given that you're using the same hyper parameters and training/evaluating on the same data).