Layer-Peeled Model: Toward Understanding Well-Trained Deep Neural Networks