Fork自 PaddlePaddle / PaddleDetection
Correctly use host_vector in LoDTensor and expose LoDTensor to Python.