Parallelizing TensorFlow