TRT::Tensor roi_align_inputs(TRT::DataType::Float);
roi_align_inputs.resize(infer_batch_size * MAX_IMAGE_BBOX * 6);
roi_align_inputs.to_cpu(false);
output_array_device.to_cpu(true);

to_cpu(true)，转cpu，同时拷贝一份。

测试执行时间：

const int ntest =1;
	auto begin_timer = iLogger::timestamp_now_float();
	for (int i = 0; i < ntest; ++i)
		boxes_array = engine->commits(images);
	// wait all result
	boxes_array.back().get();
	float inference_average_time = (iLogger::timestamp_now_float() - begin_timer) / ntest / images.size();
	auto type_name = FasterRCNN::type_name(type);
	auto mode_name = TRT::mode_string(mode);
	INFO("%s[%s] average: %.2f ms / image, FPS: %.2f", engine_file.c_str(), type_name, inference_average_time, 1000 / inference_average_time);

后面持续更新。