tensorRT C++使用pt转engine模型进行推理

本文链接：https://blog.csdn.net/wyw0000/article/details/139884247

1. 前言

本文不讲tensorRT的推理流程，因为这种文章很多，这里着重讲从标准yolov5的tensort推理代码（模型转pt->wts->engine）改造成TPH-yolov5（pt->onnx->engine）的过程。

2. 模型转换

请查看上一篇文章https://blog.csdn.net/wyw0000/article/details/139737473?spm=1001.2014.3001.5502

3. 修改Binding

如果不修改Binding，会报下图中的错误。
在这里插入图片描述
该问题是由于Binding有多个，而代码中只申请了input和output，那么如何查看engine模型有几个Bingding呢？代码如下：

int get_model_info(const string& model_path) {
    // 创建 logger
    Logger gLogger;

    // 从文件中读取 engine
    std::ifstream engineFile(model_path, std::ios::binary);
    if (!engineFile) {
        std::cerr << "Failed to open engine file." << std::endl;
        return -1;
    }

    engineFile.seekg(0, engineFile.end);
    long int fsize = engineFile.tellg();
    engineFile.seekg(0, engineFile.beg);

    std::vector<char> engineData(fsize);
    engineFile.read(engineData.data(), fsize);
    if (!engineFile) {
        std::cerr << "Failed to read engine file." << std::endl;
        return -1;
    }

    // 反序列化 engine
    auto runtime = nvinfer1::createInferRuntime(gLogger);
    auto engine = runtime->deserializeCudaEngine(engineData.data(), fsize, nullptr);

    // 获取并打印输入和输出绑定信息
    for (int i = 0; i < engine->getNbBindings(); ++i) {
        nvinfer1::Dims dims = engine->getBindingDimensions(i);
        nvinfer1::DataType type = engine->getBindingDataType(i);

        std::cout << "Binding " << i << " (" << engine->getBindingName(i) << "):" << std::endl;
        std::cout << "  Type: " << (int)type << std::endl;
        std::cout << "  Dimensions: ";
        for (int j = 0; j < dims.nbDims; ++j) {
            std::cout << (j ? "x" : "") << dims.d[j];
        }
        std::cout << std::endl;
        std::cout << "  Is Input: " << (engine->bindingIsInput(i) ? "Yes" : "No") << std::endl;
    }

    // 清理资源
    engine->destroy();
    runtime->destroy();

    return 0;
}

下图是我的tph-yolov5的Binding，可以看到有5个Binding，因此在doInference推理之前，要给5个Binding都申请空间，同时要注意获取BindingIndex时，名称和dimension与查询出来的对应。
在这里插入图片描述

//for tph-yolov5
    int Sigmoid_921_index = trt->engine->getBindingIndex("onnx::Sigmoid_921");
    int Sigmoid_1183_index = trt->engine->getBindingIndex("onnx::Sigmoid_1183");
    int Sigmoid_1367_index = trt->engine->getBindingIndex("onnx::Sigmoid_1367");
    CUDA_CHECK(cudaMalloc(&trt->buffers[Sigmoid_921_index], BATCH_SIZE * 3 * 192 * 192 * 7 * sizeof(float)));
    CUDA_CHECK(cudaMalloc(&trt->buffers[Sigmoid_1183_index], BATCH_SIZE * 3 * 96 * 96 * 7 * sizeof(float)));
    CUDA_CHECK(cudaMalloc(&trt->buffers[Sigmoid_1367_index], BATCH_SIZE * 3 * 48 * 48 * 7 * sizeof(float)));

    trt->data = new float[BATCH_SIZE * 3 * INPUT_H * INPUT_W];
    trt->prob = new float[BATCH_SIZE * OUTPUT_SIZE];
    trt->inputIndex = trt->engine->getBindingIndex(INPUT_BLOB_NAME);
    trt->outputIndex = trt->engine->getBindingIndex(OUTPUT_BLOB_NAME);

还有推理的部分也要做修改，原来只有input和output两个Binding时，那么输出是buffers[1]，而目前是有5个Binding那么输出就变成了buffers[4]

void doInference(IExecutionContext& context, cudaStream_t& stream, void **buffers, float* output, int batchSize) {
    // infer on the batch asynchronously, and DMA output back to host
    context.enqueueV2(buffers, stream, nullptr);
    //CUDA_CHECK(cudaMemcpyAsync(output, buffers[1], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream));
    CUDA_CHECK(cudaMemcpyAsync(output, buffers[4], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream));
    cudaStreamSynchronize(stream);
}

4. 修改后处理

之前的yolov5推理代码是将pt模型转为wts再转为engine的，输出维度只有一维，而TPH输出维度为145152*7，因此要对原来的后处理代码进行修改。

struct BoundingBox {
    //bbox[0],bbox[1],bbox[2],bbox[3],conf, class_id
    float x1, y1, x2, y2, score, index;
};

float iou(const BoundingBox&  box1, const BoundingBox& box2) {
	float max_x = max(box1.x1, box2.x1);  // 找出左上角坐标哪个大
	float min_x = min(box1.x2, box2.x2);  // 找出右上角坐标哪个小
	float max_y = max(box1.y1, box2.y1);
	float min_y = min(box1.y2, box2.y2);
	if (min_x <= max_x || min_y <= max_y) // 如果没有重叠
		return 0;
	float over_area = (min_x - max_x) * (min_y - max_y);  // 计算重叠面积
	float area_a = (box1.x2 - box1.x1) * (box1.y2 - box1.y1);
	float area_b = (box2.x2 - box2.x1) * (box2.y2 - box2.y1);
	float iou = over_area / (area_a + area_b - over_area);
	return iou;
}

std::vector<BoundingBox> nonMaximumSuppression(std::vector<std::vector<float>>& boxes, float overlapThreshold) {
	std::vector<BoundingBox> convertedBoxes;

	// 将数据转换为BoundingBox结构体
	for (const auto&  box: boxes) {
		if (box.size() == 6) { // Assuming [x1, y1, x2, y2, score]
			BoundingBox bbox;
			bbox.x1 = box[0];
			bbox.y1 = box[1];
			bbox.x2 = box[2];
			bbox.y2 = box[3];
			bbox.score = box[4];
			bbox.index = box[5];
			convertedBoxes.push_back(bbox);
		}
		else {
			std::cerr << "Invalid box format!" << std::endl;
		}
	}

	// 对框按照分数降序排序
	std::sort(convertedBoxes.begin(), convertedBoxes.end(), [](const BoundingBox& a, const BoundingBox&  b) {
		return a.score > b.score;
		});

	// 非最大抑制
	std::vector<BoundingBox> result;
	std::vector<bool> isSuppressed(convertedBoxes.size(), false);

	for (size_t i = 0; i < convertedBoxes.size(); ++i) {
		if (!isSuppressed[i]) {
			result.push_back(convertedBoxes[i]);

			for (size_t j = i + 1; j < convertedBoxes.size(); ++j) {
				if (!isSuppressed[j]) {
					float overlap = iou(convertedBoxes[i], convertedBoxes[j]);

					if (overlap > overlapThreshold) {
						isSuppressed[j] = true;
					}
				}
			}
		}
	}
#if 0
	// 输出结果
	std::cout << "NMS Result:" << std::endl;
	for (const auto& box: result) {
		std::cout << "x1: " << box.x1 << ", y1: " << box.y1
			<< ", x2: " << box.x2 << ", y2: " << box.y2
			<< ", score: " << box.score << ",index:" << box.index << std::endl;
	}
#endif 
	return result;
}

void post_process(float *prob_model, float conf_thres, float overlapThreshold, std::vector<Yolo::Detection> & detResult)
{
	int cols = 7, rows = 145152;
	//  ========== 8. 获取推理结果 =========
	std::vector<std::vector<float>> prediction(rows, std::vector<float>(cols));

	int index = 0;
	for (int i = 0; i < rows; ++i) {
		for (int j = 0; j < cols; ++j) {
			prediction[i][j] = prob_model[index++];
		}
	}

	//  ========== 9. 大于conf_thres加入xc =========
	std::vector<std::vector<float>> xc;
	for (const auto& row : prediction) {
		if (row[4] > conf_thres) {
			xc.push_back(row);
		}
	}
	//  ========== 10. 置信度 = obj_conf * cls_conf =========
	//std::cout << xc[0].size() << std::endl;
	for (auto& row: xc) {
		for (int i = 5; i < xc[0].size(); i++) {
			row[i] *= row[4];
		}
	}

	// ========== 11. 切片取出xywh 转为xyxy=========
	std::vector<std::vector<float>> xywh;
	for (const auto& row: xc) {
		std::vector<float> sliced_row(row.begin(), row.begin() + 4);
		xywh.push_back(sliced_row);
	}
	std::vector<std::vector<float>> box(xywh.size(), std::vector<float>(4, 0.0));

	xywhtoxxyy(xywh, box);
	
	// ========== 12. 获取置信度最高的类别和索引=========
	std::size_t mi = xc[0].size();
	std::vector<float> conf(xc.size(), 0.0);
	std::vector<float> j(xc.size(), 0.0);

	for (std::size_t i = 0; i < xc.size(); ++i) {
		// 模拟切片操作 x[:, 5:mi]
		auto sliced_x = std::vector<float>(xc[i].begin() + 5, xc[i].begin() + mi);

		// 计算 max
		auto max_it = std::max_element(sliced_x.begin(), sliced_x.end());

		// 获取 max 的索引
		std::size_t max_index = std::distance(sliced_x.begin(), max_it);

		// 将 max 的值和索引存储到相应的向量中
		conf[i] = *max_it;
		j[i] = max_index;  // 加上切片的起始索引
	}

	// ========== 13. concat x1, y1, x2, y2, score, index;======== =
	for (int i = 0; i < xc.size(); i++) {
		box[i].push_back(conf[i]);
		box[i].push_back(j[i]);
	}

	std::vector<std::vector<float>> output;
	for (int i = 0; i < xc.size(); i++) {
		output.push_back(box[i]); // 创建一个空的 float 向量并
	}

	// ==========14 应用非最大抑制 ==========
	std::vector<BoundingBox>  result = nonMaximumSuppression(output, overlapThreshold);
	for (const auto& r : result)
	{
		Yolo::Detection det;
		det.bbox[0] = r.x1;
		det.bbox[1] = r.y1;
		det.bbox[2] = r.x2;
		det.bbox[3] = r.y2;
		det.conf = r.score;
		det.class_id = r.index;
		detResult.push_back(det);
	}

}