tensorflow1.8 c++ 加载训练好的模型 踩坑

tensorflow c++下加载训练好的图和checkpoint的代码是：

const std::string pathToGraph =  "./model/model.ckpt-29520.meta";
	const std::string checkpointPath = "./model/model.ckpt-29520";

	//gpu option
	tf::SessionOptions session_options;
	session_options.config.mutable_gpu_options()->set_allow_growth(true);

	tf::Session* m_sessionFaceNet = NewSession(session_options);

	// Read in the protobuf graph we exported
	tf::MetaGraphDef graph_def;
	tf::Status m_statusFaceNet = ReadBinaryProto(tf::Env::Default(), pathToGraph, &graph_def);
	if (!m_statusFaceNet.ok())
	{
		//throw std::runtime_error("Error reading graph definition from " + pathToGraph + ": " + m_statusFaceNet.ToString());
		return 0;
	}
	// Add the graph to the session
	m_statusFaceNet = m_sessionFaceNet->Create(graph_def.graph_def());

	// Read weights from the saved checkpoint
	tf::Tensor checkpointPathTensor(tf::DT_STRING, tf::TensorShape());
	checkpointPathTensor.scalar<std::string>()() = checkpointPath;
	m_statusFaceNet = m_sessionFaceNet->Run(
	{ { graph_def.saver_def().filename_tensor_name(), checkpointPathTensor }, },
	{},
	{ graph_def.saver_def().restore_op_name() },
		nullptr);

在tensorflow1.2下上述代码工作正常，但是由于工作需要，最近将tensorflow升级到了1.8.0，重新编译得到lib，dll后，运行上述代码出现了下述两个链接bug：

error LNK2001: 无法解析的外部符号 "class tensorflow::GraphDefDefaultTypeInternal tensorflow::_GraphDef_default_instance_" (?_GraphDef_default_instance_@tensorflow@@3VGraphDefDefaultTypeInternal@1@A)

main.obj : error LNK2001: 无法解析的外部符号 "class tensorflow::SaverDefDefaultTypeInternal tensorflow::_SaverDef_default_instance_" (?_SaverDef_default_instance_@tensorflow@@3VSaverDefDefaultTypeInternal@1@A)

多种方法折腾之后，找到了解决办法，修改后的代码为：

const std::string pathToGraph =  "./model/model.ckpt-29520.meta";
    const std::string checkpointPath = "./model/model.ckpt-29520";

    //gpu option
    tf::SessionOptions session_options;
    session_options.config.mutable_gpu_options()->set_allow_growth(true);


    /*std::unique_ptr<tensorflow::Session> m_sessionFaceNet;
    Status load_graph_status = LoadGraph(pathToFrozen, &m_sessionFaceNet);
    tf::Status m_statusFaceNet;*/

    tf::Session* m_sessionFaceNet = NewSession(session_options);

    // Read in the protobuf graph we exported
    tf::MetaGraphDef graph_def;
    tf::Status m_statusFaceNet = ReadBinaryProto(tf::Env::Default(), pathToGraph, &graph_def);
    if (!m_statusFaceNet.ok())
    {
        //throw std::runtime_error("Error reading graph definition from " + pathToGraph + ": " + m_statusFaceNet.ToString());
        return 0;
    }
    // Add the graph to the session
    m_statusFaceNet = m_sessionFaceNet->Create(*(graph_def.mutable_graph_def()));

    // Read weights from the saved checkpoint
    tf::Tensor checkpointPathTensor(tf::DT_STRING, tf::TensorShape());
    checkpointPathTensor.scalar<std::string>()() = checkpointPath;
    m_statusFaceNet = m_sessionFaceNet->Run(
    { { graph_def.mutable_saver_def()->filename_tensor_name(), checkpointPathTensor }, },
    {},
    { graph_def.mutable_saver_def()->restore_op_name() },
        nullptr);

主要的修改是修改了graph_def函数为mutable_graph_def(),修改saver_def()为mutable_saver_def()。误打误撞解决了问题，原理不理解，后续有时间还要分析tensorflow底层函数。

另一个要注意的问题是：

cv::Mat转换为tensorflow::Tensor时，要注意转换的数据格式和模型要求的数据格式一致。

两种转换方式：

tf::Tensor input_tensor(tf::DT_FLOAT, tf::TensorShape({ nPersonNum,112,112,1}));

cv::Mat flaotMat;
resizeImg.convertTo(flaotMat, CV_32FC1);

tf::StringPiece tmp_data = input_tensor.tensor_data();
	memcpy(const_cast<char*>(tmp_data.data()), (flaotMat.data), flaotMat.rows * flaotMat.cols * sizeof(tf::DT_FLOAT));

另一种转换方式之灰度图：

//转换成的格式应该视自己模型的需要决定，此处转换为32F
resizeImg.convertTo(flaotMat, CV_32FC1);

for (int nTensor = 0; nTensor < nPersonNum; nTensor++)
	{
		const float * source_data = (float*)tmpVecWhitenImages[nTensor].data;
		for (int y = 0; y < tmpVecWhitenImages[nTensor].rows; ++y)
		{
			const float* source_row = source_data + (y * tmpVecWhitenImages[nTensor].cols);
			for (int x = 0; x < tmpVecWhitenImages[nTensor].cols; ++x)
			{
				const float* source_pixel = source_row + x;
				for (int c = 0; c < 1; ++c)
				{
					const float* source_value = source_pixel;
					input_tensor_mapped(nTensor, y, x, c) = *source_value;
				}
			}
		}
	}

另一种转换方式之RGB图像：

//转换成的格式应该视自己模型的需要决定，此处转换成64F
matImgByTF.convertTo(Image2, CV_64FC1);

tf::Tensor input_tensor(tf::DT_FLOAT, tf::TensorShape({ nPersonNum,m_pFaceDescriber->m_nFaceImgWidth,m_pFaceDescriber->m_nFaceImgHeight,m_pFaceDescriber->m_nFaceImgDepth }));
        auto input_tensor_mapped = input_tensor.tensor<float, 4>();

//double*
for (int nTensor = 0; nTensor < nPersonNum; nTensor++)
		{
			const double * source_data = (double*)tmpVecWhitenImages[nTensor].data;
			for (int y = 0; y < tmpVecWhitenImages[nTensor].rows; ++y)
			{
				const double* source_row = source_data + (y * tmpVecWhitenImages[nTensor].cols * m_pFaceDescriber->m_nFaceImgDepth);
				for (int x = 0; x < tmpVecWhitenImages[nTensor].cols; ++x)
				{
					const double* source_pixel = source_row + (x * m_pFaceDescriber->m_nFaceImgDepth);
					for (int c = 0; c < m_pFaceDescriber->m_nFaceImgDepth; ++c)
					{
						const double* source_value = source_pixel + (2 - c);//RGB->BGR
						input_tensor_mapped(nTensor, y, x, c) = *source_value;
					}
				}
			}
		}

遗留的问题：

可以对meta和ckpt进行冻结得到.pb文件，但是自己冻结得到的模型运行结果和原始meta、ckpt的运行结果不一致，可能是冻结图的时候某些参数设置的不合理，此问题留待后续解决。

暂时草草记录，后面有时间详细整理，如有错误，欢迎批评指正。