TensorRT Samples: MNIST(Plugin, add a custom layer)

关于TensorRT的介绍可以参考： http://blog.csdn.net/fengbingchun/article/details/78469551   

以下是参考TensorRT 2.1.2中的samplePlugin.cpp文件改写的通过IPlugin添加一个全连接层实现对手写数字0-9识别的测试代码，plugin.cpp文件内容如下：

#include <assert.h>
#include <iostream>
#include <memory>
#include <vector>
#include <string>

#include <cuda_runtime_api.h>
#include <cudnn.h>
#include <cublas_v2.h>
#include <NvInfer.h>
#include <NvCaffeParser.h>
#include <opencv2/opencv.hpp>

#include "common.hpp"

// reference: TensorRT-2.1.2/samples/samplePlugin/samplePlugin.cpp
// demonstrates how to add a custom layer to TensorRT. It replaces the final fully connected layer of the MNIST sample with a direct call to cuBLAS

namespace {

typedef std::tuple<int, int, int, std::string, std::string> DATA_INFO; // intput width, input height, output size, input blob name, output blob name

int caffeToGIEModel(const std::string& deployFile,					// name for caffe prototxt
					 const std::string& modelFile,					// name for model 
					 const std::vector<std::string>& outputs,		// network outputs
					 unsigned int maxBatchSize,						// batch size - NB must be at least as large as the batch we want to run with)
					 nvcaffeparser1::IPluginFactory* pluginFactory,	// factory for plugin layers
					 nvinfer1::IHostMemory *&gieModelStream, Logger logger)		// output stream for the GIE model
{
	// create the builder
	nvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(logger);

	// parse the caffe model to populate the network, then set the outputs
	nvinfer1::INetworkDefinition* network = builder->createNetwork();
	nvcaffeparser1::ICaffeParser* parser = nvcaffeparser1::createCaffeParser();
	parser->setPluginFactory(pluginFactory);

	bool fp16 = builder->platformHasFastFp16();
	const nvcaffeparser1::IBlobNameToTensor* blobNameToTensor = parser->parse(deployFile.c_str(),
						modelFile.c_str(), *network, fp16 ? nvinfer1::DataType::kHALF : nvinfer1::DataType::kFLOAT);

	// specify which tensors are outputs
	for (auto& s : outputs)
		network->markOutput(*blobNameToTensor->find(s.c_str()));

	// Build the engine
	builder->setMaxBatchSize(maxBatchSize);
	builder->setMaxWorkspaceSize(1 << 20);
	builder->setHalf2Mode(fp16);

	nvinfer1::ICudaEngine* engine = builder->buildCudaEngine(*network);
	CHECK(engine != nullptr);

	// we don't need the network any more, and we can destroy the parser
	network->destroy();
	parser->destroy();

	// serialize the engine, then close everything down
	gieModelStream = engine->serialize();

	engine->destroy();
	builder->destroy();
	nvcaffeparser1::shutdownProtobufLibrary();
}

int doInference(nvinfer1::IExecutionContext& context, float* input, float* output, int batchSize, const DATA_INFO& info)
{
	const nvinfer1::ICudaEngine& engine = context.getEngine();
	// input and output buffer pointers that we pass to the engine - the engine requires exactly IEngine::getNbBindings(),
	// of these, but in this case we know that there is exactly one input and one output.
	CHECK(engine.getNbBindings() == 2);
	void* buffers[2];

	// In order to bind the buffers, we need to know the names of the input and output tensors.
	// note that indices are guaranteed to be less than IEngine::getNbBindings()
	int inputIndex = engine.getBindingIndex(std::get<3>(info).c_str()), 
		outputIndex = engine.getBindingIndex(std::get<4>(info).c_str());

	// create GPU buffers and a stream
	cudaMalloc(&buffers[inputIndex], batchSize * std::get<1>(info) * std::get<0>(info) * sizeof(float));
	cudaMalloc(&buffers[outputIndex], batchSize * std::get<2>(info) * sizeof(float));

	cudaStream_t stream;
	cudaStreamCreate(&stream);

	// DMA the input to the GPU,  execute the batch asynchronously, and DMA it back:
	cudaMemcpyAsync(buffers[inputIndex], input, batchSize * std::get<1>(info) * std::get<0>(info) * sizeof(float), cudaMemcpyHostToDevice, stream);
	context.enqueue(batchSize, buffers, stream, nullptr);
	cudaMemcpyAsync(output, buffers[outputIndex], batchSize * std::get<2>(info)*sizeof(float), cudaMemcpyDeviceToHost, stream);
	cudaStreamSynchronize(stream);

	// release the stream and the buffers
	cudaStreamDestroy(stream);
	cudaFree(buffers[inputIndex]);
	cudaFree(buffers[outputIndex]);

	return 0;
}

class FCPlugin: public nvinfer1::IPlugin
{
public:
	FCPlugin(const nvinfer1::Weights* weights, int nbWeights, int nbOutputChannels) : mNbOutputChannels(nbOutputChannels)
	{
		// since we want to deal with the case where there is no bias, we can't infer
		// the number of channels from the bias weights.
		assert(nbWeights == 2);
		mKernelWeights = copyToDevice(weights[0].values, weights[0].count);
		mBiasWeights = copyToDevice(weights[1].values, weights[1].count);
		assert(mBiasWeights.count == 0 || mBiasWeights.count == nbOutputChannels);

		mNbInputChannels = int(weights[0].count / nbOutputChannels);
	}

	// create the plugin at runtime from a byte stream
	FCPlugin(const void* data, size_t length)
	{
		const char* d = reinterpret_cast<const char*>(data), *a = d;
		mNbInputChannels = read<int>(d);
		mNbOutputChannels = read<int>(d);
		int biasCount = read<int>(d);

		mKernelWeights = deserializeToDevice(d, mNbInputChannels * mNbOutputChannels);
		mBiasWeights = deserializeToDevice(d, biasCount);
		assert(d == a + length);
	}

	~FCPlugin()
	{
		cudaFree(const_cast<void*>(mKernelWeights.values));
		cudaFree(const_cast<void*>(mBiasWeights.values));
	}

	int getNbOutputs() const override
	{
		return 1;
	}

	nvinfer1::Dims getOutputDimensions(int index, const nvinfer1::Dims* inputs, int nbInputDims) override
	{
		assert(index == 0 && nbInputDims == 1 && inputs[0].nbDims == 3);
		assert(mNbInputChannels == inputs[0].d[0] * inputs[0].d[1] * inputs[0].d[2]);
		return nvinfer1::DimsCHW(mNbOutputChannels, 1, 1);
	}

	void configure(const nvinfer1::Dims* inputDims, int nbInputs, const nvinfer1::Dims* outputDims, int nbOutputs, int maxBatchSize) override
	{
	}

	int initialize() override
	{
		cudnnCreate(&mCudnn); // initialize cudnn and cublas
		cublasCreate(&mCublas);
		cudnnCreateTensorDescriptor(&mSrcDescriptor); // create cudnn tensor descriptors we need for bias addition
		cudnnCreateTensorDescriptor(&mDstDescriptor);

		return 0;
	}

	virtual void terminate() override
	{
		cublasDestroy(mCublas);
		cudnnDestroy(mCudnn);
	}

	virtual size_t getWorkspaceSize(int maxBatchSize) const override
	{
		return 0;
	}

	virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override
	{
		float kONE = 1.0f, kZERO = 0.0f;
		cublasSetStream(mCublas, stream);
		cudnnSetStream(mCudnn, stream);
		cublasSgemm(mCublas, CUBLAS_OP_T, CUBLAS_OP_N, mNbOutputChannels, batchSize, mNbInputChannels, &kONE, 
				reinterpret_cast<const float*>(mKernelWeights.values), mNbInputChannels, 
				reinterpret_cast<const float*>(inputs[0]), mNbInputChannels, &kZERO, 
				reinterpret_cast<float*>(outputs[0]), mNbOutputChannels);
		if (mBiasWeights.count) {
			cudnnSetTensor4dDescriptor(mSrcDescriptor, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, mNbOutputChannels, 1, 1);
			cudnnSetTensor4dDescriptor(mDstDescriptor, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, batchSize, mNbOutputChannels, 1, 1);
			cudnnAddTensor(mCudnn, &kONE, mSrcDescriptor, mBiasWeights.values, &kONE, mDstDescriptor, outputs[0]);
		}

		return 0;
	}

	virtual size_t getSerializationSize() override
	{
		// 3 integers (number of input channels, number of output channels, bias size), and then the weights:
		return sizeof(int)*3 + mKernelWeights.count*sizeof(float) + mBiasWeights.count*sizeof(float);
	}

	virtual void serialize(void* buffer) override
	{
		char* d = reinterpret_cast<char*>(buffer), *a = d;

		write(d, mNbInputChannels);
		write(d, mNbOutputChannels);
		write(d, (int)mBiasWeights.count);
		serializeFromDevice(d, mKernelWeights);
		serializeFromDevice(d, mBiasWeights);

		assert(d == a + getSerializationSize());
	}

private:
	template<typename T> void write(char*& buffer, const T& val)
	{
		*reinterpret_cast<T*>(buffer) = val;
		buffer += sizeof(T);
	}

	template<typename T> T read(const char*& buffer)
	{
		T val = *reinterpret_cast<const T*>(buffer);
		buffer += sizeof(T);
		return val;
	}

	nvinfer1::Weights copyToDevice(const void* hostData, size_t count)
	{
		void* deviceData;
		cudaMalloc(&deviceData, count * sizeof(float));
		cudaMemcpy(deviceData, hostData, count * sizeof(float), cudaMemcpyHostToDevice);
		return nvinfer1::Weights{ nvinfer1::DataType::kFLOAT, deviceData, int64_t(count) };
	}

	void serializeFromDevice(char*& hostBuffer, nvinfer1::Weights deviceWeights)
	{
		cudaMemcpy(hostBuffer, deviceWeights.values, deviceWeights.count * sizeof(float), cudaMemcpyDeviceToHost);
		hostBuffer += deviceWeights.count * sizeof(float);
	}

	nvinfer1::Weights deserializeToDevice(const char*& hostBuffer, size_t count)
	{
		nvinfer1::Weights w = copyToDevice(hostBuffer, count);
		hostBuffer += count * sizeof(float);
		return w;	
	}

	int mNbOutputChannels, mNbInputChannels;
	cudnnHandle_t mCudnn;
	cublasHandle_t mCublas;
	nvinfer1::Weights mKernelWeights, mBiasWeights;
	cudnnTensorDescriptor_t mSrcDescriptor, mDstDescriptor;
};

// integration for serialization
class PluginFactory : public nvinfer1::IPluginFactory, public nvcaffeparser1::IPluginFactory
{
public:
	// caffe parser plugin implementation
	bool isPlugin(const char* name) override
	{
		return !strcmp(name, "ip2");
	}

	virtual nvinfer1::IPlugin* createPlugin(const char* layerName, const nvinfer1::Weights* weights, int nbWeights) override
	{
		// there's no way to pass parameters through from the model definition, so we have to define it here explicitly
		static const int NB_OUTPUT_CHANNELS = 10;	
		assert(isPlugin(layerName) && nbWeights == 2 && weights[0].type == nvinfer1::DataType::kFLOAT && weights[1].type == nvinfer1::DataType::kFLOAT);
		assert(mPlugin.get() == nullptr);
		mPlugin = std::unique_ptr<FCPlugin>(new FCPlugin(weights, nbWeights, NB_OUTPUT_CHANNELS));
		return mPlugin.get();
	}

	// deserialization plugin implementation
	nvinfer1::IPlugin* createPlugin(const char* layerName, const void* serialData, size_t serialLength) override
	{		
		assert(isPlugin(layerName));
		assert(mPlugin.get() == nullptr);
		mPlugin = std::unique_ptr<FCPlugin>(new FCPlugin(serialData, serialLength));
		return mPlugin.get();
	}

	// the application has to destroy the plugin when it knows it's safe to do so
	void destroyPlugin()
	{
		mPlugin.release();
	}

	std::unique_ptr<FCPlugin> mPlugin{ nullptr };
};

} // namespace

int test_plugin()
{
	// stuff we know about the network and the caffe input/output blobs
	const DATA_INFO info(28, 28, 10, "data", "prob");
	const std::string deploy_file {"models/mnist.prototxt"};
	const std::string model_file {"models/mnist.caffemodel"};
	const std::string mean_file {"models/mnist_mean.binaryproto"};
	Logger logger; // multiple instances of IRuntime and/or IBuilder must all use the same logger

	// create a GIE model from the caffe model and serialize it to a stream
	PluginFactory pluginFactory;
	nvinfer1::IHostMemory* gieModelStream{ nullptr };
	caffeToGIEModel(deploy_file, model_file, std::vector<std::string>{std::get<4>(info).c_str()}, 1, &pluginFactory, gieModelStream, logger);
	pluginFactory.destroyPlugin();

	nvcaffeparser1::ICaffeParser* parser = nvcaffeparser1::createCaffeParser();
	nvcaffeparser1::IBinaryProtoBlob* meanBlob = parser->parseBinaryProto(mean_file.c_str());
	parser->destroy();

	// deserialize the engine 
	nvinfer1::IRuntime* runtime = nvinfer1::createInferRuntime(logger);
	nvinfer1::ICudaEngine* engine = runtime->deserializeCudaEngine(gieModelStream->data(), gieModelStream->size(), &pluginFactory);
	nvinfer1::IExecutionContext *context = engine->createExecutionContext();

	// parse the mean file and 	subtract it from the image
	const float* meanData = reinterpret_cast<const float*>(meanBlob->getData());

	const std::string image_path{ "images/digit/" };
	for (int i = 0; i < 10; ++i) {
		const std::string image_name = image_path + std::to_string(i) + ".png";
		cv::Mat mat = cv::imread(image_name, 0);
		if (!mat.data) {
			fprintf(stderr, "read image fail: %s\n", image_name.c_str());
			return -1;
		}

		cv::resize(mat, mat, cv::Size(std::get<0>(info), std::get<1>(info)));
		mat.convertTo(mat, CV_32FC1);

		float data[std::get<1>(info)*std::get<0>(info)];
		const float* p = (float*)mat.data;
		for (int j = 0; j < std::get<1>(info)*std::get<0>(info); ++j) {
			data[j] = p[j] - meanData[j];
		}

		// run inference
		float prob[std::get<2>(info)];
		doInference(*context, data, prob, 1, info);

		float val{-1.f};
		int idx{-1};

		for (int t = 0; t < std::get<2>(info); ++t) {
			if (val < prob[t]) {
				val = prob[t];
				idx = t;
			}
		}

		fprintf(stdout, "expected value: %d, actual value: %d, probability: %f\n", i, idx, val);
	}

	meanBlob->destroy();
	if (gieModelStream) gieModelStream->destroy();
	// destroy the engine
	context->destroy();
	engine->destroy();
	runtime->destroy();
	pluginFactory.destroyPlugin();

	return 0;
}

测试图像如下：

执行结果如下(与http://blog.csdn.net/fengbingchun/article/details/78552908 中结果一致)：

测试代码编译步骤如下(ReadMe.txt)：

在Linux下通过CMake编译TensorRT_Test中的测试代码步骤：
1. 将终端定位到CUDA_Test/prj/linux_tensorrt_cmake，依次执行如下命令：
	$ mkdir build
	$ cd build
	$ cmake ..
	$ make (生成TensorRT_Test执行文件)
	$ ln -s ../../../test_data/models  ./ (将models目录软链接到build目录下)
	$ ln -s ../../../test_data/images  ./ (将images目录软链接到build目录下)
	$ ./TensorRT_Test
2. 对于有需要用OpenCV参与的读取图像的操作，需要先将对应文件中的图像路径修改为Linux支持的路径格式

GitHub：  https://github.com/fengbingchun/CUDA_Test