如何使用TensorRT對(duì)訓(xùn)練好的PyTorch模型進(jìn)行加速？

一.簡(jiǎn)介

TensorRT是Nvidia公司出的能加速模型推理的框架，其實(shí)就是讓你訓(xùn)練的模型在測(cè)試階段的速度加快，比如你的模型測(cè)試一張圖片的速度是50ms，那么用tensorRT加速的話，可能只需要10ms。當(dāng)然具體能加速多少也不能保證，反正確實(shí)速度能提升不少。但是TensorRT坑爹的地方在于，有些模型操作是不支持的、又或者就算支持但是支持并不完善，對(duì)于這些難題，要么自己寫插件，要么就只能等待官方的更新了。

現(xiàn)在我們訓(xùn)練深度學(xué)習(xí)模型主流的框架有tensorflow，pytorch，mxnet，caffe等。這個(gè)貼子只涉及pytorch，對(duì)于tensorflow的話，可以參考TensorRT部署深度學(xué)習(xí)模型，https://zhuanlan.zhihu.com/p/84125533，這個(gè)帖子是c++如何部署TensorRT。其實(shí)原理都是一樣的，對(duì)于tensorflow模型，需要把pb模型轉(zhuǎn)化為uff模型；對(duì)于pytorch模型，需要把pth模型轉(zhuǎn)化為onnx模型；對(duì)于caffe模型，則不需要轉(zhuǎn)化，因?yàn)閠ensorRT是可以直接讀取caffe模型的。mxnet模型也是需要轉(zhuǎn)化為onnx的。

那么，這篇教學(xué)貼主要是從python和c++兩種語言環(huán)境下，嘗試將pytorch模型轉(zhuǎn)化為tensorRT，教剛接觸TensorRT的同學(xué)們?nèi)绾慰焖偕鲜帧?/p>

二.TensorRT的安裝

TensorRT的安裝并不難，推薦安裝最新版本的。由于我使用的是Centos，因此我一般是按照這個(gè)教程來安裝TensorRT的。

CentOS安裝TensorRT指南
https://tbr8.org/how-to-install-tensorrt-on-centos/

安裝完成后，在python環(huán)境下import tensorrt看能不能成功，并且編譯一下官方的sampleMnist的例子，如果都可以的話，就安裝成功了。

python環(huán)境下，成功導(dǎo)入tensorrt

運(yùn)行官方的mnist例子

三.Python環(huán)境下pytorch模型如何轉(zhuǎn)化為TensorRT

python環(huán)境下pytorch模型轉(zhuǎn)化為TensorRT有兩種路徑，一種是先把pytorch的pt模型轉(zhuǎn)化為onnx，然后再轉(zhuǎn)化為TensorRT；另一種是直接把pytorch的pt模型轉(zhuǎn)成TensorRT。

首先，我們先把pt模型轉(zhuǎn)化為onnx模型，需要安裝onnx，直接pip install onnx即可。我們以ResNet50為例，代碼如下：

import torchvisionimport torchfrom torch.autograd import Variableimport onnxprint(torch.__version__)
input_name = ['input']output_name = ['output']input = Variable(torch.randn(1, 3, 224, 224)).cuda()model = torchvision.models.resnet50(pretrained=True).cuda()torch.onnx.export(model, input, 'resnet50.onnx', input_names=input_name, output_names=output_name, verbose=True)

以上代碼使用torchvision里面預(yù)訓(xùn)練的resnet50模型為基礎(chǔ)，將resnet50的pt模型轉(zhuǎn)化成res50.onnx，其中規(guī)定onnx的輸入名是'input'，輸出名是'output'，輸入圖像的大小是3通道224x224。其中batch size是1，其實(shí)這個(gè)batch size你可以取3、4、5等。運(yùn)行這個(gè)代碼就可以生成一個(gè)名為resnet50.onnx文件。

最好檢查一下生成的onnx，代碼如下：

test = onnx.load('resnet50.onnx')
onnx.checker.check_model(test)
print("==> Passed")

接下來比較一下pytorch模型和TensorRT的結(jié)果吧：

import pycuda.autoinitimport numpy as npimport pycuda.driver as cudaimport tensorrt as trtimport torchimport osimport timefrom PIL import Imageimport cv2import torchvision
filename = 'test.jpg'max_batch_size = 1onnx_model_path = 'resnet50.onnx'
TRT_LOGGER = trt.Logger()  # This logger is required to build an engine

def get_img_np_nchw(filename):    image = cv2.imread(filename)    image_cv = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)    image_cv = cv2.resize(image_cv, (224, 224))    miu = np.array([0.485, 0.456, 0.406])    std = np.array([0.229, 0.224, 0.225])    img_np = np.array(image_cv, dtype=float) / 255.    r = (img_np[:, :, 0] - miu[0]) / std[0]    g = (img_np[:, :, 1] - miu[1]) / std[1]    b = (img_np[:, :, 2] - miu[2]) / std[2]    img_np_t = np.array([r, g, b])    img_np_nchw = np.expand_dims(img_np_t, axis=0)    return img_np_nchw
class HostDeviceMem(object):    def __init__(self, host_mem, device_mem):        """Within this context, host_mom means the cpu memory and device means the GPU memory        """        self.host = host_mem        self.device = device_mem
    def __str__(self):        return "Host:\n" + str(self.host) + "\nDevice:\n" + str(self.device)
    def __repr__(self):        return self.__str__()

def allocate_buffers(engine):    inputs = []    outputs = []    bindings = []    stream = cuda.Stream()    for binding in engine:        size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size        dtype = trt.nptype(engine.get_binding_dtype(binding))        # Allocate host and device buffers        host_mem = cuda.pagelocked_empty(size, dtype)        device_mem = cuda.mem_alloc(host_mem.nbytes)        # Append the device buffer to device bindings.        bindings.append(int(device_mem))        # Append to the appropriate list.        if engine.binding_is_input(binding):            inputs.append(HostDeviceMem(host_mem, device_mem))        else:            outputs.append(HostDeviceMem(host_mem, device_mem))    return inputs, outputs, bindings, stream

def get_engine(max_batch_size=1, onnx_file_path="", engine_file_path="", \               fp16_mode=False, int8_mode=False, save_engine=False,               ):    """Attempts to load a serialized engine if available, otherwise builds a new TensorRT engine and saves it."""
    def build_engine(max_batch_size, save_engine):        """Takes an ONNX file and creates a TensorRT engine to run inference with"""        with trt.Builder(TRT_LOGGER) as builder, \                builder.create_network() as network, \                trt.OnnxParser(network, TRT_LOGGER) as parser:
            builder.max_workspace_size = 1 << 30  # Your workspace size            builder.max_batch_size = max_batch_size            # pdb.set_trace()            builder.fp16_mode = fp16_mode  # Default: False            builder.int8_mode = int8_mode  # Default: False            if int8_mode:                # To be updated                raise NotImplementedError
            # Parse model file            if not os.path.exists(onnx_file_path):                quit('ONNX file {} not found'.format(onnx_file_path))
            print('Loading ONNX file from path {}...'.format(onnx_file_path))            with open(onnx_file_path, 'rb') as model:                print('Beginning ONNX file parsing')                parser.parse(model.read())
            print('Completed parsing of ONNX file')            print('Building an engine from file {}; this may take a while...'.format(onnx_file_path))
            engine = builder.build_cuda_engine(network)            print("Completed creating Engine")
            if save_engine:                with open(engine_file_path, "wb") as f:                    f.write(engine.serialize())            return engine
    if os.path.exists(engine_file_path):        # If a serialized engine exists, load it instead of building a new one.        print("Reading engine from file {}".format(engine_file_path))        with open(engine_file_path, "rb") as f, trt.Runtime(TRT_LOGGER) as runtime:            return runtime.deserialize_cuda_engine(f.read())    else:        return build_engine(max_batch_size, save_engine)

def do_inference(context, bindings, inputs, outputs, stream, batch_size=1):    # Transfer data from CPU to the GPU.    [cuda.memcpy_htod_async(inp.device, inp.host, stream) for inp in inputs]    # Run inference.    context.execute_async(batch_size=batch_size, bindings=bindings, stream_handle=stream.handle)    # Transfer predictions back from the GPU.    [cuda.memcpy_dtoh_async(out.host, out.device, stream) for out in outputs]    # Synchronize the stream    stream.synchronize()    # Return only the host outputs.    return [out.host for out in outputs]

def postprocess_the_outputs(h_outputs, shape_of_output):    h_outputs = h_outputs.reshape(*shape_of_output)    return h_outputs


img_np_nchw = get_img_np_nchw(filename)img_np_nchw = img_np_nchw.astype(dtype=np.float32)
# These two modes are dependent on hardwaresfp16_mode = Falseint8_mode = Falsetrt_engine_path = './model_fp16_{}_int8_{}.trt'.format(fp16_mode, int8_mode)# Build an engineengine = get_engine(max_batch_size, onnx_model_path, trt_engine_path, fp16_mode, int8_mode)# Create the context for this enginecontext = engine.create_execution_context()# Allocate buffers for input and outputinputs, outputs, bindings, stream = allocate_buffers(engine) # input, output: host # bindings
# Do inferenceshape_of_output = (max_batch_size, 1000)# Load data to the bufferinputs[0].host = img_np_nchw.reshape(-1)
# inputs[1].host = ... for multiple inputt1 = time.time()trt_outputs = do_inference(context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream) # numpy datat2 = time.time()feat = postprocess_the_outputs(trt_outputs[0], shape_of_output)
print('TensorRT ok')
model = torchvision.models.resnet50(pretrained=True).cuda()resnet_model = model.eval()
input_for_torch = torch.from_numpy(img_np_nchw).cuda()t3 = time.time()feat_2= resnet_model(input_for_torch)t4 = time.time()feat_2 = feat_2.cpu().data.numpy()print('Pytorch ok!')

mse = np.mean((feat - feat_2)**2)print("Inference time with the TensorRT engine: {}".format(t2-t1))print("Inference time with the PyTorch model: {}".format(t4-t3))print('MSE Error = {}'.format(mse))
print('All completed!')

運(yùn)行結(jié)果如下：

TensorRT okPytorch ok!Inference time with the TensorRT engine: 0.0037250518798828125Inference time with the PyTorch model: 0.3574800491333008MSE Error = 3.297184357139993e-12

這個(gè)結(jié)果Pytorch模型ResNet50竟然需要340ms，感覺有些迷，但是好像沒發(fā)現(xiàn)有啥問題。可以發(fā)現(xiàn)，TensorRT進(jìn)行inference的結(jié)果和pytorch前向的結(jié)果差距很小。代碼來源于https://github.com/RizhaoCai/PyTorch_ONNX_TensorRT

接下來介紹python環(huán)境下，直接把pytorch模型轉(zhuǎn)化為TensorRT，參考的代碼來源于NVIDIA-AI-IOT/torch2trt，https://github.com/NVIDIA-AI-IOT/torch2trt這個(gè)工程比較簡(jiǎn)單易懂，質(zhì)量很高，安裝也不難，我自己運(yùn)行的結(jié)果如下：

對(duì)于你自己的Pytorch模型，只需要把該代碼的model進(jìn)行替換即可。注意在運(yùn)行過程中經(jīng)常會(huì)出現(xiàn)"output tensor has no attribute _trt"，這是因?yàn)槟隳Ｐ彤?dāng)中有一些操作還沒有實(shí)現(xiàn)，需要自己實(shí)現(xiàn)。

四.C++環(huán)境下Pytorch模型如何轉(zhuǎn)化為TensorRT

c++環(huán)境下，以TensorRT5.1.5.0的sampleOnnxMNIST為例子，用opencv讀取一張圖片，然后讓TensorRT進(jìn)行doInference輸出(1,1000)的特征。代碼如下所示，把這個(gè)代碼替換sampleOnnxMNIST替換，然后編譯就能運(yùn)行了。

#include #include #include #include #include #include #include #include #include #include #include #include "NvInfer.h"#include "NvOnnxParser.h"#include "argsParser.h"#include "logger.h"#include "common.h"#include "image.hpp"#define DebugP(x) std::cout << "Line" << __LINE__ << "  " << #x << "=" << x << std::endl

using namespace nvinfer1;
static const int INPUT_H = 224;static const int INPUT_W = 224;static const int INPUT_C = 3;static const int OUTPUT_SIZE = 1000;
const char* INPUT_BLOB_NAME = "input";const char* OUTPUT_BLOB_NAME = "output";
const std::string gSampleName = "TensorRT.sample_onnx_image";

samplesCommon::Args gArgs;

bool onnxToTRTModel(const std::string& modelFile, // name of the onnx model                    unsigned int maxBatchSize,    // batch size - NB must be at least as large as the batch we want to run with                    IHostMemory*& trtModelStream) // output buffer for the TensorRT model{    // create the builder    IBuilder* builder = createInferBuilder(gLogger.getTRTLogger());    assert(builder != nullptr);    nvinfer1::INetworkDefinition* network = builder->createNetwork();
    auto parser = nvonnxparser::createParser(*network, gLogger.getTRTLogger());
    //Optional - uncomment below lines to view network layer information    //config->setPrintLayerInfo(true);    //parser->reportParsingInfo();
    if ( !parser->parseFromFile( locateFile(modelFile, gArgs.dataDirs).c_str(), static_cast<int>(gLogger.getReportableSeverity()) ) )    {        gLogError << "Failure while parsing ONNX file" << std::endl;        return false;    }        // Build the engine    builder->setMaxBatchSize(maxBatchSize);    //builder->setMaxWorkspaceSize(1 << 20);    builder->setMaxWorkspaceSize(10 << 20);    builder->setFp16Mode(gArgs.runInFp16);    builder->setInt8Mode(gArgs.runInInt8);
    if (gArgs.runInInt8)    {        samplesCommon::setAllTensorScales(network, 127.0f, 127.0f);    }        samplesCommon::enableDLA(builder, gArgs.useDLACore);        ICudaEngine* engine = builder->buildCudaEngine(*network);    assert(engine);
    // we can destroy the parser    parser->destroy();
    // serialize the engine, then close everything down    trtModelStream = engine->serialize();    engine->destroy();    network->destroy();    builder->destroy();
    return true;}
void doInference(IExecutionContext& context, float* input, float* output, int batchSize){    const 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.    assert(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()        const int inputIndex = engine.getBindingIndex(INPUT_BLOB_NAME);    const int outputIndex = engine.getBindingIndex(OUTPUT_BLOB_NAME);        DebugP(inputIndex); DebugP(outputIndex);    // create GPU buffers and a stream    CHECK(cudaMalloc(&buffers[inputIndex], batchSize * INPUT_C * INPUT_H * INPUT_W * sizeof(float)));    CHECK(cudaMalloc(&buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float)));
    cudaStream_t stream;    CHECK(cudaStreamCreate(&stream));
    // DMA the input to the GPU,  execute the batch asynchronously, and DMA it back:    CHECK(cudaMemcpyAsync(buffers[inputIndex], input, batchSize * INPUT_C * INPUT_H * INPUT_W * sizeof(float), cudaMemcpyHostToDevice, stream));    context.enqueue(batchSize, buffers, stream, nullptr);    CHECK(cudaMemcpyAsync(output, buffers[outputIndex], batchSize * OUTPUT_SIZE * sizeof(float), cudaMemcpyDeviceToHost, stream));    cudaStreamSynchronize(stream);
    // release the stream and the buffers    cudaStreamDestroy(stream);    CHECK(cudaFree(buffers[inputIndex]));    CHECK(cudaFree(buffers[outputIndex]));}
//!//! \brief This function prints the help information for running this sample//!void printHelpInfo(){    std::cout << "Usage: ./sample_onnx_mnist [-h or --help] [-d or --datadir=] [--useDLACore=]\n";    std::cout << "--help          Display help information\n";    std::cout << "--datadir       Specify path to a data directory, overriding the default. This option can be used multiple times to add multiple directories. If no data directories are given, the default is to use (data/samples/mnist/, data/mnist/)" << std::endl;    std::cout << "--useDLACore=N  Specify a DLA engine for layers that support DLA. Value can range from 0 to n-1, where n is the number of DLA engines on the platform." << std::endl;    std::cout << "--int8          Run in Int8 mode.\n";    std::cout << "--fp16          Run in FP16 mode." << std::endl;}
int main(int argc, char** argv){    bool argsOK = samplesCommon::parseArgs(gArgs, argc, argv);    if (gArgs.help)    {        printHelpInfo();        return EXIT_SUCCESS;    }    if (!argsOK)    {        gLogError << "Invalid arguments" << std::endl;        printHelpInfo();        return EXIT_FAILURE;    }    if (gArgs.dataDirs.empty())    {        gArgs.dataDirs = std::vector<std::string>{"data/samples/mnist/", "data/mnist/"};    }
    auto sampleTest = gLogger.defineTest(gSampleName, argc, const_cast<const char**>(argv));
    gLogger.reportTestStart(sampleTest);
    // create a TensorRT model from the onnx model and serialize it to a stream    IHostMemory* trtModelStream{nullptr};
    if (!onnxToTRTModel("resnet50.onnx", 1, trtModelStream))        gLogger.reportFail(sampleTest);
    assert(trtModelStream != nullptr);    std::cout << "Successfully parsed ONNX file!!!!" << std::endl;            std::cout << "Start reading the input image!!!!" << std::endl;        cv::Mat image = cv::imread(locateFile("test.jpg", gArgs.dataDirs), cv::IMREAD_COLOR);    if (image.empty()) {        std::cout << "The input image is empty!!! Please check....."<<std::endl;    }    DebugP(image.size());    cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
    cv::Mat dst = cv::Mat::zeros(INPUT_H, INPUT_W, CV_32FC3);    cv::resize(image, dst, dst.size());    DebugP(dst.size());
    float* data = normal(dst); 
    // deserialize the engine    IRuntime* runtime = createInferRuntime(gLogger);    assert(runtime != nullptr);    if (gArgs.useDLACore >= 0)    {        runtime->setDLACore(gArgs.useDLACore);    }
    ICudaEngine* engine = runtime->deserializeCudaEngine(trtModelStream->data(), trtModelStream->size(), nullptr);    assert(engine != nullptr);    trtModelStream->destroy();    IExecutionContext* context = engine->createExecutionContext();    assert(context != nullptr);        float prob[OUTPUT_SIZE];    typedef std::chrono::high_resolution_clock Time;    typedef std::chrono::duration<double, std::ratio<1, 1000>> ms;    typedef std::chrono::duration<float> fsec;    double total = 0.0;
    // run inference and cout time    auto t0 = Time::now();    doInference(*context, data, prob, 1);    auto t1 = Time::now();    fsec fs = t1 - t0;    ms d = std::chrono::duration_cast(fs);    total += d.count();    // destroy the engine    context->destroy();    engine->destroy();    runtime->destroy();        std::cout << std::endl << "Running time of one image is:" << total << "ms" << std::endl;       gLogInfo << "Output:\n";    for (int i = 0; i < OUTPUT_SIZE; i++)    {        gLogInfo << prob[i] << " ";    }    gLogInfo << std::endl;
    return gLogger.reportTest(sampleTest, true);}

其中image.cpp的代碼為：

#include #include "image.hpp"
static const float kMean[3] = { 0.485f, 0.456f, 0.406f };static const float kStdDev[3] = { 0.229f, 0.224f, 0.225f };static const int map_[7][3] = { {0,0,0} ,        {128,0,0},        {0,128,0},        {0,0,128},        {128,128,0},        {128,0,128},        {0,128,0}};

float* normal(cv::Mat img) {  //cv::Mat image(img.rows, img.cols, CV_32FC3);  float * data;  data = (float*)calloc(img.rows*img.cols * 3, sizeof(float));
  for (int c = 0; c < 3; ++c)  {            for (int i = 0; i < img.rows; ++i)    { //獲取第i行首像素指針       cv::Vec3b *p1 = img.ptr(i);      //cv::Vec3b *p2 = image.ptr(i);      for (int j = 0; j < img.cols; ++j)      {        data[c * img.cols * img.rows + i * img.cols + j] = (p1[j][c] / 255.0f - kMean[c]) / kStdDev[c];             }    }          }  return data;}

image.hpp的內(nèi)容為：

#pragma oncetypedef struct {  int w;  int h;  int c;  float *data;} image;float* normal(cv::Mat img);

運(yùn)行結(jié)果為：

同樣的test.jpg在python環(huán)境下的運(yùn)行結(jié)果為：

可以發(fā)現(xiàn)，c++環(huán)境下resnet50輸出的(1,1000)的特征與python環(huán)境下feat1(TensorRT)和feat2(pytorch)的結(jié)果差距很小。

上面的是將pytorch首先轉(zhuǎn)化為onnx，然后讓TensorRT解析onnx從而構(gòu)建TensorRT引擎。那么我們?nèi)绾巫孴ensorRT直接加載引擎文件呢，也就是說，我們先把onnx轉(zhuǎn)化為TensorRT的trt文件，然后讓c++環(huán)境下的TensorRT直接加載trt文件，從而構(gòu)建engine。

在這里我們首先使用onnx-tensorrt這個(gè)項(xiàng)目來使resnet50.onnx轉(zhuǎn)化為resnet50.trt。采用的項(xiàng)目是https://github.com/onnx/onnx-tensorrt這個(gè)項(xiàng)目的安裝也不難。按要求安裝好protobuf就可以。安裝成功的結(jié)果如下：

運(yùn)行如下命令，就可以獲得rensnet50.trt這個(gè)引擎文件

onnx2trt resnet50.onnx -o resnet50.trt

需要注意的是，onnx-tensort這個(gè)項(xiàng)目在編譯的時(shí)有一個(gè)指定GPU計(jì)算能力的選項(xiàng)，如下圖所示：

https://developer.nvidia.com/cuda-gpus可以查看不同顯卡的計(jì)算能力，比如你用7.5計(jì)算力生成的trt文件，是不能用6.5的顯卡來解析的。

另外在onnx2trt命令有個(gè)-b操作，是指定生成的trt文件的batch size的。在實(shí)際test過程中，你的batch size是多少，這個(gè)就設(shè)置成多少。我記得我當(dāng)時(shí)trt文件的batch size是1，但是我實(shí)際的batch size是8，運(yùn)行后，只有一張圖片有結(jié)果，其他7張圖片都是0。

如果能順利生成trt文件的話，在代碼中可以直接添加以下函數(shù)，來生成engine, 其他就不需要改變。

bool read_TRT_File(const std::string& engineFile, IHostMemory*& trtModelStream){     std::fstream file;     std::cout << "loading filename from:" << engineFile << std::endl;     nvinfer1::IRuntime* trtRuntime;     //nvonnxparser::IPluginFactory* onnxPlugin = createPluginFactory(gLogger.getTRTLogger());     file.open(engineFile, std::ios::binary | std::ios::in);     file.seekg(0, std::ios::end);     int length = file.tellg();     std::cout << "length:" << length << std::endl;     file.seekg(0, std::ios::beg);     std::unique_ptr<char[]> data(new char[length]);     file.read(data.get(), length);     file.close();     std::cout << "load engine done" << std::endl;     std::cout << "deserializing" << std::endl;     trtRuntime = createInferRuntime(gLogger.getTRTLogger());     //ICudaEngine* engine = trtRuntime->deserializeCudaEngine(data.get(), length, onnxPlugin);     ICudaEngine* engine = trtRuntime->deserializeCudaEngine(data.get(), length, nullptr);     std::cout << "deserialize done" << std::endl;     assert(engine != nullptr);     std::cout << "The engine in TensorRT.cpp is not nullptr" <<std::endl;     trtModelStream = engine->serialize();     return true; }

如果想保存引擎文件的話，可以在自己的代碼中添加這幾句話，就可以生成trt文件，然后下次直接調(diào)用trt文件。

  nvinfer1::IHostMemory* data = engine->serialize();  std::ofstream file;  file.open(filename, std::ios::binary | std::ios::out);  cout << "writing engine file..." << endl;  file.write((const char*)data->data(), data->size());  cout << "save engine file done" << endl;  file.close();

五.總結(jié)

TensorRT的部署并不難，難的是模型轉(zhuǎn)化，在這個(gè)過程中有太多的操作是TensorRT不支持的，或者pytorch模型轉(zhuǎn)化成的onnx本身就有問題。經(jīng)常會(huì)出現(xiàn)，expand, Gather, reshape不支持等。感覺TensorRT對(duì)pytorch的維度變化特別不友好，我自己在模型轉(zhuǎn)化過程中絕大多數(shù)bug都出在維度變化上。如果你有什么問題的話，請(qǐng)?jiān)谙路搅粞园桑『冒桑瑫簳r(shí)就先寫這么多，以后再補(bǔ)充吧。

最近TensorRT7出來了，支持了挺多5版本，6版本無法支持的操作。發(fā)現(xiàn)了一個(gè)tiny-tensorrt，貌似在C++和python環(huán)境下部署很easy，暫時(shí)還沒測(cè)試過，但是先記錄一下。https://github.com/zerollzeng/tiny-tensorrt