基于PyTorch与TensorRT的cifar10推理加速引擎(C++)

一、写在开头

1、基于PyTorch训练出cifar10模型

2、以ONNX(Open Neural Network Exchange)格式导出模型cifar10.onnx

3、下载cifar10二进制版本数据集

4、创建TensorRT(vs c++)项目，解析模型，进行推理

 二、基于PyTorch的cifar10神经网络模型

import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import datetime

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Assuming that we are on a CUDA machine, this should print a CUDA device:
print(device)


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


# functions to show an image
def imshow(img):
    img = img / 2 + 0.5     # unnormalize
    npimg = img.numpy()
    plt.imshow(np.transpose(npimg, (1, 2, 0)))
    plt.show()

# load dataset
transform = transforms.Compose(
    [transforms.ToTensor()])
       # transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) 

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
                                        download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=4,
                                          shuffle=True, num_workers=0)

testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                       download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=4,
                                         shuffle=False, num_workers=0)

classes = ('plane', 'car', 'bird', 'cat',
           'deer', 'dog', 'frog', 'horse', 'ship', 'truck')


# get some random training images
dataiter = iter(trainloader)
images, labels = dataiter.next()
# show images
imshow(torchvision.utils.make_grid(images))
# print labels
print(' '.join('%5s' % classes[labels[j]] for j in range(4)))
net = Net()
net = Net().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

for epoch in range(2):

    running_loss = 0.0
    for i, data in enumerate(trainloader, 0):
        # get the inputs; data is a list of [inputs, labels]
        # inputs, labels = data
        inputs, labels = data[0].to(device), data[1].to(device)
        # zero the parameter gradients
        optimizer.zero_grad()

        # forward + backward + optimize
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

        # print statistics
        running_loss += loss.item()
        if i % 2000 == 1999:  # print every 2000 mini-batches
            print('[%d, %5d] loss: %.3f' %
                  (epoch + 1, i + 1, running_loss / 2000))
            running_loss = 0.0

    print('Finished Training')

correct = 0
total = 0
start = datetime.datetime.now()
with torch.no_grad():
    for data in trainloader:
        images, labels = data
        images, labels = images.to(device), labels.to(device)
        outputs = net(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()
print('Accuracy of the network on the 50000 test images: %d %%' % (
    100 * correct / total))

end = datetime.datetime.now()
print((end - start).seconds, 's')
# 导出onnx模型
dummy_input = torch.randn(1, 3, 32, 32, device='cuda')
input_names = ['input']
output_names = ['output']
torch.onnx.export(net, dummy_input, "cifar10.onnx", verbose='True', input_names=input_names, output_names=output_names) 

三、下载cifar10二进制数据集

https://tianchi.aliyun.com/dataset/dataDetail?dataId=43780

四、构建cifar10的TensorRT推理引擎C++

TensorRT给的samples中有好几版mnist模型的例子，mnist和cifar10非常类似，主要区别在于cifar10数据是RGB三通道的，所以我们在mnist的基础上把它改成cifar10推理引擎。

其实，主要的修改在于输入与输出部分，因为TensorRT的核心函都是封装好的。要看懂下面代码，必须参考TensorRT给的samples

std::vector cifarbinary;//存储每个batch文件的标签、RGB数据

inline void readBinaryFile(const std::string& filename, vector& tempbinary) {//用来读取每个batch文件

    std::ifstream infile(filename, std::ifstream::binary);
    assert(infile.is_open() && "Attempting to read from a file that is not open.");
    infile.seekg(0, ios::end);
    const int length = 10000 * (32 * 32 * 3 + 1);
    gLogInfo << filename << " : " << length << " bytes" << std::endl;
    tempbinary.resize(length);

    infile.seekg(0, ios::beg);
    infile.read(reinterpret_cast<char*>(tempbinary.data()), length);
    infile.close();
}

//这里对processInput函数做了修改，它直接完成数据读取并进行推理
bool SampleOnnxCIFAR10::processInput(samplesCommon::BufferManager& buffers, SampleUniquePtr& context)
{   
    const int inputC = mInputDims.d[1];
    const int inputH = mInputDims.d[2];
    const int inputW = mInputDims.d[3];
    const int batchSize = mParams.batchSize;
    const int volImg = inputC * inputH * inputW;
    const int imageSize = volImg + 1;
    const int outputSize = mOutputDims.d[1];
    float* hostDataBuffer = static_cast<float*>(buffers.getHostBuffer(mParams.inputTensorNames[0]));
    int temp[10];
    int maxposition{0};
    int count{ 0 };
    // 5 batchbinary files
    auto starttime = clock();
    for (int index = 0; index < 5; ++index) {
        // Read cifar10 original binary file
        readBinaryFile(locateFile("data_batch_" + std::to_string(index + 1) + ".bin", mParams.dataDirs), cifarbinary);
      
         for (int i = 0; i < 10000; ++i) {

             for (int j = 0; j < 32 * 32 * 3; ++j) {
                 //RGB format
                 hostDataBuffer[j] = float(cifarbinary[i * imageSize + j])/255.0;
             }
             // Memcpy from host input buffers to device input buffers
             buffers.copyInputToDevice();
             //execute inference on every image
             bool status = context->executeV2(buffers.getDeviceBindings().data());
             assert(status == true);
             // Memcpy from device output buffers to host output buffers
             buffers.copyOutputToHost();
             //verifyoutput
             float* output = static_cast<float*>(buffers.getHostBuffer(mParams.outputTensorNames[0]));

             maxposition = max_element(output, output + 10) - output;
             //predict correctly
             if (maxposition == int(cifarbinary[i * imageSize])) {

                 ++count;
             }
         }
    }
    auto endtime = clock();
    gLogInfo << "The accuracy of the TRT Engine on 50000 data is :" << float(count) / 50000.0 << endl;
    gLogInfo << "TotalUse time is :" << double(endtime - starttime) / CLOCKS_PER_SEC << "s" << std::endl;
    return true;
}

五、TensorRT与PyTorch的推理速度对比

在50000个image上进行了测试，TRT(c++)引擎的速度要比PyTorch快一倍多，效率提升非常明显。

模型重构以后，准确率略有下降。