TensorRT优化部署（一）--TensorRT和ONNX基础

系列文章目录

第一章 TensorRT优化部署（一）–TensorRT和ONNX基础
第二章 TensorRT优化部署（二）–剖析ONNX架构
第三章 TensorRT优化部署（三）–ONNX注册算子
第四章 TensorRT模型优化部署（四）–Roofline model
第五章 TensorRT模型优化部署（五）–模型优化部署重点注意
第六章 TensorRT模型优化部署（六）–Quantization量化基础（一）
第七章 TensorRT模型优化部署（七）–Quantization量化（PTQ and QAT)（二）
第八章 TensorRT模型优化部署 (八）–模型剪枝Pruning

---

---

前言

自学视频笔记，专题内容后续有补充。

---

一、模型部署目的

模型训练 (Training)

$$\{\begin{array}{l}精度越高越好\\ \\ 把模型做的深一点，宽一点\\ \\ 使用丰富的Dataaugmentation\\ \\ 使用各种Trainingtrick\end{array}$$

模型部署 (Deploy)

$$\{\begin{array}{l}以精度不变，或者精度掉点很小的情况下尽量压缩模型\\ \\ 减少计算量\\ \\ 减少memoryaccess\\ \\ 提高计算密度\end{array}$$

自动驾驶中模型部署所关注的东西

• RealTime(实时性)
• Power Consumption(消耗电力)（我们需要的算力越大，消耗的电力越大）
• Long range Accuracy(远距离)

Attention:
TensorRT虽然很方便，但是不能过分依赖。需要结合部署的硬件特性，做一些Benchmark和Profiling,学习使用一些Profile tools去分析模型的计算瓶颈，分析预测优化策略和实际优化策略产生分歧的原因。

• Benchmark通常指对某个软件系统或者部分系统进行性能测试，目的是评估其性能表现。
• Profiling则是指对代码的性能进行分析和优化。

二、TensorRT的模块

TensorRT内部各类优化策略如下图：

2.1 Layer fusion(层融合）

层融合可以减少启动kernel的开销与memort操作，从而提高效率同时，有些计算可以通过层融合优化后，跟其他计算合并。

1. Vertical layer fusion(垂直层融合)

conv+BN+ReLu进行层融合

Step1: Step2:
注意：conv + BN + ReLU的计算时间和conv的计算时间差不多

2. Horizontal layer fusion(水平层融合)

Step3:

2.2 Kernel auto-tuning

TensorRT内部对于同一个层使用各种不同kernel函数进行性能测试

比如对于FC层中的矩阵乘法，根据tile size有很多中kernel function 。(e.g. 32x32, 32x64, 64x64, 64x128, 128x128，针对不同硬件有不同策略)

2.3 Quantization

量化，压缩模型的很重要的策略，将单精度类型(FP32)训练权重转变为半精度(FP16)或者整型(INT8, INT4)

三、ONNX是什么？

ONNX是一种神经网络的格式，采用Protobuf(*)二进制形式进行序列化模型。Protobuf会根据用于定义的数据结构来进行序列化存储。
下面是一些简单的例子，代码链接：

3.1 example.py

import torch
import torch.nn as nn
import torch.onnx

class Model(torch.nn.Module):
    def __init__(self, in_features, out_features, weights, bias=False):
        super().__init__()
        self.linear = nn.Linear(in_features, out_features, bias)
        with torch.no_grad():
            self.linear.weight.copy_(weights)
    
    def forward(self, x):
        x = self.linear(x)
        return x

def infer():
    in_features = torch.tensor([1, 2, 3, 4], dtype=torch.float32)
    weights = torch.tensor([
        [1, 2, 3, 4],
        [2, 3, 4, 5],
        [3, 4, 5, 6]
    ],dtype=torch.float32)
    
    model = Model(4, 3, weights)
    x = model(in_features)
    print("result is: ", x)

def export_onnx():
    input   = torch.zeros(1, 1, 1, 4)
    weights = torch.tensor([
        [1, 2, 3, 4],
        [2, 3, 4, 5],
        [3, 4, 5, 6]
    ],dtype=torch.float32)
    model   = Model(4, 3, weights)
    model.eval() #添加eval防止权重继续更新

    # pytorch导出onnx的方式，参数有很多，也可以支持动态size
    # 重点部分===========================================
    torch.onnx.export(
        model         = model, 
        args          = (input,),
        f             = "../models/example.onnx",
        input_names   = ["input0"],
        output_names  = ["output0"],
        opset_version = 12)  #onnx指令集版本
    print("Finished onnx export")


if __name__ == "__main__":
    infer()
    export_onnx()

输出结果

使用Netron打开

3.2 example_two_head.py 两个输出

import torch
import torch.nn as nn
import torch.onnx
#模型定义=================
class Model(torch.nn.Module):
    def __init__(self, in_features, out_features, weights1, weights2, bias=False):
        super().__init__()
        self.linear1 = nn.Linear(in_features, out_features, bias)
        self.linear2 = nn.Linear(in_features, out_features, bias)
        with torch.no_grad():
            self.linear1.weight.copy_(weights1)
            self.linear2.weight.copy_(weights2)

    
    def forward(self, x):
        x1 = self.linear1(x)
        x2 = self.linear2(x)
        return x1, x2
#模型推理=====================
def infer():
    in_features = torch.tensor([1, 2, 3, 4], dtype=torch.float32)
    weights1 = torch.tensor([
        [1, 2, 3, 4],
        [2, 3, 4, 5],
        [3, 4, 5, 6]
    ],dtype=torch.float32)
    weights2 = torch.tensor([
        [2, 3, 4, 5],
        [3, 4, 5, 6],
        [4, 5, 6, 7]
    ],dtype=torch.float32)
    
    model = Model(4, 3, weights1, weights2)
    x1, x2 = model(in_features)
    print("result is: \n")
    print(x1)
    print(x2)
#导出onnx
def export_onnx():
    input    = torch.zeros(1, 1, 1, 4)
    weights1 = torch.tensor([
        [1, 2, 3, 4],
        [2, 3, 4, 5],
        [3, 4, 5, 6]
    ],dtype=torch.float32)
    weights2 = torch.tensor([
        [2, 3, 4, 5],
        [3, 4, 5, 6],
        [4, 5, 6, 7]
    ],dtype=torch.float32)
    model   = Model(4, 3, weights1, weights2)
    model.eval() #添加eval防止权重继续更新

    # pytorch导出onnx的方式，参数有很多，也可以支持动态size
    torch.onnx.export(
        model         = model, 
        args          = (input,),
        f             = "../models/example_two_head.onnx",
        input_names   = ["input0"],
        output_names  = ["output0", "output1"],
        opset_version = 12)
    print("Finished onnx export")


if __name__ == "__main__":
    infer()
    export_onnx()

3.3 动态batch，example_dynamic_shape.py

import torch
import torch.nn as nn
import torch.onnx

class Model(torch.nn.Module):
    def __init__(self, in_features, out_features, weights, bias=False):
        super().__init__()
        self.linear = nn.Linear(in_features, out_features, bias)
        with torch.no_grad():
            self.linear.weight.copy_(weights)
    
    def forward(self, x):
        x = self.linear(x)
        return x

def infer():
    in_features = torch.tensor([1, 2, 3, 4], dtype=torch.float32)
    weights = torch.tensor([
        [1, 2, 3, 4],
        [2, 3, 4, 5],
        [3, 4, 5, 6]
    ],dtype=torch.float32)

    model = Model(4, 3, weights)
    x = model(in_features)
    print("result of {1, 1, 1 ,4} is ", x.data)

def export_onnx():
    input   = torch.zeros(1, 1, 1, 4)
    weights = torch.tensor([
        [1, 2, 3, 4],
        [2, 3, 4, 5],
        [3, 4, 5, 6]
    ],dtype=torch.float32)
    model   = Model(4, 3, weights)
    model.eval() #添加eval防止权重继续更新

    # pytorch导出onnx的方式，参数有很多，也可以支持动态size
    torch.onnx.export(
        model         = model, 
        args          = (input,),
        f             = "../models/example_dynamic_shape.onnx",
        input_names   = ["input0"],
        output_names  = ["output0"],
        ######################修改了这一部分###########################
        dynamic_axes  = {
            'input0':  {0: 'batch'},
            'output0': {0: 'batch'}
        },#可以自己设置batch值
        ####################################################################
        opset_version = 12)
    print("Finished onnx export")



if __name__ == "__main__":
    infer()
    export_onnx()

3.5 sample_cbr.py

import torch
import torch.nn as nn
import torch.onnx

class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3)
        self.bn1   = nn.BatchNorm2d(num_features=16)
        self.act1  = nn.ReLU()
    
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.act1(x)
        return x

def export_norm_onnx():
    input   = torch.rand(1, 3, 5, 5)
    model   = Model()
    model.eval()

    # 通过这个案例可以发现onnx导出的时候，其实有一些节点已经被融合了
    # batchNorm不见了
    file    = "../models/sample-cbr.onnx"
    torch.onnx.export(
        model         = model, 
        args          = (input,),
        f             = file,
        input_names   = ["input0"],
        output_names  = ["output0"],
        opset_version = 12)
    print("Finished normal onnx export")

if __name__ == "__main__":
    export_norm_onnx()

我们可以发现BatchNormalization不见了，说明在torch导出onnx的时候已经被融合了。

3.6 sample_reshape.py

import torch
import torch.nn as nn
import torch.onnx
import onnxsim
import onnx

class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1   = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3, padding=1)
        self.bn1     = nn.BatchNorm2d(num_features=16)
        self.act1    = nn.ReLU()
        self.conv2   = nn.Conv2d(in_channels=16, out_channels=64, kernel_size=5, padding=2)
        self.bn2     = nn.BatchNorm2d(num_features=64)
        self.act2    = nn.ReLU()
        self.avgpool = nn.AdaptiveAvgPool1d(1)
        self.head    = nn.Linear(in_features=64, out_features=10)
    
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.act1(x)
        x = self.conv2(x)
        x = self.bn2(x)
        x = self.act2(x) 
        x = torch.flatten(x, 2, 3)  # B, C, H, W -> B, C, L (这一个过程产生了shape->slice->concat->reshape这一系列计算节点)


        # b, c, w, h = x.shape
        # x = x.reshape(b, c, w * h)
        # x = x.view(b, c, -1)

        x = self.avgpool(x)         # B, C, L    -> B, C, 1
        x = torch.flatten(x, 1)     # B, C, 1    -> B, C
        x = self.head(x)            # B, C      -> B, 10
        return x

def export_norm_onnx():
    input   = torch.rand(1, 3, 64, 64)
    model   = Model()
    file    = "../models/sample-reshape.onnx"
    torch.onnx.export(
        model         = model, 
        args          = (input,),
        f             = file,
        input_names   = ["input0"],
        output_names  = ["output0"],
        opset_version = 12)
    print("Finished normal onnx export")

    model_onnx = onnx.load(file)

    # 检查导入的onnx model
    onnx.checker.check_model(model_onnx)


    # 使用onnx-simplifier来进行onnx的简化。
    # 可以试试把这个简化给注释掉，看看flatten操作在简化前后的区别
    # onnx中其实会有一些constant value，以及不需要计算图跟踪的节点
    # 大家可以一起从netron中看看这些节点都在干什么
#=======================使用onnx-smilifier把下面代码注释取消即可=============#
    # print(f"Simplifying with onnx-simplifier {onnxsim.__version__}...")
    # model_onnx, check = onnxsim.simplify(model_onnx)
    # assert check, "assert check failed"
    onnx.save(model_onnx, file)

if __name__ == "__main__":
    export_norm_onnx()

3.7 load_torchvision.py

给出torchvision内置的模型，如何转onnx

import torch
import torchvision
import onnxsim
import onnx
import argparse

def get_model(type, dir):
    if type == "resnet":
        model = torchvision.models.resnet50()
        file  = dir + "resnet50.onnx"
    elif type == "vgg":
        model = torchvision.models.vgg11()
        file  = dir + "vgg11.onnx"
    elif type == "mobilenet":
        model = torchvision.models.mobilenet_v3_small()
        file  = dir + "mobilenetV3.onnx"
    elif type == "efficientnet":
        model = torchvision.models.efficientnet_b0()
        file  = dir + "efficientnetb0.onnx"
    elif type == "efficientnetv2":
        model = torchvision.models.efficientnet_v2_s()
        file  = dir + "efficientnetV2.onnx"
    elif type == "regnet":
        model = torchvision.models.regnet_x_1_6gf()
        file  = dir + "regnet1.6gf.onnx"
    return model, file

def export_norm_onnx(model, file, input):
    
    torch.onnx.export(
        model         = model, 
        args          = (input,),
        f             = file,
        input_names   = ["input0"],
        output_names  = ["output0"],
        opset_version = 12)
    print("Finished normal onnx export")

    model_onnx = onnx.load(file)

    # 检查导入的onnx model
    onnx.checker.check_model(model_onnx)

    # 使用onnx-simplifier来进行onnx的简化。
    print(f"Simplifying with onnx-simplifier {onnxsim.__version__}...")
    model_onnx, check = onnxsim.simplify(model_onnx)
    assert check, "assert check failed"
    onnx.save(model_onnx, file)


def main(args):
    type        = args.type
    dir         = args.dir
    input       = torch.rand(1, 3, 224, 224)
    model, file = get_model(type, dir)

    export_norm_onnx(model, file, input)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("-t", "--type", type=str, default="vgg")
    parser.add_argument("-d", "--dir", type=str, default="../models/")
    opt = parser.parse_args()
    main(opt)

总结

本章为专题内容，主要介绍TensorRT优化部署，可移步专题查看其他内容。
参考链接
视频推荐