• 针对不同平台对生成的模型进行转换，也就是俗称的parse、convert，即前端解释器
• 针对转化后的模型进行优化
• 在特定的平台(嵌入端或者服务端)成功运行已经转化好的模型
• 在模型可以运行的基础上，保证模型的速度、精度和稳定性
• 用C++、cuda写算子(预处理、op、后处理等等)去实现一些独特的算子
• 调bug、联合编译、动态静态库混搭

• 好用的开源推理框架：Caffe、NCNN、MNN、TVM(移动端首选)、OpenVino（libtorch）、libtorch、PaddlePaddle
• 好用的半开源推理框架：TensorRT（GPU服务器首选）
• 好用的开源服务器框架：triton-inference-server
• 基础知识：计算机原理、编译原理等

1. 模型结构

• 使用现有的网络结构
• 模型重参化: 训练时采用多分支的网络进行训练，尽可能利用多分支结构的优势来提升模型性能，而在使用时先进行一个等价转换，将多分支网络转换为一个单路网络，然后再进行推理使用，利用单路网络的优势提升推理的速度，从而实现高性能和高速度的平衡。

2. 剪枝

• yolov3-channel-and-layer-pruning[16]
• YOLOv3-model-pruning[17]
• centernet_prune[18]
• ResRep[19]

3. 蒸馏

• centerX[20]

4. 稀疏化训练

• 稀疏化就是随机将Tensor的部分元素置为0，类似于我们常见的dropout，附带正则化作用的同时也减少了模型的容量，从而加快了模型的推理速度。

def prune(model, amount=0.3):
    # Prune model to requested global sparsity
    import torch.nn.utils.prune as prune
    print('Pruning model... ', end='')
    for name, m in model.named_modules():
        if isinstance(m, nn.Conv2d):
            prune.l1_unstructured(m, name='weight', amount=amount)  # prune
            prune.remove(m, 'weight')  # make permanent
    print(' %.3g global sparsity' % sparsity(model))

5. 量化训练QTA(Quantization Aware Training)

• 量化后的模型在特定CPU或者GPU上相比FP32、FP16有更高的速度和吞吐，也是部署提速方法之一。

• 量化训练是在模型训练中量化的，与PTQ(训练后量化)不同, 这种量化方式对模型的精度影响不大，量化后的模型速度基本与量化前的相同(另一种量化方式PTQ，TensorRT或者NCNN中使用交叉熵进行校准量化的方式，在一些结构中会对模型的精度造成比较大的影响)。

• Pytorch->ONNX->trt onnx2trt

• NVIDIA 方案： r=s*q+b； r,s,b是浮点数，q是int8 类型定点数。b可以去掉而没有明显精度损失； r=s * q

• Google 方案： r=S（q-Z）；S是scaling factor，浮点数；Z是零点，是一个定点数； 为求得scaling factor，我们需要知道r rr的最大值与最小值，分别记为r m a x r_{max}rmax和r m i n r_{min}rmin。同样地，q qq也有最大值与最小值，记为q m a x q_{max}qmax和q m i n q_{min}qmin。q qq的位宽虽然是8比特，但是可能是有符号形式或者无符号形式，因此取值范围不同。对于无符号形式，取值范围为[ q m i n , q m a x ] = [ 0 , 255 ] [q_{min},q_{max}]=[0,255][qmin,qmax]=[0,255]；无符号形式有[ q m i n , q m a x ] = [ − 128 , 127 ] [q_{min},q_{max}]=[-128,127][qmin,qmax]=[−128,127]。

1. pytorch量化-Dynamic quantization

• 动态量化是对权重进行线下量化，即拿到训练好的浮点模型后量化权重。但是对于activation，是根据推理过程中activation的取值范围进行量化，这意味着同一个模型在处理不同输入图片时，activation采用的量化参数很可能是不一样的。
• 系统自动选择最合适的scale （标度）和 zero_point（零点位置），不需要自定义。量化后的模型，可以推理运算，但不能训练（不能反向传播）
  

import torch

# define a floating point model
class M(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = torch.nn.Linear(4, 4)

    def forward(self, x):
        x = self.fc(x)
        return x

# create a model instance
model_fp32 = M()
# create a quantized model instance
model_int8 = torch.quantization.quantize_dynamic(
    model_fp32,  # the original model
    {torch.nn.Linear},  # a set of layers to dynamically quantize
    dtype=torch.qint8)  # the target dtype for quantized weights

# run the model
input_fp32 = torch.randn(4, 4, 4, 4)
res = model_int8(input_fp32)

# import the modules used here in this recipe
import torch
import torch.quantization
import torch.nn as nn
import copy
import os
import time

# define a very, very simple LSTM for demonstration purposes
# in this case, we are wrapping nn.LSTM, one layer, no pre or post processing
# inspired by
# https://pytorch.org/tutorials/beginner/nlp/sequence_models_tutorial.html, by Robert Guthrie
# and https://pytorch.org/tutorials/advanced/dynamic_quantization_tutorial.html
class lstm_for_demonstration(nn.Module):
  """Elementary Long Short Term Memory style model which simply wraps nn.LSTM
     Not to be used for anything other than demonstration.
  """
  def __init__(self,in_dim,out_dim,depth):
     super(lstm_for_demonstration,self).__init__()
     self.lstm = nn.LSTM(in_dim,out_dim,depth)

  def forward(self,inputs,hidden):
     out,hidden = self.lstm(inputs,hidden)
     return out, hidden


torch.manual_seed(29592)  # set the seed for reproducibility

#shape parameters
model_dimension=8
sequence_length=20
batch_size=1
lstm_depth=1

# random data for input
inputs = torch.randn(sequence_length,batch_size,model_dimension)
# hidden is actually is a tuple of the initial hidden state and the initial cell state
hidden = (torch.randn(lstm_depth,batch_size,model_dimension), torch.randn(lstm_depth,batch_size,model_dimension))

torch.quantization.quantize_dynamic()

 # here is our floating point instance
float_lstm = lstm_for_demonstration(model_dimension, model_dimension,lstm_depth)

# this is the call that does the work
quantized_lstm = torch.quantization.quantize_dynamic(   #在这里调用函数不一样
    float_lstm, {nn.LSTM, nn.Linear}, dtype=torch.qint8
)

def print_size_of_model(model, label=""):
    torch.save(model.state_dict(), "temp.p")
    size=os.path.getsize("temp.p")
    print("model: ",label,' \t','Size (KB):', size/1e3)
    os.remove('temp.p')
    return size

# compare the sizes
f=print_size_of_model(float_lstm,"fp32")
q=print_size_of_model(quantized_lstm,"int8")
print("{0:.2f} times smaller".format(f/q))

out1, hidden1 = float_lstm(inputs, hidden)
mag1 = torch.mean(abs(out1)).item()
print('mean absolute value of output tensor values in the FP32 model is {0:.5f} '.format(mag1))

# run the quantized model
out2, hidden2 = quantized_lstm(inputs, hidden)
mag2 = torch.mean(abs(out2)).item()
print('mean absolute value of output tensor values in the INT8 model is {0:.5f}'.format(mag2))

# compare them
mag3 = torch.mean(abs(out1-out2)).item()
print('mean absolute value of the difference between the output tensors is {0:.5f} or {1:.2f} percent'.format(mag3,mag3/mag1*100))

• https://pytorch.org/docs/stable/quantization.html#post-training-dynamic-quantization
• https://pytorch.org/tutorials/recipes/recipes/dynamic_quantization.html

2. pytorch量化-Static quantization

• 量化后的模型，不能训练（不能反向传播），也不能推理，需要解量化后，才能进行运算
• 量化与反量化模块的添加：QuantStub()，DeQuantStub()
• 加法以及concatenate单元的量化替换
• 模块融合

# original model
# all tensors and computations are in floating point
previous_layer_fp32 -- linear_fp32 -- activation_fp32 -- next_layer_fp32
                    /
    linear_weight_fp32

# statically quantized model
# weights and activations are in int8
previous_layer_int8 -- linear_with_activation_int8 -- next_layer_int8
                    /
  linear_weight_int8

import torch

# define a floating point model where some layers could be statically quantized
class M(torch.nn.Module):
    def __init__(self):
        super().__init__()
        # QuantStub converts tensors from floating point to quantized
        self.quant = torch.quantization.QuantStub()
        self.conv = torch.nn.Conv2d(1, 1, 1)
        self.relu = torch.nn.ReLU()
        # DeQuantStub converts tensors from quantized to floating point
        self.dequant = torch.quantization.DeQuantStub()

    def forward(self, x):
        # manually specify where tensors will be converted from floating
        # point to quantized in the quantized model
        x = self.quant(x)
        x = self.conv(x)
        x = self.relu(x)
        # manually specify where tensors will be converted from quantized
        # to floating point in the quantized model
        x = self.dequant(x)
        return x

# create a model instance
model_fp32 = M()

# model must be set to eval mode for static quantization logic to work
model_fp32.eval()

# attach a global qconfig, which contains information about what kind
# of observers to attach. Use 'fbgemm' for server inference and
# 'qnnpack' for mobile inference. Other quantization configurations such
# as selecting symmetric or assymetric quantization and MinMax or L2Norm
# calibration techniques can be specified here.
model_fp32.qconfig = torch.quantization.get_default_qconfig('fbgemm')

# Fuse the activations to preceding layers, where applicable.
# This needs to be done manually depending on the model architecture.
# Common fusions include `conv + relu` and `conv + batchnorm + relu`
model_fp32_fused = torch.quantization.fuse_modules(model_fp32, [['conv', 'relu']])

# Prepare the model for static quantization. This inserts observers in
# the model that will observe activation tensors during calibration.
model_fp32_prepared = torch.quantization.prepare(model_fp32_fused)

# calibrate the prepared model to determine quantization parameters for activations
# in a real world setting, the calibration would be done with a representative dataset
input_fp32 = torch.randn(4, 1, 4, 4)
model_fp32_prepared(input_fp32)

# Convert the observed model to a quantized model. This does several things:
# quantizes the weights, computes and stores the scale and bias value to be
# used with each activation tensor, and replaces key operators with quantized
# implementations.
model_int8 = torch.quantization.convert(model_fp32_prepared)

# run the model, relevant calculations will happen in int8
res = model_int8(input_fp32)

• https://pytorch.org/docs/stable/quantization.html#post-training-static-quantization

from torch.quantization import QuantStub, DeQuantStub

def _make_divisible(v, divisor, min_value=None):
    """
    This function is taken from the original tf repo.
    It ensures that all layers have a channel number that is divisible by 8
    It can be seen here:
    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
    :param v:
    :param divisor:
    :param min_value:
    :return:
    """
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    # Make sure that round down does not go down by more than 10%.
    if new_v < 0.9 * v:
        new_v += divisor
    return new_v


class ConvBNReLU(nn.Sequential):
    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1):
        padding = (kernel_size - 1) // 2
        super(ConvBNReLU, self).__init__(
            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
            nn.BatchNorm2d(out_planes, momentum=0.1),
            # Replace with ReLU
            nn.ReLU(inplace=False)
        )


class InvertedResidual(nn.Module):
    def __init__(self, inp, oup, stride, expand_ratio):
        super(InvertedResidual, self).__init__()
        self.stride = stride
        assert stride in [1, 2]

        hidden_dim = int(round(inp * expand_ratio))
        self.use_res_connect = self.stride == 1 and inp == oup

        layers = []
        if expand_ratio != 1:
            # pw
            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
        layers.extend([
            # dw
            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
            # pw-linear
            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
            nn.BatchNorm2d(oup, momentum=0.1),
        ])
        self.conv = nn.Sequential(*layers)
        # Replace torch.add with floatfunctional
        self.skip_add = nn.quantized.FloatFunctional()

    def forward(self, x):
        if self.use_res_connect:
            return self.skip_add.add(x, self.conv(x))
        else:
            return self.conv(x)


class MobileNetV2(nn.Module):
    def __init__(self, num_classes=1000, width_mult=1.0, inverted_residual_setting=None, round_nearest=8):
        """
        MobileNet V2 main class
        Args:
            num_classes (int): Number of classes
            width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
            inverted_residual_setting: Network structure
            round_nearest (int): Round the number of channels in each layer to be a multiple of this number
            Set to 1 to turn off rounding
        """
        super(MobileNetV2, self).__init__()
        block = InvertedResidual
        input_channel = 32
        last_channel = 1280

        if inverted_residual_setting is None:
            inverted_residual_setting = [
                # t, c, n, s
                [1, 16, 1, 1],
                [6, 24, 2, 2],
                [6, 32, 3, 2],
                [6, 64, 4, 2],
                [6, 96, 3, 1],
                [6, 160, 3, 2],
                [6, 320, 1, 1],
            ]

        # only check the first element, assuming user knows t,c,n,s are required
        if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
            raise ValueError("inverted_residual_setting should be non-empty "
                             "or a 4-element list, got {}".format(inverted_residual_setting))

        # building first layer
        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
        self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
        features = [ConvBNReLU(3, input_channel, stride=2)]
        # building inverted residual blocks
        for t, c, n, s in inverted_residual_setting:
            output_channel = _make_divisible(c * width_mult, round_nearest)
            for i in range(n):
                stride = s if i == 0 else 1
                features.append(block(input_channel, output_channel, stride, expand_ratio=t))
                input_channel = output_channel
        # building last several layers
        features.append(ConvBNReLU(input_channel, self.last_channel, kernel_size=1))
        # make it nn.Sequential
        self.features = nn.Sequential(*features)
        self.quant = QuantStub()
        self.dequant = DeQuantStub()
        # building classifier
        self.classifier = nn.Sequential(
            nn.Dropout(0.2),
            nn.Linear(self.last_channel, num_classes),
        )

        # weight initialization
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out')
                if m.bias is not None:
                    nn.init.zeros_(m.bias)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.ones_(m.weight)
                nn.init.zeros_(m.bias)
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 0.01)
                nn.init.zeros_(m.bias)

    def forward(self, x):

        x = self.quant(x)

        x = self.features(x)
        x = x.mean([2, 3])
        x = self.classifier(x)
        x = self.dequant(x)
        return x

    # Fuse Conv+BN and Conv+BN+Relu modules prior to quantization
    # This operation does not change the numerics
    def fuse_model(self):
        for m in self.modules():
            if type(m) == ConvBNReLU:
                torch.quantization.fuse_modules(m, ['0', '1', '2'], inplace=True)
            if type(m) == InvertedResidual:
                for idx in range(len(m.conv)):
                    if type(m.conv[idx]) == nn.Conv2d:
                        torch.quantization.fuse_modules(m.conv, [str(idx), str(idx + 1)], inplace=True)

• 在forward中添加了量化和反量化函数
• 将浮点模型的加法x + self.conv(x)替换为skip_add.add(x, self.conv(x))

3. pytorch量化-Quantization aware quantization

• 系统自动选择最合适的scale （标度）和 zero_point（零点位置），不需要自定义。但这是一种伪量化，量化后的模型权重仍然是32位浮点数，但大小和8位定点数权重的大小相同。伪量化后的模型可以进行训练。虽然是以32位浮点数进行的训练，但结果与8位定点数的结果一致。

import torch

# define a floating point model where some layers could benefit from QAT
class M(torch.nn.Module):
    def __init__(self):
        super().__init__()
        # QuantStub converts tensors from floating point to quantized
        self.quant = torch.quantization.QuantStub()
        self.conv = torch.nn.Conv2d(1, 1, 1)
        self.bn = torch.nn.BatchNorm2d(1)
        self.relu = torch.nn.ReLU()
        # DeQuantStub converts tensors from quantized to floating point
        self.dequant = torch.quantization.DeQuantStub()

    def forward(self, x):
        x = self.quant(x)
        x = self.conv(x)
        x = self.bn(x)
        x = self.relu(x)
        x = self.dequant(x)
        return x

# create a model instance
model_fp32 = M()

# model must be set to eval for fusion to work
model_fp32.eval()

# attach a global qconfig, which contains information about what kind
# of observers to attach. Use 'fbgemm' for server inference and
# 'qnnpack' for mobile inference. Other quantization configurations such
# as selecting symmetric or assymetric quantization and MinMax or L2Norm
# calibration techniques can be specified here.
model_fp32.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')

# fuse the activations to preceding layers, where applicable
# this needs to be done manually depending on the model architecture
model_fp32_fused = torch.quantization.fuse_modules(model_fp32,
    [['conv', 'bn', 'relu']])

# Prepare the model for QAT. This inserts observers and fake_quants in
# the model needs to be set to train for QAT logic to work
# the model that will observe weight and activation tensors during calibration.
model_fp32_prepared = torch.quantization.prepare_qat(model_fp32_fused.train())

# run the training loop (not shown)
training_loop(model_fp32_prepared)

# Convert the observed model to a quantized model. This does several things:
# quantizes the weights, computes and stores the scale and bias value to be
# used with each activation tensor, fuses modules where appropriate,
# and replaces key operators with quantized implementations.
model_fp32_prepared.eval()
model_int8 = torch.quantization.convert(model_fp32_prepared)

# run the model, relevant calculations will happen in int8
res = model_int8(input_fp32)

• https://pytorch.org/docs/stable/quantization.html#quantization-aware-training-for-static-quantization

6. 算子融合

• 算子拆解、算子聚合、算子重建，以便达到在硬件设备上更好的性能。
• 比如常见Conv+ReLu的两个算子，因为Conv需要做大量卷积计算，需要密集的计算单元支持，而Relu几乎不需要计算，如果Relu算子单独运算，则不仅需要一个计算单元支持其实不需要怎么计算的算子，同时又要对前端的数据进行一次读操作，很浪费资源和增加I/O操作； 此时，可以将Conv和Relu合并融合成一个算子，可以节省I/O访问和带宽开销，也可以节省计算单元。

7. 计算图优化

• 在计算图中，存在某些算子是串行依赖，而某些算子是不依赖性；这些相互独立的子计算图，就可以进行并行计算，提高推理速度，这就是计算图的调度。

8. 底层优化

.1. 内存优化

• 静态内存分配：比如一些固定的算子在整个计算图中都会使用，此时需要再模型初始化时一次性申请完内存空间，在实际推理时不需要频繁申请操作，提高性能
• 动态内存分配：对于中间临时的内存需求，可以进行临时申请和释放，节省内存使用，提高模型并发能力
• 内存复用：对于同一类同一个大小的内存形式，又满足临时性，可以复用内存地址，减少内存申请。