Target detection - pytorch uses mobilenet series (V1, V2, V3) to build yolov4 target detection platform

Learn foreword

Let's see how to use mobilenet Series construction yolov4 Target detection platform .

Source code download

https://github.com/bubbliiiing/mobilenet-yolov4-pytorch

Network replacement implementation ideas

1、 Network structure analysis and replacement thinking analysis

about YoloV4 for , The whole network structure can be divided into three parts .
Namely ：
1、 Backbone feature extraction network Backbone, Corresponding to... On the image CSPdarknet53
2、 Strengthen the feature extraction network , Corresponding to... On the image SPP and PANet
3、 Prediction network YoloHead, Use the obtained features to predict

among ：
The first part Backbone feature extraction network The function of is to Preliminary feature extraction , Extract network information using backbone features , We can get three Preliminary effective characteristic layer .
The second part Strengthen the feature extraction network The function of is to Enhanced feature extraction , Using enhanced feature extraction network , We can talk about three Preliminary effective characteristic layer Feature fusion , Extract better features , Get three More effective feature layer .
The third part Prediction network The function of is to use more effective special layer to obtain prediction results .

In these three parts , The first 1 Section and section 2 Parts can be modified more easily . The first 3 Some modifiable contents are not big , After all, it's just 3x3 Convolution sum 1x1 The combination of convolutions .

mobilenet A series of networks can be used for classification , Its main part is used for feature extraction , We can use mobilenet Series network replaces yolov4 In the middle of CSPdarknet53 Feature extraction , Will three Preliminary effective characteristic layer identical shape Feature layer to enhance feature extraction , Can will mobilenet Series replacement into yolov4 In the middle .

2、mobilenet Series network introduction

This paper shares three backbone feature extraction networks , Namely mobilenetV1、mobilenetV2、mobilenetV3.

a、mobilenetV1 Introduce

MobileNet The model is Google A lightweight deep neural network is proposed for embedded devices such as mobile phones , The core idea of its use is depthwise separable convolution（ Deep separable convolution block ）.

For a convolution point ：
Let's say I have a 3×3 The size of the convolution layer , Its input channel is 16、 The output channel is 32. Specific for ,32 individual 3×3 The size of the convolution kernel will traverse 16 Each data in one channel , Finally, the required 32 Output channels , The required parameters are 16×32×3×3=4608 individual .

Application depth separable convolution structure block , use 16 individual 3×3 The convolution kernels of size traverse 16 Channel data , Got it 16 A feature map . Before the fusion operation , Then use 32 individual 1×1 The size of the convolution kernel traverses this 16 A feature map , The required parameters are 16×3×3+16×32×1×1=656 individual .
You can see it depthwise separable convolution It can reduce the parameters of the model .

The following picture is depthwise separable convolution Structure

When building the model , You can convolute group Set to in_filters The layer realizes deep separable convolution , And then reuse it 1x1 Convolution adjustment channels Count .

Commonly understood as 3x3 The convolution kernel is only one layer thick , Then slide on the input tensor layer by layer , Each convolution generates an output channel , When the convolution is complete , Using 1x1 Convolution adjustment thickness of .

Here is MobileNet Structure , among Conv dw It's layered convolution , After that, there will be one 1x1 Convolution for channel processing ,

The picture above shows mobilenetV1-1 Structure , Because I can't find pytorch Of mobilenetv1 Weight resources for , I only mobilenetV1-0.25 The weight of , So this article uses mobilenetV1 The version is mobilenetV1-0.25.

mobilenetV1-0.25 yes mobilenetV1-1 The number of channels is compressed to the original 1/4 Network of .
about yolov4 Speaking of , We need to take out its last three shape Effective feature layer to enhance feature extraction .

In the code , We took out out1、out2、out3.

import time
import torch
import torch.nn as nn
import torchvision.models._utils as _utils
import torchvision.models as models
import torch.nn.functional as F
from torch.autograd import Variable

def conv_bn(inp, oup, stride = 1):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
        nn.BatchNorm2d(oup),
        nn.ReLU6(inplace=True)
    )
    
def conv_dw(inp, oup, stride = 1):
    return nn.Sequential(
        nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=False),
        nn.BatchNorm2d(inp),
        nn.ReLU6(inplace=True),

        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
        nn.BatchNorm2d(oup),
        nn.ReLU6(inplace=True),
    )

class MobileNetV1(nn.Module):
    def __init__(self):
        super(MobileNetV1, self).__init__()
        self.stage1 = nn.Sequential(
            # 640,640,3 -> 320,320,32
            conv_bn(3, 32, 2),
            # 320,320,32 -> 320,320,64
            conv_dw(32, 64, 1), 

            # 320,320,64 -> 160,160,128
            conv_dw(64, 128, 2),
            conv_dw(128, 128, 1),

            # 160,160,128 -> 80,80,256
            conv_dw(128, 256, 2),
            conv_dw(256, 256, 1), 
        )
            # 80,80,256 -> 40,40,512
        self.stage2 = nn.Sequential(
            conv_dw(256, 512, 2),
            conv_dw(512, 512, 1),
            conv_dw(512, 512, 1),
            conv_dw(512, 512, 1), 
            conv_dw(512, 512, 1),
            conv_dw(512, 512, 1),
        )
            # 40,40,512 -> 20,20,1024
        self.stage3 = nn.Sequential(
            conv_dw(512, 1024, 2),
            conv_dw(1024, 1024, 1),
        )
        self.avg = nn.AdaptiveAvgPool2d((1,1))
        self.fc = nn.Linear(1024, 1000)

    def forward(self, x):
        x = self.stage1(x)
        x = self.stage2(x)
        x = self.stage3(x)
        x = self.avg(x)
        # x = self.model(x)
        x = x.view(-1, 1024)
        x = self.fc(x)
        return x

def mobilenet_v1(pretrained=False, progress=True):
    model = MobileNetV1()
    if pretrained:
        print("mobilenet_v1 has no pretrained model")
    return model

if __name__ == "__main__":
    import torch
    from torchsummary import summary

    #  Need to use device To specify the network at GPU still CPU function 
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model = mobilenet_v1().to(device)
    summary(model, input_size=(3, 416, 416))

b、mobilenetV2 Introduce

MobileNetV2 yes MobileNet Upgraded version , It has a very important feature that it uses Inverted resblock, Whole mobilenetv2 All by Inverted resblock form .

Inverted resblock It can be divided into two parts ：
On the left is the trunk , The first use of 1x1 Convolution is used to increase the dimension , And then use it 3x3 Depth separable convolution for feature extraction , And then reuse it 1x1 Convolution dimensionality reduction .
On the right is the residual side , The input and output are directly connected .

The overall network structure is as follows ：（ among Inverted resblock The operation is the above structure ）

about yolov4 Speaking of , We need to take out its last three shape Effective feature layer to enhance feature extraction .

In the code , We took out out1、out2、out3.

from torch import nn
from torchvision.models.utils import load_state_dict_from_url

model_urls = {
    
    'mobilenet_v2': 'https://download.pytorch.org/models/mobilenet_v2-b0353104.pth',
}


def _make_divisible(v, divisor, min_value=None):
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    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),
            nn.ReLU6(inplace=True)
        )

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:
            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1))
        layers.extend([
            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim),
            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
            nn.BatchNorm2d(oup),
        ])
        self.conv = nn.Sequential(*layers)

    def forward(self, x):
        if self.use_res_connect:
            return 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):
        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],
            ]

        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))

        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)]

        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

        features.append(ConvBNReLU(input_channel, self.last_channel, kernel_size=1))
        self.features = nn.Sequential(*features)

        self.classifier = nn.Sequential(
            nn.Dropout(0.2),
            nn.Linear(self.last_channel, num_classes),
        )

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

def mobilenet_v2(pretrained=False, progress=True):
    model = MobileNetV2()
    if pretrained:
        state_dict = load_state_dict_from_url(model_urls['mobilenet_v2'], model_dir="model_data",
                                              progress=progress)
        model.load_state_dict(state_dict)

    return model

if __name__ == "__main__":
    print(mobilenet_v2())

c、mobilenetV3 Introduce

mobilenetV3 Used a special bneck structure .

bneck The structure is shown in the following figure ：

It combines the following four features ：
a、MobileNetV2 Inverse residual structure with linear bottleneck (the inverted residual with linear bottleneck).

That is, first use 1x1 Convolution carries out ascending dimension , Then perform the following operations , And has residual edges .

b、MobileNetV1 The depth separable convolution （depthwise separable convolutions）.

In the input 1x1 After the convolution is carried out to raise the dimension , Conduct 3x3 Depth separates the convolution .

c、 Lightweight attention model .

This attention mechanism works by adjusting the weight of each channel .

d、 utilize h-swish Instead of swish function .
Used in the structure h-swishj Activation function , Instead of swish function , Reduce the amount of computation , Improve performance .

The following figure shows the whole mobilenetV3 Structure diagram ：

How to understand this watch ？ We start from each train ：
First column Input representative mobilenetV3 Of each feature layer shape change ;
Second column Operator Represents what each feature layer will experience block structure , We can see in the MobileNetV3 in , Feature extraction has gone through a lot of bneck structure ;
Third 、 The four columns represent bneck The number of channels after the rise of the internal inverse residual structure 、 Input to bneck The number of channels in the feature layer .
The fifth column SE It represents whether to introduce attention mechanism in this layer .
The sixth column NL Represents the type of activation function ,HS representative h-swish,RE representative RELU.
The seventh column s Represents every time block The step size used by the structure .

about yolov4 Speaking of , We need to take out its last three shape Effective feature layer to enhance feature extraction .

In the code , We took out out1、out2、out3.

import torch.nn as nn
import math
import torch
def _make_divisible(v, divisor, min_value=None):
    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 h_sigmoid(nn.Module):
    def __init__(self, inplace=True):
        super(h_sigmoid, self).__init__()
        self.relu = nn.ReLU6(inplace=inplace)

    def forward(self, x):
        return self.relu(x + 3) / 6


class h_swish(nn.Module):
    def __init__(self, inplace=True):
        super(h_swish, self).__init__()
        self.sigmoid = h_sigmoid(inplace=inplace)

    def forward(self, x):
        return x * self.sigmoid(x)


class SELayer(nn.Module):
    def __init__(self, channel, reduction=4):
        super(SELayer, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
                nn.Linear(channel, _make_divisible(channel // reduction, 8)),
                nn.ReLU(inplace=True),
                nn.Linear(_make_divisible(channel // reduction, 8), channel),
                h_sigmoid()
        )

    def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        return x * y


def conv_3x3_bn(inp, oup, stride):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
        nn.BatchNorm2d(oup),
        h_swish()
    )


def conv_1x1_bn(inp, oup):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
        nn.BatchNorm2d(oup),
        h_swish()
    )


class InvertedResidual(nn.Module):
    def __init__(self, inp, hidden_dim, oup, kernel_size, stride, use_se, use_hs):
        super(InvertedResidual, self).__init__()
        assert stride in [1, 2]

        self.identity = stride == 1 and inp == oup

        if inp == hidden_dim:
            self.conv = nn.Sequential(
                # dw
                nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, (kernel_size - 1) // 2, groups=hidden_dim, bias=False),
                nn.BatchNorm2d(hidden_dim),
                h_swish() if use_hs else nn.ReLU(inplace=True),
                # Squeeze-and-Excite
                SELayer(hidden_dim) if use_se else nn.Identity(),
                # pw-linear
                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup),
            )
        else:
            self.conv = nn.Sequential(
                # pw
                nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
                nn.BatchNorm2d(hidden_dim),
                h_swish() if use_hs else nn.ReLU(inplace=True),
                # dw
                nn.Conv2d(hidden_dim, hidden_dim, kernel_size, stride, (kernel_size - 1) // 2, groups=hidden_dim, bias=False),
                nn.BatchNorm2d(hidden_dim),
                # Squeeze-and-Excite
                SELayer(hidden_dim) if use_se else nn.Identity(),
                h_swish() if use_hs else nn.ReLU(inplace=True),
                # pw-linear
                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup),
            )

    def forward(self, x):
        if self.identity:
            return x + self.conv(x)
        else:
            return self.conv(x)


class MobileNetV3(nn.Module):
    def __init__(self, num_classes=1000, width_mult=1.):
        super(MobileNetV3, self).__init__()
        # setting of inverted residual blocks
        self.cfgs = [
            #` k, t, c, SE, HS, s 
            [3,   1,  16, 0, 0, 1],
            [3,   4,  24, 0, 0, 2],
            [3,   3,  24, 0, 0, 1],
            [5,   3,  40, 1, 0, 2],
            [5,   3,  40, 1, 0, 1],
            [5,   3,  40, 1, 0, 1],
            [3,   6,  80, 0, 1, 2],
            [3, 2.5,  80, 0, 1, 1],
            [3, 2.3,  80, 0, 1, 1],
            [3, 2.3,  80, 0, 1, 1],
            [3,   6, 112, 1, 1, 1],
            [3,   6, 112, 1, 1, 1],
            [5,   6, 160, 1, 1, 2],
            [5,   6, 160, 1, 1, 1],
            [5,   6, 160, 1, 1, 1]
        ]

        input_channel = _make_divisible(16 * width_mult, 8)
        layers = [conv_3x3_bn(3, input_channel, 2)]

        block = InvertedResidual
        for k, t, c, use_se, use_hs, s in self.cfgs:
            output_channel = _make_divisible(c * width_mult, 8)
            exp_size = _make_divisible(input_channel * t, 8)
            layers.append(block(input_channel, exp_size, output_channel, k, s, use_se, use_hs))
            input_channel = output_channel
        self.features = nn.Sequential(*layers)

        self.conv = conv_1x1_bn(input_channel, exp_size)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        output_channel = _make_divisible(1280 * width_mult, 8) if width_mult > 1.0 else 1280
        self.classifier = nn.Sequential(
            nn.Linear(exp_size, output_channel),
            h_swish(),
            nn.Dropout(0.2),
            nn.Linear(output_channel, num_classes),
        )

        self._initialize_weights()

    def forward(self, x):
        x = self.features(x)
        x = self.conv(x)
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.classifier(x)
        return x

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                n = m.weight.size(1)
                m.weight.data.normal_(0, 0.01)
                m.bias.data.zero_()

def mobilenet_v3(pretrained=False, **kwargs):
    model = MobileNetV3(**kwargs)
    if pretrained:
        state_dict = torch.load('./model_data/mobilenetv3-large-1cd25616.pth')
        model.load_state_dict(state_dict, strict=True)
    return model

3、 Integrate forecast results into yolov4 In the Internet

about yolov4 Speaking of , We need to use the backbone feature to extract the information obtained by the network Three effective features are used to strengthen the construction of feature pyramid .

Use the... Defined in the previous step MobilenetV1、MobilenetV2、MobilenetV3 Three functions, we can get each Mobilenet Three effective feature layers corresponding to the network .

We can use these three effective feature layers to replace the original yolov4 Backbone network CSPdarknet53 Effective feature layer .

In order to further reduce the number of parameters , We can use deep separable convolution instead of yoloV3 The common convolution used in .

The implementation code is as follows ：

import torch
import torch.nn as nn
from collections import OrderedDict
from nets.mobilenet_v1 import mobilenet_v1
from nets.mobilenet_v2 import mobilenet_v2
from nets.mobilenet_v3 import mobilenet_v3

class MobileNetV1(nn.Module):
    def __init__(self, pretrained = False):
        super(MobileNetV1, self).__init__()
        self.model = mobilenet_v1(pretrained=pretrained)

    def forward(self, x):
        out3 = self.model.stage1(x)
        out4 = self.model.stage2(out3)
        out5 = self.model.stage3(out4)
        return out3, out4, out5

class MobileNetV2(nn.Module):
    def __init__(self, pretrained = False):
        super(MobileNetV2, self).__init__()
        self.model = mobilenet_v2(pretrained=pretrained)

    def forward(self, x):
        out3 = self.model.features[:7](x)
        out4 = self.model.features[7:14](out3)
        out5 = self.model.features[14:18](out4)
        return out3, out4, out5

class MobileNetV3(nn.Module):
    def __init__(self, pretrained = False):
        super(MobileNetV3, self).__init__()
        self.model = mobilenet_v3(pretrained=pretrained)

    def forward(self, x):
        out3 = self.model.features[:7](x)
        out4 = self.model.features[7:13](out3)
        out5 = self.model.features[13:16](out4)
        return out3, out4, out5

def conv2d(filter_in, filter_out, kernel_size, groups=1, stride=1):
    pad = (kernel_size - 1) // 2 if kernel_size else 0
    return nn.Sequential(OrderedDict([
        ("conv", nn.Conv2d(filter_in, filter_out, kernel_size=kernel_size, stride=stride, padding=pad, groups=groups, bias=False)),
        ("bn", nn.BatchNorm2d(filter_out)),
        ("relu", nn.ReLU6(inplace=True)),
    ]))

def conv_dw(filter_in, filter_out, stride = 1):
    return nn.Sequential(
        nn.Conv2d(filter_in, filter_in, 3, stride, 1, groups=filter_in, bias=False),
        nn.BatchNorm2d(filter_in),
        nn.ReLU6(inplace=True),

        nn.Conv2d(filter_in, filter_out, 1, 1, 0, bias=False),
        nn.BatchNorm2d(filter_out),
        nn.ReLU6(inplace=True),
    )

#---------------------------------------------------#
# SPP structure , Pool using pool cores of different sizes 
#  Stack after pooling 
#---------------------------------------------------#
class SpatialPyramidPooling(nn.Module):
    def __init__(self, pool_sizes=[5, 9, 13]):
        super(SpatialPyramidPooling, self).__init__()

        self.maxpools = nn.ModuleList([nn.MaxPool2d(pool_size, 1, pool_size//2) for pool_size in pool_sizes])

    def forward(self, x):
        features = [maxpool(x) for maxpool in self.maxpools[::-1]]
        features = torch.cat(features + [x], dim=1)

        return features

#---------------------------------------------------#
#  Convolution  +  On the sampling 
#---------------------------------------------------#
class Upsample(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(Upsample, self).__init__()

        self.upsample = nn.Sequential(
            conv2d(in_channels, out_channels, 1),
            nn.Upsample(scale_factor=2, mode='nearest')
        )

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

#---------------------------------------------------#
#  Cubic convolution block 
#---------------------------------------------------#
def make_three_conv(filters_list, in_filters):
    m = nn.Sequential(
        conv2d(in_filters, filters_list[0], 1),
        conv_dw(filters_list[0], filters_list[1]),
        conv2d(filters_list[1], filters_list[0], 1),
    )
    return m

#---------------------------------------------------#
#  Quintic convolution block 
#---------------------------------------------------#
def make_five_conv(filters_list, in_filters):
    m = nn.Sequential(
        conv2d(in_filters, filters_list[0], 1),
        conv_dw(filters_list[0], filters_list[1]),
        conv2d(filters_list[1], filters_list[0], 1),
        conv_dw(filters_list[0], filters_list[1]),
        conv2d(filters_list[1], filters_list[0], 1),
    )
    return m

#---------------------------------------------------#
#  At last get yolov4 Output 
#---------------------------------------------------#
def yolo_head(filters_list, in_filters):
    m = nn.Sequential(
        conv_dw(in_filters, filters_list[0]),
        
        nn.Conv2d(filters_list[0], filters_list[1], 1),
    )
    return m

#---------------------------------------------------#
# yolo_body
#---------------------------------------------------#
class YoloBody(nn.Module):
    def __init__(self, num_anchors, num_classes, backbone="mobilenetv2", pretrained=False):
        super(YoloBody, self).__init__()
        # backbone
        if backbone == "mobilenetv1":
            self.backbone = MobileNetV1(pretrained=pretrained)
            alpha = 1
            in_filters = [256,512,1024]
        elif backbone == "mobilenetv2":
            self.backbone = MobileNetV2(pretrained=pretrained)
            alpha = 1
            in_filters = [32,96,320]
        elif backbone == "mobilenetv3":
            self.backbone = MobileNetV3(pretrained=pretrained)
            alpha = 1
            in_filters = [40,112,160]
        else:
            raise ValueError('Unsupported backbone - `{}`, Use mobilenetv1, mobilenetv2, mobilenetv3.'.format(backbone))

        self.conv1           = make_three_conv([int(512*alpha), int(1024*alpha)], in_filters[2])
        self.SPP             = SpatialPyramidPooling()
        self.conv2           = make_three_conv([int(512*alpha), int(1024*alpha)], int(2048*alpha))

        self.upsample1       = Upsample(int(512*alpha), int(256*alpha))
        self.conv_for_P4     = conv2d(in_filters[1], int(256*alpha),1)
        self.make_five_conv1 = make_five_conv([int(256*alpha), int(512*alpha)], int(512*alpha))

        self.upsample2       = Upsample(int(256*alpha), int(128*alpha))
        self.conv_for_P3     = conv2d(in_filters[0], int(128*alpha),1)
        self.make_five_conv2 = make_five_conv([ int(128*alpha), int(256*alpha)], int(256*alpha))
        # 3*(5+num_classes)=3*(5+20)=3*(4+1+20)=75
        # 4+1+num_classes
        final_out_filter2    = num_anchors * (5 + num_classes)
        self.yolo_head3      = yolo_head([int(256*alpha), final_out_filter2],int(128*alpha))

        self.down_sample1    = conv_dw(int(128*alpha), int(256*alpha),stride=2)
        self.make_five_conv3 = make_five_conv([int(256*alpha), int(512*alpha)],int(512*alpha))
        # 3*(5+num_classes)=3*(5+20)=3*(4+1+20)=75
        final_out_filter1    = num_anchors * (5 + num_classes)
        self.yolo_head2      = yolo_head([int(512*alpha), final_out_filter1], int(256*alpha))


        self.down_sample2    = conv_dw(int(256*alpha), int(512*alpha),stride=2)
        self.make_five_conv4 = make_five_conv([int(512*alpha), int(1024*alpha)], int(1024*alpha))
        # 3*(5+num_classes)=3*(5+20)=3*(4+1+20)=75
        final_out_filter0    = num_anchors * (5 + num_classes)
        self.yolo_head1      = yolo_head([int(1024*alpha), final_out_filter0], int(512*alpha))


    def forward(self, x):
        # backbone
        x2, x1, x0 = self.backbone(x)

        P5 = self.conv1(x0)
        P5 = self.SPP(P5)
        P5 = self.conv2(P5)

        P5_upsample = self.upsample1(P5)
        P4 = self.conv_for_P4(x1)
        P4 = torch.cat([P4,P5_upsample],axis=1)
        P4 = self.make_five_conv1(P4)

        P4_upsample = self.upsample2(P4)
        P3 = self.conv_for_P3(x2)
        P3 = torch.cat([P3,P4_upsample],axis=1)
        P3 = self.make_five_conv2(P3)

        P3_downsample = self.down_sample1(P3)
        P4 = torch.cat([P3_downsample,P4],axis=1)
        P4 = self.make_five_conv3(P4)

        P4_downsample = self.down_sample2(P4)
        P5 = torch.cat([P4_downsample,P5],axis=1)
        P5 = self.make_five_conv4(P5)

        out2 = self.yolo_head3(P3)
        out1 = self.yolo_head2(P4)
        out0 = self.yolo_head1(P5)

        return out0, out1, out2

Train yourself YoloV4 Model

First, go to Github Download the corresponding warehouse , After downloading, use the decompression software to decompress , Then open the folder with programming software .
Note that the open root directory must be correct , Otherwise, if the relative directory is incorrect , The code will not run .
Be sure to note that the root directory after opening is the directory where the files are stored .

One 、 Data set preparation

This article USES the VOC Format for training , Before training, you need to make your own data set , If you don't have your own data set , Can pass Github Connect to download VOC12+07 Try the following data set .
Put the label file in before training VOCdevkit Under folder VOC2007 Under folder Annotation in .

Put the picture file in before training VOCdevkit Under folder VOC2007 Under folder JPEGImages in .

At this point, the placement of the data set has ended .

Two 、 Data set processing

After the data set is placed , We need to do the next step on the dataset , The purpose is to obtain... For training 2007_train.txt as well as 2007_val.txt, You need to use... Under the root directory voc_annotation.py.

voc_annotation.py There are some parameters that need to be set .
Namely annotation_mode、classes_path、trainval_percent、train_percent、VOCdevkit_path, The first training can only modify classes_path

''' annotation_mode Used to specify the contents of the file to be calculated at run time  annotation_mode by 0 Represents the entire label processing process , Including getting VOCdevkit/VOC2007/ImageSets Inside txt And for training 2007_train.txt、2007_val.txt annotation_mode by 1 On behalf of VOCdevkit/VOC2007/ImageSets Inside txt annotation_mode by 2 Stands for... For training 2007_train.txt、2007_val.txt '''
annotation_mode     = 0
'''  It has to be changed , Used to generate 2007_train.txt、2007_val.txt Target information for   And the... Used for training and prediction classes_path It's OK to be consistent   If generated 2007_train.txt There is no target information in it   So it's because classes Not set correctly   Only in annotation_mode by 0 and 2 In force  '''
classes_path        = 'model_data/voc_classes.txt'
''' trainval_percent Is used to specify the ( Training set + Verification set ) Ratio to test set , By default  ( Training set + Verification set ): Test set  = 9:1 train_percent Is used to specify the ( Training set + Verification set ) The ratio of training set to verification set in , By default   Training set : Verification set  = 9:1  Only in annotation_mode by 0 and 1 In force  '''
trainval_percent    = 0.9
train_percent       = 0.9
'''  Point to VOC The folder where the dataset is located   By default, it points to... In the root directory VOC Data sets  '''
VOCdevkit_path  = 'VOCdevkit'

classes_path Used to point to... Corresponding to the detection category txt, With voc Data sets, for example , We use txt by ：

When training your dataset , You can build your own cls_classes.txt, Write the categories you need to distinguish .

3、 ... and 、 Start network training

adopt voc_annotation.py We have generated 2007_train.txt as well as 2007_val.txt, Now we can start training .
There are many training parameters , You can read the notes carefully after downloading the library , The most important part is still train.py Inside classes_path.

classes_path Used to point to... Corresponding to the detection category txt, This txt and voc_annotation.py Inside txt equally ！ Training your own data set must be modified ！

After revising classes_path Then you can run train.py It's training , Training multiple epoch after , Weights will be generated in logs In the folder .

in addition ,backbone Parameter is used to specify the backbone feature extraction network used , Can be in mobilenetv1, mobilenetv2, mobilenetv3 To choose from .

Pay attention to what you use before training mobilenet Alignment of version and pre training weights .

The functions of other parameters are as follows ：

#-------------------------------#
#  Whether to use Cuda
#  No, GPU It can be set to False
#-------------------------------#
Cuda = True
#--------------------------------------------------------#
#  Be sure to modify before training classes_path, Make it correspond to its own data set 
#--------------------------------------------------------#
classes_path    = 'model_data/voc_classes.txt'
#---------------------------------------------------------------------#
# anchors_path Represents the corresponding a priori box txt file , Generally, it is not modified .
# anchors_mask Used to help the code find the corresponding a priori box , Generally, it is not modified .
#---------------------------------------------------------------------#
anchors_path    = 'model_data/yolo_anchors.txt'
anchors_mask    = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
#------------------------------------------------------------------------------------------------------#
#  For the weight file, please see README, Baidu online disk download . The pre training weight of data is universal for different data sets , Because features are universal 
#  The pre training weight is for 99% In all cases, you must use , If not, the weight is too random , The effect of feature extraction is not obvious , The results of online training will not be good .
#  When training your dataset, you will be prompted that the dimensions do not match , The predictions are different, and the natural dimensions don't match 
#  If you want to continue practicing, you will model_path Set to logs The weight files that have been trained under the folder . 
#------------------------------------------------------------------------------------------------------#
model_path      = 'model_data/yolov4_mobilenet_v1_voc.pth'
#------------------------------------------------------#
#  Input shape size , Must be 32 Multiple 
#------------------------------------------------------#
input_shape     = [416, 416]
#-------------------------------#
#  Backbone feature extraction network used 
# mobilenetv1
# mobilenetv2
# mobilenetv3
# ghostnet
#-------------------------------#
backbone        = "mobilenetv1"
#----------------------------------#
#  Whether to use the pre training weight of the backbone network 
#  Only the trunk part , And model_path irrelevant 
#----------------------------------#
pretrained      = False
#------------------------------------------------------#
# Yolov4 Of tricks application 
# mosaic  Mosaic data enhancement  True or False 
#  In the actual test mosaic Data enhancement is not stable , So the default is False
# Cosine_scheduler  Cosine annealing learning rate  True or False
# label_smoothing  Label smoothing  0.01 The following general   Such as 0.01、0.005
#------------------------------------------------------#
mosaic              = False
Cosine_lr           = False
label_smoothing     = 0

Four 、 Prediction of training results

The prediction of training results requires two files , Namely yolo.py and predict.py.
We need to go first yolo.py Modify inside model_path as well as classes_path, These two parameters must be modified .

in addition ,backbone Parameter is used to specify the backbone feature extraction network used , Can be in mobilenetv1, mobilenetv2, mobilenetv3 To choose from .

model_path Point to the trained weight file , stay logs Folder .
classes_path Point to the... Corresponding to the detection category txt.

After the modification, you can run predict.py It's been tested . After running, enter the picture path to detect .