Darknet53原理
Darknet53是一个卷积神经网络模型，在2018年由Joseph Redmon在论文"YOLOv3: An Incremental Improvement"中提出，用于目标检测和分类任务。它是YOLOv3的核心网络模型，其设计思路是通过堆叠多个卷积和残差连接层来提高特征提取的效果。
Darknet53包含53个卷积层和5个max-pooling层组成。Darknet53的结构可以被划分为三组：前段主要由卷积层和max-pooling层组成，中段主要由残差块组成，后段主要由全局平均池化层和全连接层组成。
具体来说，前段的7个卷积层每层都是卷积核大小为3x3，步长为1x1的卷积，接着是5个max-pooling层。这部分的目的是将图像的空间信息不断缩小并增加深度。
接下来是中段的残差块，由于YOLOv3的作者发现，在上述设计下，YOLOv3的学习非常的慢，于是便提出了一种名为"残差块（Residual Block）"的结构。残差块类似于一个短路，它可以让神经网络学习到“残差”，即当前输入和期望输出之间的差异。它使用残差结构来加速收敛、减少梯度消失的问题，以及提高精度。这种结构使得网络能够更好地学习输入图像中的特征，并且增强了模型的表达能力。网络框架图如下：
Darknet53的输入为416x416像素大小的RGB图像，经过一系列卷积操作后，输出为13x13x1024的张量。这个张量可以传递给YOLOv3中的检测头，用于检测图像中的物体。
Darknet53网络框架的特点包括：
模型较浅：相比其他深度神经网络框架，Darknet53只有53层，因此训练速度快，性能高。 网络深度可扩展：使用残差结构，可以轻松地增加更多的卷积层来提高网络的性能。 参数量小：Darknet53的参数量比其他网络少很多，尤其是与ResNet等模型相比，这使得模型更快，更容易训练和优化。 高效推理：Darknet53使用低精度计算和并行计算技术，使得模型的推理速度很快，并且可以在低功耗设备上执行。 适用于大量类别的物体分类：Darknet53的优势在于学习图像特征，并可以高效地解决大量类别的物体分类问题。
总的来说，Darknet53是一个非常有效的卷积神经网络，具有较高的准确性和效率，成为了现代计算机视觉任务中广泛使用的骨干网络之一，并且它已被广泛应用于计算机视觉领域的各种任务，包括图像分类、目标检测、人脸识别等。
Darknet53源码（torch版）
数据集运行代码时自动下载，如果网络比较慢，可以自行点击我分享的链接下载cifar数据集。
链接：百度网盘 提取码：kd9a
此代码是使用的GPU运行的，如果没有GPU导致运行失败，就把代码中的device、.to(device)删除，使用默认CPU运行。
如果GPU显存小导致运行报错，就将主函数main()里面的batch_size调小即可。
import torch
import torch.nn as nn
from torch.nn import modules
from torch.utils.data import DataLoader
from torchvision.datasets import CIFAR10
from torchvision.transforms import transforms
from torch.autograd import Variable
class ConvBnLeaky(modules.Module):
def __init__(self,in_it,out_it,kernels,padding = 0,strides = 1):
super(ConvBnLeaky, self).__init__()
self.convs = nn.Sequential(
nn.Conv2d(in_it,out_it,kernels,padding=padding,stride=strides),
nn.BatchNorm2d(out_it),
nn.LeakyReLU(0.1,True)
)
def forward(self,x):
x = self.convs(x)
return x
pass
class Resnet_Block(modules.Module):
def __init__(self,ch,num_block = 1):
super(Resnet_Block, self).__init__()
self.module_list = modules.ModuleList()
for _ in range(num_block):
resblock = nn.Sequential(
ConvBnLeaky(ch,ch // 2,1),
ConvBnLeaky(ch // 2,ch,kernels=3,padding=1)
)
self.module_list.append(resblock)
def forward(self,x):
for i in self.module_list:
x = i(x) + x
return x
class Darknet_53(modules.Module):
def __init__(self,num_classes = 10):
super(Darknet_53, self).__init__()
self.later_1 = nn.Sequential(
ConvBnLeaky(3,32,3,padding=1),
ConvBnLeaky(32,64,3,padding=1,strides=2),
Resnet_Block(64,1)
)
self.later_2 = nn.Sequential(
ConvBnLeaky(64, 128, 3, padding=1,strides=2),
Resnet_Block(128, 2)
)
self.later_3 = nn.Sequential(
ConvBnLeaky(128, 256, 3, padding=1, strides=2),
Resnet_Block(256, 8)
)
self.later_4 = nn.Sequential(
ConvBnLeaky(256, 512, 3, padding=1, strides=2),
Resnet_Block(512, 8)
)
self.later_5 = nn.Sequential(
ConvBnLeaky(512, 1024, 3, padding=1, strides=2),
Resnet_Block(1024, 4)
)
self.pool = nn.AdaptiveAvgPool2d((1,1))
self.fc = nn.Linear(1024,num_classes)
def forward(self,x):
x = self.later_1(x)
x = self.later_2(x)
x = self.later_3(x)
x = self.later_4(x)
x = self.later_5(x)
x = self.pool(x)
x = torch.squeeze(x)
x = self.fc(x)
return x
def main():
Root_file = 'cifar'
train_data = CIFAR10(Root_file,train=True,transform = transforms.ToTensor())
data = DataLoader(train_data,batch_size=64,shuffle=True)
device = torch.device('cuda')
net = Darknet_53().to(device)
print(net)
Cross = nn.CrossEntropyLoss().to(device)
optimizer = torch.optim.Adam(net.parameters(),0.001)
for epoch in range(10):
for img,label in data:
img = Variable(img).to(device)
label = Variable(label).to(device)
output = net.forward(img)
loss = Cross(output,label)
loss.backward()
optimizer.zero_grad()
optimizer.step()
pre = torch.argmax(output,1)
num = (pre == label).sum().item()
acc = num / img.shape[0]
loss_val = loss.item()
print("epoch:",epoch)
print("acc:",acc)
print("loss:",loss_val)
pass
if __name__ == '__main__':
main()
训练10个epoch的效果
效果图
框架图
Darknet_53( (later_1): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)) (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (2): Resnet_Block( (module_list): ModuleList( (0): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) ) ) ) (later_2): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): Resnet_Block( (module_list): ModuleList( (0): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (1): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) ) ) ) (later_3): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): Resnet_Block( (module_list): ModuleList( (0): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (1): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (2): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (3): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (4): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (5): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (6): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (7): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) ) ) ) (later_4): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): Resnet_Block( (module_list): ModuleList( (0): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (1): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (2): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (3): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (4): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (5): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (6): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (7): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) ) ) ) (later_5): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)) (1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): Resnet_Block( (module_list): ModuleList( (0): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (1): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (2): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) (3): Sequential( (0): ConvBnLeaky( (convs): Sequential( (0): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1)) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) (1): ConvBnLeaky( (convs): Sequential( (0): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (2): LeakyReLU(negative_slope=0.1, inplace=True) ) ) ) ) ) ) (pool): AdaptiveAvgPool2d(output_size=(1, 1)) (fc): Linear(in_features=1024, out_features=10, bias=True) )
推荐链接
您阅读本篇文章共花了：
发表评论