图像识别算法--VGG16

[TOC]

1、参考文献

VGG16:[1]SIMONYAN K, ZISSERMAN A. Very Deep Convolutional Networks for Large-Scale Image Recognition[M/OL]. arXiv, 2015[2023-04-07]. http://arxiv.org/abs/1409.1556.

Dropout:[2]SRIVASTAVA N, HINTON G, KRIZHEVSKY A, 等. Dropout: A Simple Way to Prevent Neural Networks from Overfitting[J].

2、VGG16实践

2.1 VGG16 长处

1、应用3x3的卷积核而非7x7的卷积核

First, we incorporate three non-linear rectification layers instead of a single one, which makes the decision function more discriminative. Second, we decrease the number of parameters.

也就是说VGG16一方面缩小了参数(绝对于7x7),另外一方面通过3非线性层,更加具备非线性表达能力

2.2 VGG16网络结构图

VGG设计的神经网络结构图:

D:dropout

图片变动过程
1:输出咱们的川建国(224x224x3)--->224x224x64--->112x112x64
2:112x112x64--->112x112x128--->56x56x128
3:56x56x128--->56x56x256--->56x56x256--->28x28x256
4:28x28x256--->28x28x512--->28x28x512--->14x14x512
5:14x14x512--->14x14x512--->14x14x512--->7x7x512

变动过程第一个数字代表输出,最初一个数字代表这一层的输入,下一层的输出

全连贯层:
1、输出:7x7x512(25088),输入:4096
2、输出:4096,输入4096
3、输出:4096,输入1000 (因为进行的是1000个分类
在参考文献1中作者同时比拟了上面几种不同构造(VGG11、VGG16与VGG19):


倡议

we have found that our conceptually much simpler scheme already provides a speedup of 3.75 times on an off-the-shelf 4-GPU system, as compared to using a single GPU. On a system equipped with four NVIDIA Titan Black GPUs, training a single net took 2–3 weeks depending on the architecture.

咱们发现,与应用单个GPU相比,咱们在概念上更简略的计划曾经在现成的4 - GPU零碎上提供了3.75倍的减速比。在搭载4个NVIDIA Titan Black GPU的零碎上,依据架构的不同,训练单个网络须要2 ~ 3周。

拜访链接:https://www.image-net.org/challenges/LSVRC/2012/index.php

如果想复现VGG16,间接应用论文作者数据是不要切合实际的:1、数据过大;2、没有这么高的电脑配置。

举荐应用数据集:https://download.pytorch.org/tutorial/hymenoptera_data.zip

url = "https://download.pytorch.org/tutorial/hymenoptera_data.zip"save_path = os.path.join(data_dir, "hymenoptera_data.zip")if not os.path.exists(save_path):    urllib.request.urlretrieve(url, save_path)    zip = zipfile.ZipFile(save_path)    zip.extractall(data_dir)    zip.close()    os.remove(save_path)

pytorch官网数据,次要是实现蜜蜂和蚂蚁分类,不过在应用前必须对图片进行解决,因为他提供的图片并非都是224x224x3,所以须要对图片进行转换。

"""图片预处理:1、图片裁剪2、标准化3、图片旋转"""class ImageTransform():    def __init__(self, resize, mean, std):        self.data_transform = {            'train': transforms.Compose([                transforms.RandomResizedCrop(resize, scale=(0.5, 1.0)),                 #scale在调整大小之前,指定裁剪的随机区域的上限和下限。规模是绝对于原始图像的面积来定义的                transforms.RandomHorizontalFlip(), #以给定的概率随机程度翻转给定的图像                transforms.ToTensor(), #将图片转化为张量                transforms.Normalize(mean, std) #将图片进行正则化            ]),            'val': transforms.Compose([                transforms.Resize(resize), #扭转尺寸                transforms.CenterCrop(resize), #核心裁剪图像                transforms.ToTensor(),                 transforms.Normalize(mean, std)            ])        }    def __call__(self, img, phase='train'):        return self.data_transform[phase](img)

上述代码波及到一个实践:在卷积神经网络中(VGG也是一种卷积神经网络),在对于训练集数据有余的时候,能够尝试对图片进行旋转等操作来补充训练集数据。比方咱们川建国,我旋转他就相当于又减少了一个训练集数据。

如果实验室电脑配置不够:倡议间接租算力(如果只是轻微应用深度学习+实验室没钱)

举荐网站:AutoDL-品质GPU租用平台-租GPU就上AutoDL,学生认证价格也还ok。网站提供GPU(局部):


多尺度评估的试验后果:

作者操作过程中还应用了:1、$L_{2}$范数;2、设置0.5的dropout

2.2.1 复现代码

#定义训练网络 VGG-16import torch.nn.functional as Fclass vgg16(nn.Module):    def __init__(self):        super().__init__()        #开始定义网络结构        self.conv1 = torch.nn.Conv2d(3, 64, 3, padding=(1,1))        self.conv2 = torch.nn.Conv2d(64, 64, 3, padding=(1, 1))        self.pool1 = torch.nn.MaxPool2d((2, 2), padding=(1, 1)) #64x112x112        self.conv3 = torch.nn.Conv2d(64, 128,3,padding=(1,1))        self.conv4 = torch.nn.Conv2d(128, 128, 3, padding=(1, 1))        self.pool2 = torch.nn.MaxPool2d((2, 2), padding=(1, 1))        self.conv5 = torch.nn.Conv2d(128, 256,3, padding=(1,1))        self.conv6 = torch.nn.Conv2d(256, 256,3, padding=(1, 1))        self.conv7 = torch.nn.Conv2d(256, 256,3, padding=(1, 1))        self.pool3 = torch.nn.MaxPool2d((2,2), padding=(1, 1))                       self.conv8 = torch.nn.Conv2d(256, 512,3, padding=(1,1))        self.conv9 = torch.nn.Conv2d(512, 512,3, padding=(1, 1))        self.conv10 = torch.nn.Conv2d(512, 512,3, padding=(1, 1))        self.pool4 = torch.nn.MaxPool2d((2,2),padding=(1, 1))                 self.conv11 = torch.nn.Conv2d(512, 512,3)        self.conv12 = torch.nn.Conv2d(512, 512,3, padding=(1, 1))        self.conv13 = torch.nn.Conv2d(512, 512,3, padding=(1, 1))        self.pool5 = torch.nn.MaxPool2d((2,2),padding=(1, 1))         self.fc1 = nn.Linear(512*7*7, 4096)        self.dropout1 = nn.Dropout(0.5)        self.fc2 = nn.Linear(4096, 4096)        self.dropout2 = nn.Dropout(0.5)        self.fc3 = nn.Linear(4096, 2)    def forward(self, x):        insize = x.size(0)        out = F.relu(self.conv1(x))        out = self.pool1(F.relu(self.conv2(out)))        out = F.relu(self.conv3(out))        out = self.pool2(F.relu(self.conv4(out)))        out = F.relu(self.conv5(out))        out = F.relu(self.conv6(out))        out = self.pool3(F.relu(self.conv7(out)))        out = F.relu(self.conv8(out))        out = F.relu(self.conv9(out))        out = self.pool4(F.relu(self.conv10(out)))        out = F.relu(self.conv11(out))        out = F.relu(self.conv12(out))        out = self.pool5(F.relu(self.conv13(out)))        out = out.view(insize, -1) #这里对于不同数据处理会有不一样,-1位于前面相当于间接将数据进行平铺-->1*n;        # -1位于后面则--->n*1        out = self.dropout1(self.act1(self.fc1(out)))        out = self.dropout2(self.act1(self.fc2(out)))        out = self.fc3(out)        out = F.log_softmax(out, dim=1)        return outdevice = torch.device('cuda:0' if torch.cuda.is_available() else 'CPU')vgg = vgg16()x = torch.rand(size=(4, 3, 224, 224)) #相当于4张224x224的图片,所以旋转out.view(insize, -1)"""x = torch.rand(size=(3, 224, 224)) out.view(224*224*3, -1)"""vgg(x)