Going Deeper with Convolutions (GoogLeNet)

目录

• 代码

Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]. computer vision and pattern recognition, 2015: 1-9.

@article{szegedy2015going,

title={Going deeper with convolutions},

author={Szegedy, Christian and Liu, Wei and Jia, Yangqing and Sermanet, Pierre and Reed, Scott and Anguelov, Dragomir and Erhan, Dumitru and Vanhoucke, Vincent and Rabinovich, Andrew},

pages={1--9},

year={2015}}

这里讲的很细, 不多赘诉了.

代码

"""
代码虽然是"copy"源代码, 但是收获不少.
虽然参数少, 但是训练得很慢, 是因为要传三次梯度?
测试集上正确率维0.8682
"""

import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
import numpy as np
import os

class BasicConv2d(nn.Module):

    def __init__(self, in_channels, out_channels, **kwargs):
        super(BasicConv2d, self).__init__()
        self.conv = nn.Conv2d(in_channels, out_channels,
                              bias=False, **kwargs) #不要偏置
        #eps 为了数值稳定 默认是1e-5
        self.bn = nn.BatchNorm2d(out_channels, eps=0.001)
        self.relu = nn.ReLU(inplace=True)

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

class Inception(nn.Module):

    def __init__(self, in_channels, ch1x1, ch3x3red, ch3x3,
                 ch5x5red, ch5x5, pool_proj):
        """
        :param in_channels: 输入的通道数
        :param ch1x1:   1x1卷积核的输出通道数
        :param ch3x3red: 3x3一开始的1x1部分的通道数
        :param ch3x3: 3x3后半的3x3部分的通道数
        :param ch5x5: ...
        :param ch5x5red:  ...
        :param pool_proj:  池化层的通道数
        """
        super(Inception, self).__init__()

        self.branch1 = BasicConv2d(in_channels, ch1x1, kernel_size=1)

        self.branch2 = nn.Sequential(
            BasicConv2d(in_channels, ch3x3red, kernel_size=1),
            BasicConv2d(ch3x3red, ch3x3, kernel_size=3, padding=1)
        )

        #pytorch 这里用的3x3卷积核?
        self.branch3 = nn.Sequential(
            BasicConv2d(in_channels, ch5x5red, kernel_size=1),
            BasicConv2d(ch5x5red, ch5x5, kernel_size=5, padding=2)
        )

        self.branch4 = nn.Sequential(
            nn.MaxPool2d(kernel_size=3, stride=1, padding=1, ceil_mode=True),
            BasicConv2d(in_channels, pool_proj, kernel_size=1)
        )

    def forward(self, x):
        x1 = self.branch1(x)
        x2 = self.branch2(x)
        x3 = self.branch3(x)
        x4 = self.branch4(x)
        out = (x1, x2, x3, x4)
        return torch.cat(out, 1)

class InceptionAux(nn.Module):

    def __init__(self, in_channels, num_classes):
        super(InceptionAux, self).__init__()
        self.avgpool = nn.AdaptiveAvgPool2d((4, 4))
        self.conv = BasicConv2d(in_channels, 128, kernel_size=1)
        #N x 128 x 4 x 4
        self.dense = nn.Sequential(
            nn.Linear(2048, 1024),
            nn.ReLU(inplace=True),
            nn.Dropout(0.7),
            nn.Linear(1024, num_classes)
        )

    def forward(self, x):
        x = self.avgpool(x)
        x = self.conv(x)
        x = torch.flatten(x, 1)
        out = self.dense(x)
        return out

class GoogLeNet(nn.Module):

    def __init__(self, num_classes=10, aux_logits=True):
        """
        :param num_classes: 类别个数
        :param aux_logits: 是否需要添加辅助训练器
        """
        super(GoogLeNet, self).__init__()
        self.aux_logits =aux_logits

        # N x 3 x 224 x 224
        self.conv1 = BasicConv2d(3, 64, kernel_size=7, stride=2, padding=3)
        # N x 64 x 112 x 112
        self.maxpool1 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
        # N x 64 x 56 x 56
        self.conv2 = BasicConv2d(64, 64, kernel_size=1)
        self.conv3 = BasicConv2d(64, 192, kernel_size=3, padding=1)
        # N x 192 x 56 x 56
        self.maxpool2 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
        # N x 192 x 28 x 28

        self.inception3a = Inception(192, 64, 96, 128, 16, 32, 32)
        #N x 256 x 28 x 28
        self.inception3b = Inception(256, 128, 128, 192, 32, 96, 64)
        #N x 480 x 28 x 28
        self.maxpool3 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
        #N x 480 x 14 x 14

        self.inception4a = Inception(480, 192, 96, 208, 16, 48, 64)
        #N x 512 x 14 x 14
        self.inception4b = Inception(512, 160, 112, 224, 24, 64, 64)
        #N x 512 x 14 x 14
        self.inception4c = Inception(512, 128, 128, 256, 24, 64, 64)
        #N x 512 x 14 x 14
        self.inception4d = Inception(512, 112, 144, 288, 32, 64, 64)
        #N x 528 x 14 x 14
        self.inception4e = Inception(528, 256, 160, 320, 32, 128, 128)
        #N x 832 x 14 x 14
        self.maxpool4 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
        #N x 832 x 7 x 7

        self.inception5a = Inception(832, 256, 160, 320, 32, 128, 128)
        #N x 832 x 7 x 7
        self.inception5b = Inception(832, 384, 192, 384, 48, 128, 128)
        #N x 1024 x 7 x 7
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        #N x 1024 x 1 x 1
        self.drop = nn.Dropout(0.4)
        self.fc = nn.Linear(1024, num_classes)

        if self.aux_logits:
            self.aux1 = InceptionAux(512, num_classes)
            self.aux2 = InceptionAux(528, num_classes)

    def forward(self, x):
        x = self.conv1(x)
        x = self.maxpool1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.maxpool2(x)

        x = self.inception3a(x)
        x = self.inception3b(x)
        x = self.maxpool3(x)

        x = self.inception4a(x)
        if self.aux_logits and self.training:
            aux1 = self.aux1(x)
        x = self.inception4b(x)
        x = self.inception4c(x)
        x = self.inception4d(x)
        if self.aux_logits and self.training:
            aux2 = self.aux2(x)
        x = self.inception4e(x)
        x = self.maxpool4(x)

        x = self.inception5a(x)
        x = self.inception5b(x)

        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.drop(x)
        out = self.fc(x)

        if self.aux_logits and self.training:
            return (out, aux1, aux2)
        return out

class Train:

    def __init__(self, lr=0.01, momentum=0.9, weight_decay=0.0001):
        self.net = GoogLeNet()
        self.criterion = nn.CrossEntropyLoss()
        self.opti = torch.optim.SGD(self.net.parameters(),
                                    lr=lr, momentum=momentum,
                                    weight_decay=weight_decay)
        self.gpu()
        self.generate_path()
        self.acc_rates = []
        self.errors = []

    def gpu(self):
        self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
        if torch.cuda.device_count() > 1:
            print("Let'us use %d GPUs" % torch.cuda.device_count())
            self.net = nn.DataParallel(self.net)
        self.net = self.net.to(self.device)

    def generate_path(self):
        """
        生成保存数据的路径
        :return:
        """
        try:
            os.makedirs('./paras')
            os.makedirs('./logs')
            os.makedirs('./infos')
        except FileExistsError as e:
            pass
        name = self.net.__class__.__name__
        paras = os.listdir('./paras')
        logs = os.listdir('./logs')
        infos = os.listdir('./infos')
        number = max((len(paras), len(logs), len(infos)))
        self.para_path = "./paras/{0}{1}.pt".format(
            name,
            number
        )

        self.log_path = "./logs/{0}{1}.txt".format(
            name,
            number
        )
        self.info_path = "./infos/{0}{1}.npy".format(
            name,
            number
        )

    def log(self, strings):
        """
        运行日志
        :param strings:
        :return:
        """
        # a 往后添加内容
        with open(self.log_path, 'a', encoding='utf8') as f:
            f.write(strings)

    def save(self):
        """
        保存网络参数
        :return:
        """
        torch.save(self.net.state_dict(), self.para_path)

    def derease_lr(self, multi=0.96):
        """
        降低学习率
        :param multi:
        :return:
        """
        self.opti.param_groups[0]['lr'] *= multi

    def train(self, trainloder, epochs=50):
        data_size = len(trainloder) * trainloder.batch_size
        part = int(trainloder.batch_size / 2)
        for epoch in range(epochs):
            running_loss = 0.
            total_loss = 0.
            acc_count = 0.
            if (epoch + 1) % 8 is 0:
                self.derease_lr()
                self.log(#日志记录
                    "learning rate change!!!\n"
                )
            for i, data in enumerate(trainloder):
                imgs, labels = data
                imgs = imgs.to(self.device)
                labels = labels.to(self.device)
                (out, aux1, aux2) = self.net(imgs)
                loss1 = self.criterion(out, labels)
                loss2 = self.criterion(aux1, labels)
                loss3 = self.criterion(aux2, labels)
                loss = 0.4 * loss1 + 0.3 * loss2 + 0.3 * loss3
                _, pre = torch.max(out, 1)  #判断是否判断正确
                acc_count += (pre == labels).sum().item() #加总对的个数

                self.opti.zero_grad()
                loss.backward()
                self.opti.step()

                running_loss += loss.item()

                if (i+1) % part is 0:
                    strings = "epoch {0:<3} part {1:<5} loss: {2:<.7f}\n".format(
                        epoch, i, running_loss / part
                    )
                    self.log(strings)#日志记录
                    total_loss += running_loss
                    running_loss = 0.
            self.acc_rates.append(acc_count / data_size)
            self.errors.append(total_loss / data_size)
            self.log( #日志记录
                "Accuracy of the network on %d train images: %d %%\n" %(
                    data_size, acc_count / data_size * 100
                )
            )
            self.save() #保存网络参数
        #保存一些信息画图用
        np.save(self.info_path, {
            'acc_rates': np.array(self.acc_rates),
            'errors': np.array(self.errors)
        })

if __name__ == "__main__":

    root = "../../data"

    trainset = torchvision.datasets.CIFAR10(root=root, train=True,
                                          download=False,
                                          transform=transforms.Compose(
                                              [transforms.Resize(224),
                                               transforms.ToTensor(),
                                               transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
                                          ))

    train_loader = torch.utils.data.DataLoader(trainset, batch_size=128,
                                              shuffle=True, num_workers=8,
                                               pin_memory=True)

    dog = Train()
    dog.train(train_loader, epochs=1000)