pytorch识别CIFAR10：训练ResNet-34（准确率80%）

版权声明：本文为博主原创文章，欢迎转载，并请注明出处。联系方式：460356155@qq.com

CNN的层数越多，能够提取到的特征越丰富，但是简单地增加卷积层数，训练时会导致梯度弥散或梯度爆炸。

何凯明2015年提出了残差神经网络，即Reset，并在ILSVRC-2015的分类比赛中获得冠军。

ResNet可以有效的消除卷积层数增加带来的梯度弥散或梯度爆炸问题。

ResNet的核心思想是网络输出分为2部分恒等映射（identity mapping）、残差映射（residual mapping），即y = x + F（x），图示如下：

ResNet通过改变学习目标，即由学习完整的输出变为学习残差，解决了传统卷积在信息传递时存在的信息丢失核损耗问题，通过将输入直接绕道传递到输出，保护了信息的完整性。此外学习目标的简化也降低了学习难度。

常见的ResNet结构有：

34层的ResNet图示如下：

pytorch实现核训练ResNet-34的代码如下：

# -*- coding:utf-8 -*-

u"""ResNet训练学习CIFAR10"""

__author__ = 'zhengbiqing 460356155@qq.com'


import torch as t
import torchvision as tv
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
from torchvision.transforms import ToPILImage
import torch.backends.cudnn as cudnn

import matplotlib.pyplot as plt

import datetime
import argparse


# 样本读取线程数
WORKERS = 4

# 网络参赛保存文件名
PARAS_FN = 'cifar_resnet_params.pkl'

# minist数据存放位置
ROOT = '/home/zbq/PycharmProjects/cifar'

# 目标函数
loss_func = nn.CrossEntropyLoss()

# 最优结果
best_acc = 0

# 记录准确率，显示曲线
global_train_acc = []
global_test_acc = []


'''
残差块
in_channels, out_channels：残差块的输入、输出通道数
对第一层，in out channel都是64，其他层则不同
对每一层，如果in out channel不同， stride是1，其他层则为2
'''
class ResBlock(nn.Module):
    def __init__(self, in_channels, out_channels, stride=1):
        super(ResBlock, self).__init__()

        # 残差块的第一个卷积
        # 通道数变换in->out，每一层（除第一层外）的第一个block
        # 图片尺寸变换：stride=2时，w-3+2 / 2 + 1 = w/2，w/2 * w/2
        # stride=1时尺寸不变，w-3+2 / 1 + 1 = w
        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1)
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.relu = nn.ReLU(inplace=True)

        # 残差块的第二个卷积
        # 通道数、图片尺寸均不变
        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
        self.bn2 = nn.BatchNorm2d(out_channels)

        # 残差块的shortcut
        # 如果残差块的输入输出通道数不同，则需要变换通道数及图片尺寸，以和residual部分相加
        # 输出：通道数*2 图片尺寸/2
        if in_channels != out_channels:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=2),
                nn.BatchNorm2d(out_channels)
            )
        else:
            # 通道数相同，无需做变换，在forward中identity = x
            self.downsample = None

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out


'''
定义网络结构
'''
class ResNet34(nn.Module):
    def __init__(self, block):
        super(ResNet34, self).__init__()

        # 初始卷积层核池化层
        self.first = nn.Sequential(
            # 卷基层1：7*7kernel，2stride，3padding，outmap：32-7+2*3 / 2 + 1，16*16
            nn.Conv2d(3, 64, 7, 2, 3),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),

            # 最大池化，3*3kernel，1stride（32的原始输入图片较小，不再缩小尺寸），1padding，
            # outmap：16-3+2*1 / 1 + 1，16*16
            nn.MaxPool2d(3, 1, 1)
        )

        # 第一层，通道数不变
        self.layer1 = self.make_layer(block, 64, 64, 3, 1)

        # 第2、3、4层，通道数*2，图片尺寸/2
        self.layer2 = self.make_layer(block, 64, 128, 4, 2)  # 输出8*8
        self.layer3 = self.make_layer(block, 128, 256, 6, 2)  # 输出4*4
        self.layer4 = self.make_layer(block, 256, 512, 3, 2)  # 输出2*2

        self.avg_pool = nn.AvgPool2d(2)  # 输出512*1
        self.fc = nn.Linear(512, 10)

    def make_layer(self, block, in_channels, out_channels, block_num, stride):
        layers = []

        # 每一层的第一个block，通道数可能不同
        layers.append(block(in_channels, out_channels, stride))

        # 每一层的其他block，通道数不变，图片尺寸不变
        for i in range(block_num - 1):
            layers.append(block(out_channels, out_channels, 1))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.first(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        x = self.avg_pool(x)

        # x.size()[0]: batch size
        x = x.view(x.size()[0], -1)
        x = self.fc(x)

        return x


'''
训练并测试网络
net：网络模型
train_data_load：训练数据集
optimizer：优化器
epoch：第几次训练迭代
log_interval：训练过程中损失函数值和准确率的打印频率
'''
def net_train(net, train_data_load, optimizer, epoch, log_interval):
    net.train()

    begin = datetime.datetime.now()

    # 样本总数
    total = len(train_data_load.dataset)

    # 样本批次训练的损失函数值的和
    train_loss = 0

    # 识别正确的样本数
    ok = 0

    for i, data in enumerate(train_data_load, 0):
        img, label = data
        img, label = img.cuda(), label.cuda()

        optimizer.zero_grad()

        outs = net(img)
        loss = loss_func(outs, label)
        loss.backward()
        optimizer.step()

        # 累加损失值和训练样本数
        train_loss += loss.item()

        _, predicted = t.max(outs.data, 1)
        # 累加识别正确的样本数
        ok += (predicted == label).sum()

        if (i + 1) % log_interval == 0:
            # 训练结果输出

            # 已训练的样本数
            traind_total = (i + 1) * len(label)

            # 准确度
            acc = 100. * ok / traind_total

            # 记录训练准确率以输出变化曲线
            global_train_acc.append(acc)

    end = datetime.datetime.now()
    print('one epoch spend: ', end - begin)


'''
用测试集检查准确率
'''
def net_test(net, test_data_load, epoch):
    net.eval()

    ok = 0

    for i, data in enumerate(test_data_load):
        img, label = data
        img, label = img.cuda(), label.cuda()

        outs = net(img)
        _, pre = t.max(outs.data, 1)
        ok += (pre == label).sum()

    acc = ok.item() * 100. / (len(test_data_load.dataset))
    print('EPOCH:{}, ACC:{}
'.format(epoch, acc))

    # 记录测试准确率以输出变化曲线
    global_test_acc.append(acc)

    # 最好准确度记录
    global best_acc
    if acc > best_acc:
        best_acc = acc


'''
显示数据集中一个图片
'''
def img_show(dataset, index):
    classes = ('plane', 'car', 'bird', 'cat',
               'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

    show = ToPILImage()

    data, label = dataset[index]
    print('img is a ', classes[label])
    show((data + 1) / 2).resize((100, 100)).show()


'''
显示训练准确率、测试准确率变化曲线
'''
def show_acc_curv(ratio):
    # 训练准确率曲线的x、y
    train_x = list(range(len(global_train_acc)))
    train_y = global_train_acc

    # 测试准确率曲线的x、y
    # 每ratio个训练准确率对应一个测试准确率
    test_x = train_x[ratio-1::ratio]
    test_y = global_test_acc

    plt.title('CIFAR10 RESNET34 ACC')

    plt.plot(train_x, train_y, color='green', label='training accuracy')
    plt.plot(test_x, test_y, color='red', label='testing accuracy')

    # 显示图例
    plt.legend()
    plt.xlabel('iterations')
    plt.ylabel('accs')

    plt.show()


def main():
    # 训练超参数设置，可通过命令行设置
    parser = argparse.ArgumentParser(description='PyTorch CIFA10 ResNet34 Example')
    parser.add_argument('--batch-size', type=int, default=128, metavar='N',
                        help='input batch size for training (default: 128)')
    parser.add_argument('--test-batch-size', type=int, default=100, metavar='N',
                        help='input batch size for testing (default: 100)')
    parser.add_argument('--epochs', type=int, default=200, metavar='N',
                        help='number of epochs to train (default: 200)')
    parser.add_argument('--lr', type=float, default=0.1, metavar='LR',
                        help='learning rate (default: 0.1)')
    parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                        help='SGD momentum (default: 0.9)')
    parser.add_argument('--log-interval', type=int, default=10, metavar='N',
                        help='how many batches to wait before logging training status (default: 10)')
    parser.add_argument('--no-train', action='store_true', default=False,
                        help='If train the Model')
    parser.add_argument('--save-model', action='store_true', default=False,
                        help='For Saving the current Model')
    args = parser.parse_args()

    # 图像数值转换，ToTensor源码注释
    """Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor.
        Converts a PIL Image or numpy.ndarray (H x W x C) in the range
        [0, 255] to a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0].
        """
    # 归一化把[0.0, 1.0]变换为[-1,1], ([0, 1] - 0.5) / 0.5 = [-1, 1]
    transform = tv.transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])

    # 定义数据集
    train_data = tv.datasets.CIFAR10(root=ROOT, train=True, download=True, transform=transform)
    test_data = tv.datasets.CIFAR10(root=ROOT, train=False, download=False, transform=transform)

    train_load = t.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=WORKERS)
    test_load = t.utils.data.DataLoader(test_data, batch_size=args.test_batch_size, shuffle=False, num_workers=WORKERS)

    net = ResNet34(ResBlock).cuda()
    print(net)

    # 并行计算提高运行速度
    net = nn.DataParallel(net)
    cudnn.benchmark = True

    # 如果不训练，直接加载保存的网络参数进行测试集验证
    if args.no_train:
        net.load_state_dict(t.load(PARAS_FN))
        net_test(net, test_load, 0)
        return

    optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum)

    start_time = datetime.datetime.now()

    for epoch in range(1, args.epochs + 1):
        net_train(net, train_load, optimizer, epoch, args.log_interval)

        # 每个epoch结束后用测试集检查识别准确度
        net_test(net, test_load, epoch)

    end_time = datetime.datetime.now()

    global best_acc
    print('CIFAR10 pytorch ResNet34 Train: EPOCH:{}, BATCH_SZ:{}, LR:{}, ACC:{}'.format(args.epochs, args.batch_size, args.lr, best_acc))
    print('train spend time: ', end_time - start_time)

    # 每训练一个迭代记录的训练准确率个数
    ratio = len(train_data) / args.batch_size / args.log_interval
    ratio = int(ratio)

    # 显示曲线
    show_acc_curv(ratio)

    if args.save_model:
        t.save(net.state_dict(), PARAS_FN)


if __name__ == '__main__':
    main()

运行结果：

Files already downloaded and verified
ResNet34(
  (first): Sequential(
    (0): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3))
    (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (2): ReLU(inplace)
    (3): MaxPool2d(kernel_size=3, stride=1, padding=1, dilation=1, ceil_mode=False)
  )
  (layer1): Sequential(
    (0): ResBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (1): ResBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (2): ResBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (layer2): Sequential(
    (0): ResBlock(
      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2))
        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResBlock(
      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (2): ResBlock(
      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (3): ResBlock(
      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (layer3): Sequential(
    (0): ResBlock(
      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2))
        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (2): ResBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (3): ResBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (4): ResBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (5): ResBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (layer4): Sequential(
    (0): ResBlock(
      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2))
        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): ResBlock(
      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (2): ResBlock(
      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (avg_pool): AvgPool2d(kernel_size=2, stride=2, padding=0)
  (fc): Linear(in_features=512, out_features=10, bias=True)
)
one epoch spend:  0:00:23.971338
EPOCH:1, ACC:48.54

one epoch spend:  0:00:22.339190
EPOCH:2, ACC:61.06

......

one epoch spend:  0:00:22.023034
EPOCH:199, ACC:79.84

one epoch spend:  0:00:22.057692
EPOCH:200, ACC:79.6

CIFAR10 pytorch ResNet34 Train: EPOCH:200, BATCH_SZ:128, LR:0.1, ACC:80.19
train spend time:  1:18:40.948080

运行200个迭代，每个迭代耗时22秒，准确率不高，只有80%。准确率变化曲线如下：