A股上市公司传智教育(股票代码 003032)旗下技术交流社区北京昌平校区

 找回密码
 加入黑马

QQ登录

只需一步,快速开始

本帖最后由 梦缠绕的时候 于 2019-6-19 14:42 编辑

1.引文

深度学习的比赛中,图片分类是很常见的比赛,同时也是很难取得特别高名次的比赛,因为图片分类已经被大家研究的很透彻,一些开源的网络很容易取得高分。如果大家还掌握不了使用开源的网络进行训练,再慢慢去模型调优,很难取得较好的成绩。

我们在[PyTorch小试牛刀]实战六·准备自己的数据集用于训练讲解了如何制作自己的数据集用于训练,这个教程在此基础上,进行训练与应用。

2.数据介绍

数据下载地址:
https://download.csdn.net/download/xiaosongshine/11128410

这次的实战使用的数据是交通标志数据集,共有62类交通标志。其中训练集数据有4572张照片(每个类别大概七十个),测试数据集有2520张照片(每个类别大概40个)。数据包含两个子目录分别train与test:

为什么还需要测试数据集呢?这个测试数据集不会拿来训练,是用来进行模型的评估与调优。


train与test每个文件夹里又有62个子文件夹,每个类别在同一个文件夹内:

我从中打开一个文件间,把里面图片展示出来:


其中每张照片都类似下面的例子,100*100*3的大小。100是照片的照片的长和宽,3是什么呢?这其实是照片的色彩通道数目,RGB。彩色照片存储在计算机里就是以三维数组的形式。我们送入网络的也是这些数组。

3.网络构建

1.导入Python包,定义一些参数

1import torch as t
2import torchvision as tv
3import os
4import time
5import numpy as np
6from tqdm import tqdm
7
8
9class DefaultConfigs(object):
10
11    data_dir = "./traffic-sign/"
12    data_list = ["train","test"]
13
14    lr = 0.001
15    epochs = 10
16    num_classes = 62
17    image_size = 224
18    batch_size = 40
19    channels = 3
20    gpu = "0"
21    train_len = 4572
22    test_len = 2520
23    use_gpu = t.cuda.is_available()
24
25config = DefaultConfigs()


2.数据准备,采用PyTorch提供的读取方式

注意一点Train数据需要进行随机裁剪,Test数据不要进行裁剪了

1normalize = tv.transforms.Normalize(mean = [0.485, 0.456, 0.406],
2                                    std = [0.229, 0.224, 0.225]
3                                    )
4
5transform = {
6    config.data_list[0]:tv.transforms.Compose(
7        [tv.transforms.Resize([224,224]),tv.transforms.CenterCrop([224,224]),
8        tv.transforms.ToTensor(),normalize]#tv.transforms.Resize 用于重设图片大小
9    ) ,
10    config.data_list[1]:tv.transforms.Compose(
11        [tv.transforms.Resize([224,224]),tv.transforms.ToTensor(),normalize]
12    )
13}
14
15datasets = {
16    x:tv.datasets.ImageFolder(root = os.path.join(config.data_dir,x),transform=transform[x])
17    for x in config.data_list
18}
19
20dataloader = {
21    x:t.utils.data.DataLoader(dataset= datasets[x],
22        batch_size=config.batch_size,
23        shuffle=True
24    )
25    for x in config.data_list
26}


3.构建网络模型(使用resnet18进行迁移学习,训练参数为最后一个全连接层 t.nn.Linear(512,num_classes))

1def get_model(num_classes):
2
3    model = tv.models.resnet18(pretrained=True)
4    for parma in model.parameters():
5        parma.requires_grad = False
6    model.fc = t.nn.Sequential(
7        t.nn.Dropout(p=0.3),
8        t.nn.Linear(512,num_classes)
9    )
10    return(model)


如果电脑硬件支持,可以把下述代码屏蔽,则训练整个网络,最终准确率会上升,训练数据会变慢。

1for parma in model.parameters():
2    parma.requires_grad = False


模型输出

1ResNet(
2  (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
3  (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
4  (relu): ReLU(inplace)
5  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
6  (layer1): Sequential(
7    (0): BasicBlock(
8      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
9      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
10      (relu): ReLU(inplace)
11      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
12      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
13    )
14    (1): BasicBlock(
15      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
16      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
17      (relu): ReLU(inplace)
18      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
19      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
20    )
21  )
22  (layer2): Sequential(
23    (0): BasicBlock(
24      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
25      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
26      (relu): ReLU(inplace)
27      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
28      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
29      (downsample): Sequential(
30        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
31        (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
32      )
33    )
34    (1): BasicBlock(
35      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
36      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
37      (relu): ReLU(inplace)
38      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
39      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
40    )
41  )
42  (layer3): Sequential(
43    (0): BasicBlock(
44      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
45      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
46      (relu): ReLU(inplace)
47      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
48      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
49      (downsample): Sequential(
50        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
51        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
52      )
53    )
54    (1): BasicBlock(
55      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
56      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
57      (relu): ReLU(inplace)
58      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
59      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
60    )
61  )
62  (layer4): Sequential(
63    (0): BasicBlock(
64      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
65      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
66      (relu): ReLU(inplace)
67      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
68      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
69      (downsample): Sequential(
70        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
71        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
72      )
73    )
74    (1): BasicBlock(
75      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
76      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
77      (relu): ReLU(inplace)
78      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
79      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
80    )
81  )
82  (avgpool): AvgPool2d(kernel_size=7, stride=1, padding=0)
83  (fc): Sequential(
84    (0): Dropout(p=0.3)
85    (1): Linear(in_features=512, out_features=62, bias=True)
86  )
87)


4.训练模型(支持自动GPU加速)

1def train(epochs):
2
3    model = get_model(config.num_classes)
4    print(model)
5    loss_f = t.nn.CrossEntropyLoss()
6    if(config.use_gpu):
7        model = model.cuda()
8        loss_f = loss_f.cuda()
9
10    opt = t.optim.Adam(model.fc.parameters(),lr = config.lr)
11    time_start = time.time()
12
13    for epoch in range(epochs):
14        train_loss = []
15        train_acc = []
16        test_loss = []
17        test_acc = []
18        model.train(True)
19        print("Epoch {}/{}".format(epoch+1,epochs))
20        for batch, datas in tqdm(enumerate(iter(dataloader["train"]))):
21            x,y = datas
22            if (config.use_gpu):
23                x,y = x.cuda(),y.cuda()
24            y_ = model(x)
25            #print(x.shape,y.shape,y_.shape)
26            _, pre_y_ = t.max(y_,1)
27            pre_y = y
28            #print(y_.shape)
29            loss = loss_f(y_,pre_y)
30            #print(y_.shape)
31            acc = t.sum(pre_y_ == pre_y)
32
33            loss.backward()
34            opt.step()
35            opt.zero_grad()
36            if(config.use_gpu):
37                loss = loss.cpu()
38                acc = acc.cpu()
39            train_loss.append(loss.data)
40            train_acc.append(acc)
41            #if((batch+1)%5 ==0):
42        time_end = time.time()
43        print("Batch {}, Train loss:{:.4f}, Train acc:{:.4f}, Time: {}"\
44            .format(batch+1,np.mean(train_loss)/config.batch_size,np.mean(train_acc)/config.batch_size,(time_end-time_start)))
45        time_start = time.time()
46
47        model.train(False)
48        for batch, datas in tqdm(enumerate(iter(dataloader["test"]))):
49            x,y = datas
50            if (config.use_gpu):
51                x,y = x.cuda(),y.cuda()
52            y_ = model(x)
53            #print(x.shape,y.shape,y_.shape)
54            _, pre_y_ = t.max(y_,1)
55            pre_y = y
56            #print(y_.shape)
57            loss = loss_f(y_,pre_y)
58            acc = t.sum(pre_y_ == pre_y)
59
60            if(config.use_gpu):
61                loss = loss.cpu()
62                acc = acc.cpu()
63
64            test_loss.append(loss.data)
65            test_acc.append(acc)
66        print("Batch {}, Test loss:{:.4f}, Test acc:{:.4f}".format(batch+1,np.mean(test_loss)/config.batch_size,np.mean(test_acc)/config.batch_size))
67
68        t.save(model,str(epoch+1)+"ttmodel.pkl")
69
70
71
72if __name__ == "__main__":
73    train(config.epochs)


训练结果如下:

1Epoch 1/10
2115it [00:48,  2.63it/s]
3Batch 115, Train loss:0.0590, Train acc:0.4635, Time: 48.985504150390625
463it [00:24,  2.62it/s]
5Batch 63, Test loss:0.0374, Test acc:0.6790, Time :24.648272275924683
6Epoch 2/10
7115it [00:45,  3.22it/s]
8Batch 115, Train loss:0.0271, Train acc:0.7576, Time: 45.68823838233948
963it [00:23,  2.62it/s]
10Batch 63, Test loss:0.0255, Test acc:0.7524, Time :23.271782875061035
11Epoch 3/10
12115it [00:45,  3.19it/s]
13Batch 115, Train loss:0.0181, Train acc:0.8300, Time: 45.92648506164551
1463it [00:23,  2.60it/s]
15Batch 63, Test loss:0.0212, Test acc:0.7861, Time :23.80789279937744
16Epoch 4/10
17115it [00:45,  3.28it/s]
18Batch 115, Train loss:0.0138, Train acc:0.8767, Time: 45.27525019645691
1963it [00:23,  2.57it/s]
20Batch 63, Test loss:0.0173, Test acc:0.8385, Time :23.736321449279785
21Epoch 5/10
22115it [00:44,  3.22it/s]
23Batch 115, Train loss:0.0112, Train acc:0.8950, Time: 44.983638286590576
2463it [00:22,  2.69it/s]
25Batch 63, Test loss:0.0156, Test acc:0.8520, Time :22.790074348449707
26Epoch 6/10
27115it [00:44,  3.19it/s]
28Batch 115, Train loss:0.0095, Train acc:0.9159, Time: 45.10426950454712
2963it [00:22,  2.77it/s]
30Batch 63, Test loss:0.0158, Test acc:0.8214, Time :22.80412459373474
31Epoch 7/10
32115it [00:45,  2.95it/s]
33Batch 115, Train loss:0.0081, Train acc:0.9280, Time: 45.30439043045044
3463it [00:23,  2.66it/s]
35Batch 63, Test loss:0.0139, Test acc:0.8528, Time :23.122379541397095
36Epoch 8/10
37115it [00:44,  3.23it/s]
38Batch 115, Train loss:0.0073, Train acc:0.9300, Time: 44.304762840270996
3963it [00:22,  2.74it/s]
40Batch 63, Test loss:0.0142, Test acc:0.8496, Time :22.801835536956787
41Epoch 9/10
42115it [00:43,  3.19it/s]
43Batch 115, Train loss:0.0068, Train acc:0.9361, Time: 44.08414030075073
4463it [00:23,  2.44it/s]
45Batch 63, Test loss:0.0142, Test acc:0.8437, Time :23.604419231414795
46Epoch 10/10
47115it [00:46,  3.12it/s]
48Batch 115, Train loss:0.0063, Train acc:0.9337, Time: 46.76597046852112
4963it [00:24,  2.65it/s]
50Batch 63, Test loss:0.0130, Test acc:0.8591, Time :24.64351773262024


训练10个Epoch,测试集准确率可以到达0.86,已经达到不错效果。通过修改参数,增加训练,可以达到更高的准确率。

3 个回复

倒序浏览
有任何问题欢迎联系学姐或者在下方留言哦

回复 使用道具 举报
感谢分享~~
回复 使用道具 举报
梦缠绕的时候 发表于 2019-6-19 14:43
有任何问题欢迎联系学姐或者在下方留言哦

学姐,怎么把分错的照片找出来你知道吗?
回复 使用道具 举报
您需要登录后才可以回帖 登录 | 加入黑马