Pytorch实战:8层神经网络实现Cifar-10图像分类验证集准确率94.71%

实验环境:

  1. Pytorch 1.7.0
  2. torchvision 0.8.2
  3. Python 3.8
  4. CUDA10.2 + cuDNN v7.6.5
  5. Win10 + Pycharm
  6. GTX1660, 6G

网络结构采用最简洁的类VGG结构,即全部由3*3卷积和最大池化组成,后面接一个全连接层用于分类,网络大小仅18M左右。

神经网络结构图:

 

 

Pytorch上搭建网络:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
class Block(nn.Module):
    def __init__(self, inchannel, outchannel, res=True):
        super(Block, self).__init__()
        self.res = res     # 是否带残差连接
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, outchannel, kernel_size=3, padding=1, bias=False),
            nn.BatchNorm2d(outchannel),
            nn.ReLU(inplace=True),
            nn.Conv2d(outchannel, outchannel, kernel_size=3, padding=1, bias=False),
            nn.BatchNorm2d(outchannel),
        )
        if stride != 1 or inchannel != outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, outchannel, kernel_size=1, bias=False),
                nn.BatchNorm2d(outchannel),
            )
        else:
            self.shortcut = nn.Sequential()
 
        self.relu = nn.Sequential(
            nn.ReLU(inplace=True),
        )
 
    def forward(self, x):
        out = self.left(x)
        if self.res:
            out += self.shortcut(x)
        out = self.relu(out)
        return out
 
 
class myModel(nn.Module):
    def __init__(self, cfg=[64, 'M', 128'M', 256, 'M', 512, 'M'], res=True):
        super(myModel, self).__init__()
        self.res = res       # 是否带残差连接
        self.cfg = cfg       # 配置列表
        self.inchannel = 3   # 初始输入通道数
        self.futures = self.make_layer()
        # 构建卷积层之后的全连接层以及分类器:
        self.classifier = nn.Sequential(nn.Dropout(0.4),            # 两层fc效果还差一些
                                        nn.Linear(4 * 512, 10), )   # fc,最终Cifar10输出是10类
 
    def make_layer(self):
        layers = []
        for v in self.cfg:
            if v == 'M':
                layers.append(nn.MaxPool2d(kernel_size=2, stride=2))
            else:
                layers.append(Block(self.inchannel, v, self.res))
                self.inchannel = v    # 输入通道数改为上一层的输出通道数
        return nn.Sequential(*layers)
 
    def forward(self, x):
        out = self.futures(x)
        # view(out.size(0), -1): change tensor size from (N ,H , W) to (N, H*W)
        out = out.view(out.size(0), -1)
        out = self.classifier(out)
        return out

 

该网络可以很方便的改造成带残差的,只要在初始化网络时,将参数res设为True即可,并可改变cfg配置列表来方便的修改网络层数。

Pytorch上训练:
所选数据集为Cifar-10,该数据集共有60000张带标签的彩色图像,这些图像尺寸32*32,分为10个类,每类6000张图。这里面有50000张用于训练,每个类5000张,另外10000用于测试,每个类1000张。
训练策略如下:

1.优化器:momentum=0.9 的 optim.SGD,adam在很多情况下能加速收敛,但因为是自适应学习率,在训练后期存在不能收敛到全局极值点的问题,所以采用能手动调节学习率的SGD,现在很多比赛和论文中也是采用该策略。设置weight_decay=5e-3,即设置较大的L2正则来降低过拟合。

1
2
3
# 定义损失函数和优化器
loss_func = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9, weight_decay=5e-3)

  

2.学习率:optim.lr_scheduler.MultiStepLR,参数设为:milestones=[int(num_epochs * 0.56), int(num_epochs * 0.78)], gamma=0.1,即在0.56倍epochs和0.78时分别下降为前一阶段学习率的0.1倍。

1
2
3
4
# 学习率调整策略 MultiStep:
scheduler = optim.lr_scheduler.MultiStepLR(optimizer=optimizer,
                   milestones=[int(num_epochs * 0.56), int(num_epochs * 0.78)],
                   gamma=0.1, last_epoch=-1)

在每个epoch训练完的时候一定要记得step一下,不然不会更新学习率,可以通过get_last_lr()来查看最新的学习率

1
2
3
# 更新学习率并查看当前学习率
scheduler.step()
print('\t last_lr:', scheduler.get_last_lr())

3.数据策略:
实验表明,针对cifar10数据集,随机水平翻转、随机遮挡、随机中心裁剪能有效提高验证集准确率,而旋转、颜色抖动等则无效。

1
2
3
4
5
6
norm_mean = [0.485, 0.456, 0.406]      # 均值
norm_std = [0.229, 0.224, 0.225]       # 方差     
transforms.Normalize(norm_mean, norm_std),                    #将[0,1]归一化到[-1,1]
transforms.RandomHorizontalFlip(),                            # 随机水平镜像
transforms.RandomErasing(scale=(0.04, 0.2), ratio=(0.5, 2)),  # 随机遮挡
transforms.RandomCrop(32, padding=4)                          # 随机中心裁剪

4.超参数:

1
2
3
batch_size = 512     # 约占用显存4G
num_epochs = 200     # 训练轮数
LR = 0.01            # 初始学习率   

实验结果:best_acc= 94.71%

 

 

 

  

另外,将网络改成14层的带残差结构后,准确率上升到了95.56%,但是网络大小也从18M到了43M。以下是14层残差网络的全部代码,8层的只需修改cfg和初始化时的res参数:
cfg=[64, ‘M’, 128, 128, ‘M’, 256, 256, ‘M’, 512, 512,‘M’] 修改为 [64, ‘M’, 128, ‘M’, 256, ‘M’, 512, ‘M’]

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
# *_* coding : UTF-8 *_*
# 开发人员: Capsule
# 开发时间: 2022/9/28 15:17
# 文件名称: battey_class.py
# 开发工具: PyCharm
# 功能描述: 自建CNN对cifar10进行分类
 
import torch
from torchvision import datasets, transforms
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import onnx
import time
import numpy as np
import matplotlib.pyplot as plt
 
 
class Block(nn.Module):
    def __init__(self, inchannel, outchannel, res=True, stride=1):
        super(Block, self).__init__()
        self.res = res     # 是否带残差连接
        self.left = nn.Sequential(
            nn.Conv2d(inchannel, outchannel, kernel_size=3, padding=1, stride=stride, bias=False),
            nn.BatchNorm2d(outchannel),
            nn.ReLU(inplace=True),
            nn.Conv2d(outchannel, outchannel, kernel_size=3, padding=1, stride=1, bias=False),
            nn.BatchNorm2d(outchannel),
        )
        if stride != 1 or inchannel != outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, outchannel, kernel_size=1, bias=False),
                nn.BatchNorm2d(outchannel),
            )
        else:
            self.shortcut = nn.Sequential()
 
        self.relu = nn.Sequential(
            nn.ReLU(inplace=True),
        )
 
    def forward(self, x):
        out = self.left(x)
        if self.res:
            out += self.shortcut(x)
        out = self.relu(out)
        return out
 
 
class myModel(nn.Module):
    def __init__(self, cfg=[64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512,'M'], res=True):
        super(myModel, self).__init__()
        self.res = res       # 是否带残差连接
        self.cfg = cfg       # 配置列表
        self.inchannel = 3   # 初始输入通道数
        self.futures = self.make_layer()
        # 构建卷积层之后的全连接层以及分类器:
        self.classifier = nn.Sequential(nn.Dropout(0.4),           # 两层fc效果还差一些
                                        nn.Linear(4 * 512, 10), )   # fc,最终Cifar10输出是10类
 
    def make_layer(self):
        layers = []
        for v in self.cfg:
            if v == 'M':
                layers.append(nn.MaxPool2d(kernel_size=2, stride=2))
            else:
                layers.append(Block(self.inchannel, v, self.res))
                self.inchannel = v    # 输入通道数改为上一层的输出通道数
        return nn.Sequential(*layers)
 
    def forward(self, x):
        out = self.futures(x)
        # view(out.size(0), -1): change tensor size from (N ,H , W) to (N, H*W)
        out = out.view(out.size(0), -1)
        out = self.classifier(out)
        return out
 
all_start = time.time()
# 使用torchvision可以很方便地下载Cifar10数据集,而torchvision下载的数据集为[0,1]的PILImage格式
# 我们需要将张量Tensor归一化到[-1,1]
norm_mean = [0.485, 0.456, 0.406# 均值
norm_std = [0.229, 0.224, 0.225# 方差
transform_train = transforms.Compose([transforms.ToTensor(),  # 将PILImage转换为张量
                                      # 将[0,1]归一化到[-1,1]
                                      transforms.Normalize(norm_mean, norm_std),
                                      transforms.RandomHorizontalFlip(),  # 随机水平镜像
                                      transforms.RandomErasing(scale=(0.04, 0.2), ratio=(0.5, 2)),  # 随机遮挡
                                      transforms.RandomCrop(32, padding=4# 随机中心裁剪
                                      ])
 
transform_test = transforms.Compose([transforms.ToTensor(),
                                     transforms.Normalize(norm_mean, norm_std)])
 
# 超参数:
batch_size = 256
num_epochs = 200   # 训练轮数
LR = 0.01          # 初始学习率
 
# 选择数据集:
trainset = datasets.CIFAR10(root='Datasets', train=True, download=True, transform=transform_train)
testset = datasets.CIFAR10(root='Datasets', train=False, download=True, transform=transform_test)
# 加载数据:
train_data = DataLoader(dataset=trainset, batch_size=batch_size, shuffle=True)
valid_data = DataLoader(dataset=testset, batch_size=batch_size, shuffle=False)
cifar10_classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
 
train_data_size = len(trainset)
valid_data_size = len(testset)
 
print('train_size: {:4d}  valid_size:{:4d}'.format(train_data_size, valid_data_size))
 
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
model = myModel(res=True)
 
# 定义损失函数和优化器
loss_func = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9, weight_decay=5e-3)
 
# 学习率调整策略 MultiStep:
scheduler = optim.lr_scheduler.MultiStepLR(optimizer=optimizer,
                                           milestones=[int(num_epochs * 0.56), int(num_epochs * 0.78)],
                                           gamma=0.1, last_epoch=-1)
 
# 训练和验证:
def train_and_valid(model, loss_function, optimizer, epochs=10):
    model.to(device)
    history = []
    best_acc = 0.0
    best_epoch = 0
 
    for epoch in range(epochs):
        epoch_start = time.time()
        print("Epoch: {}/{}".format(epoch + 1, epochs))
 
        model.train()
 
        train_loss = 0.0
        train_acc = 0.0
        valid_loss = 0.0
        valid_acc = 0.0
 
        for i, (inputs, labels) in enumerate(train_data):
            inputs = inputs.to(device)
            labels = labels.to(device)
 
            # 因为这里梯度是累加的,所以每次记得清零
            optimizer.zero_grad()
 
            outputs = model(inputs)
 
            loss = loss_function(outputs, labels)
 
            loss.backward()
 
            optimizer.step()
 
            train_loss += loss.item() * inputs.size(0)
 
            ret, predictions = torch.max(outputs.data, 1)
            correct_counts = predictions.eq(labels.data.view_as(predictions))
 
            acc = torch.mean(correct_counts.type(torch.FloatTensor))
 
            train_acc += acc.item() * inputs.size(0)
 
        with torch.no_grad():
            model.eval()
 
            for j, (inputs, labels) in enumerate(valid_data):
                inputs = inputs.to(device)
                labels = labels.to(device)
 
                outputs = model(inputs)
 
                loss = loss_function(outputs, labels)
 
                valid_loss += loss.item() * inputs.size(0)
 
                ret, predictions = torch.max(outputs.data, 1)
                correct_counts = predictions.eq(labels.data.view_as(predictions))
 
                acc = torch.mean(correct_counts.type(torch.FloatTensor))
 
                valid_acc += acc.item() * inputs.size(0)
        # 更新学习率并查看当前学习率
        scheduler.step()
        print('\t last_lr:', scheduler.get_last_lr())
 
        avg_train_loss = train_loss / train_data_size
        avg_train_acc = train_acc / train_data_size
 
        avg_valid_loss = valid_loss / valid_data_size
        avg_valid_acc = valid_acc / valid_data_size
 
        history.append([avg_train_loss, avg_valid_loss, avg_train_acc, avg_valid_acc])
 
        if best_acc < avg_valid_acc:
            best_acc = avg_valid_acc
            best_epoch = epoch + 1
 
        epoch_end = time.time()
 
        print(
            "\t Training: Loss: {:.4f}, Accuracy: {:.4f}%, "
            "\n\t Validation: Loss: {:.4f}, Accuracy: {:.4f}%, Time: {:.3f}s".format(
                avg_train_loss, avg_train_acc * 100, avg_valid_loss, avg_valid_acc * 100,
                                epoch_end - epoch_start
            ))
        print("\t Best Accuracy for validation : {:.4f} at epoch {:03d}".format(best_acc, best_epoch))
 
        torch.save(model, '%s/' % 'cifar10_my' + '%02d' % (epoch + 1) + '.pt'# 保存模型
 
        # # 存储模型为onnx格式:
        # d_cuda = torch.rand(1, 3, 32, 32, dtype=torch.float).to(device='cuda')
        # onnx_path = '%s/' % 'cifar10_shuffle' + '%02d' % (epoch + 1) + '.onnx'
        # torch.onnx.export(model.to('cuda'), d_cuda, onnx_path)
        # shape_path = '%s/' % 'cifar10_shuffle' + '%02d' % (epoch + 1) + '_shape.onnx'
        # onnx.save(onnx.shape_inference.infer_shapes(onnx.load(onnx_path)), shape_path)
        # print('\t export shape success...')
 
    return model, history
 
 
trained_model, history = train_and_valid(model, loss_func, optimizer, num_epochs)
 
history = np.array(history)
# Loss曲线
plt.figure(figsize=(10, 10))
plt.plot(history[:, 0:2])
plt.legend(['Tr Loss', 'Val Loss'])
plt.xlabel('Epoch Number')
plt.ylabel('Loss')
# 设置坐标轴刻度
plt.xticks(np.arange(0, num_epochs + 1, step=10))
plt.yticks(np.arange(0, 2.05, 0.1))
plt.grid()  # 画出网格
plt.savefig('cifar10_shuffle_' + '_loss_curve1.png')
 
# 精度曲线
plt.figure(figsize=(10, 10))
plt.plot(history[:, 2:4])
plt.legend(['Tr Accuracy', 'Val Accuracy'])
plt.xlabel('Epoch Number')
plt.ylabel('Accuracy')
# 设置坐标轴刻度
plt.xticks(np.arange(0, num_epochs + 1, step=10))
plt.yticks(np.arange(0, 1.05, 0.05))
plt.grid()  # 画出网格
plt.savefig('cifar10_shuffle_' + '_accuracy_curve1.png')
 
all_end = time.time()
all_time = round(all_end - all_start)
print('all time: ', all_time, ' 秒')
print("All Time: {:d} 分 {:d} 秒".format(all_time // 60, all_time % 60))

  

posted @   咖啡陪你  阅读(3077)  评论(0编辑  收藏  举报
相关博文:
阅读排行:
· DeepSeek “源神”启动!「GitHub 热点速览」
· 我与微信审核的“相爱相杀”看个人小程序副业
· 微软正式发布.NET 10 Preview 1:开启下一代开发框架新篇章
· C# 集成 DeepSeek 模型实现 AI 私有化(本地部署与 API 调用教程)
· spring官宣接入deepseek,真的太香了~
点击右上角即可分享
微信分享提示