图1 Resnet的残差块
图2 Resnet18 网络架构
Cifar10 数据集的Resnet10的框架实现（Pytorch）:
1 import torch 2 from torch import nn 3 4 # ResNet18_BasicBlock-残差单元 5 class ResNet18_BasicBlock(nn.Module): 6 def __init__(self, input_channel, output_channel, stride, use_conv1_1): 7 super(ResNet18_BasicBlock, self).__init__() 8 9 # 第一层卷积 10 self.conv1 = nn.Conv2d(input_channel, output_channel, kernel_size=3, stride=stride, padding=1) 11 # 第二层卷积 12 self.conv2 = nn.Conv2d(output_channel, output_channel, kernel_size=3, stride=1, padding=1) 13 14 # 1*1卷积核，在不改变图片尺寸的情况下给通道升维 15 self.extra = nn.Sequential( 16 nn.Conv2d(input_channel, output_channel, kernel_size=1, stride=stride, padding=0), 17 nn.BatchNorm2d(output_channel) 18 ) 19 20 self.use_conv1_1 = use_conv1_1 21 22 self.bn = nn.BatchNorm2d(output_channel) 23 self.relu = nn.ReLU(inplace=True) 24 25 def forward(self, x): 26 out = self.bn(self.conv1(x)) 27 out = self.relu(out) 28 29 out = self.bn(self.conv2(out)) 30 31 # 残差连接-(B,C,H,W)维度一致才能进行残差连接 32 if self.use_conv1_1: 33 out = self.extra(x) + out 34 35 out = self.relu(out) 36 return out 37 38 39 # 构建 ResNet18 网络模型 40 class ResNet18(nn.Module): 41 def __init__(self): 42 super(ResNet18, self).__init__() 43 44 self.conv1 = nn.Sequential( 45 nn.Conv2d(3, 64, kernel_size=3, stride=3, padding=1), 46 nn.BatchNorm2d(64) 47 ) 48 self.block1_1 = ResNet18_BasicBlock(input_channel=64, output_channel=64, stride=1, use_conv1_1=False) 49 self.block1_2 = ResNet18_BasicBlock(input_channel=64, output_channel=64, stride=1, use_conv1_1=False) 50 51 self.block2_1 = ResNet18_BasicBlock(input_channel=64, output_channel=128, stride=2, use_conv1_1=True) 52 self.block2_2 = ResNet18_BasicBlock(input_channel=128, output_channel=128, stride=1, use_conv1_1=False) 53 54 self.block3_1 = ResNet18_BasicBlock(input_channel=128, output_channel=256, stride=2, use_conv1_1=True) 55 self.block3_2 = ResNet18_BasicBlock(input_channel=256, output_channel=256, stride=1, use_conv1_1=False) 56 57 self.block4_1 = ResNet18_BasicBlock(input_channel=256, output_channel=512, stride=2, use_conv1_1=True) 58 self.block4_2 = ResNet18_BasicBlock(input_channel=512, output_channel=512, stride=1, use_conv1_1=False) 59 60 self.FC_layer = nn.Linear(512 * 1 * 1, 10) 61 62 self.adaptive_avg_pool2d = nn.AdaptiveAvgPool2d((1,1)) 63 self.relu = nn.ReLU(inplace=True) 64 65 def forward(self, x): 66 67 x = self.relu(self.conv1(x)) 68 69 # ResNet18-网络模型 70 x = self.block1_1(x) 71 x = self.block1_2(x) 72 x = self.block2_1(x) 73 x = self.block2_2(x) 74 x = self.block3_1(x) 75 x = self.block3_2(x) 76 x = self.block4_1(x) 77 x = self.block4_2(x) 78 79 # 平均值池化 80 x = self.adaptive_avg_pool2d(x) 81 82 # 数据平坦化处理，为接下来的全连接层做准备 83 x = x.view(x.size(0), -1) 84 x = self.FC_layer(x) 85 86 return xView Code
classfyNet_main.py
1 import torch
2 from torch.utils.data import DataLoader
3 from torch import nn, optim
4 from torchvision import datasets, transforms
5
6 from matplotlib import pyplot as plt
7
8
9 import time
10
11 from Lenet5 import Lenet5_new
12 from Resnet18 import ResNet18
13
14 def main():
15
16 print("Load datasets...")
17
18 # transforms.RandomHorizontalFlip(p=0.5)---以0.5的概率对图片做水平横向翻转
19 # transforms.ToTensor()---shape从(H,W,C)->(C,H,W), 每个像素点从(0-255)映射到(0-1):直接除以255
20 # transforms.Normalize---先将输入归一化到(0,1),像素点通过"(x-mean)/std",将每个元素分布到(-1,1)
21 transform_train = transforms.Compose([
22 transforms.RandomHorizontalFlip(p=0.5),
23 transforms.ToTensor(),
24 transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
25 ])
26
27 transform_test = transforms.Compose([
28 transforms.ToTensor(),
29 transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
30 ])
31
32 # 内置函数下载数据集
33 train_dataset = datasets.CIFAR10(root="./data/Cifar10/", train=True,
34 transform = transform_train,
35 download=True)
36 test_dataset = datasets.CIFAR10(root = "./data/Cifar10/",
37 train = False,
38 transform = transform_test,
39 download=True)
40
41 print(len(train_dataset), len(test_dataset))
42
43 Batch_size = 64
44 train_loader = DataLoader(train_dataset, batch_size=Batch_size, shuffle = True)
45 test_loader = DataLoader(test_dataset, batch_size = Batch_size, shuffle = False)
46
47 # 设置CUDA
48 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
49
50 # 初始化模型
51 # 直接更换模型就行，其他无需操作
52 # model = Lenet5_new().to(device)
53 model = ResNet18().to(device)
54
55 # 构造损失函数和优化器
56 criterion = nn.CrossEntropyLoss() # 多分类softmax构造损失
57 opt = optim.SGD(model.parameters(), lr=0.001, momentum=0.8, weight_decay=0.001)
58
59 # 动态更新学习率 ------每隔step_size : lr = lr * gamma
60 schedule = optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.6, last_epoch=-1)
61
62 # 开始训练
63 print("Start Train...")
64
65 epochs = 100
66
67 loss_list = []
68 train_acc_list =[]
69 test_acc_list = []
70 epochs_list = []
71
72 for epoch in range(0, epochs):
73
74 start = time.time()
75
76 model.train()
77
78 running_loss = 0.0
79 batch_num = 0
80
81 for i, (inputs, labels) in enumerate(train_loader):
82
83 inputs, labels = inputs.to(device), labels.to(device)
84
85 # 将数据送入模型训练
86 outputs = model(inputs)
87 # 计算损失
88 loss = criterion(outputs, labels).to(device)
89
90 # 重置梯度
91 opt.zero_grad()
92 # 计算梯度，反向传播
93 loss.backward()
94 # 根据反向传播的梯度值优化更新参数
95 opt.step()
96
97 # 100个batch的 loss 之和
98 running_loss += loss.item()
99 # loss_list.append(loss.item())
100 batch_num+=1
101
102
103 epochs_list.append(epoch)
104
105 # 每一轮结束输出一下当前的学习率 lr
106 lr_1 = opt.param_groups[0]['lr']
107 print("learn_rate:%.15f" % lr_1)
108 schedule.step()
109
110 end = time.time()
111 print('epoch = %d/100, batch_num = %d, loss = %.6f, time = %.3f' % (epoch+1, batch_num, running_loss/batch_num, end-start))
112 running_loss=0.0
113
114 # 每个epoch训练结束，都进行一次测试验证
115 model.eval()
116 train_correct = 0.0
117 train_total = 0
118
119 test_correct = 0.0
120 test_total = 0
121
122 # 训练模式不需要反向传播更新梯度
123 with torch.no_grad():
124
125 # print("=======================train=======================")
126 for inputs, labels in train_loader:
127 inputs, labels = inputs.to(device), labels.to(device)
128 outputs = model(inputs)
129
130 pred = outputs.argmax(dim=1) # 返回每一行中最大值元素索引
131 train_total += inputs.size(0)
132 train_correct += torch.eq(pred, labels).sum().item()
133
134
135 # print("=======================test=======================")
136 for inputs, labels in test_loader:
137 inputs, labels = inputs.to(device), labels.to(device)
138 outputs = model(inputs)
139
140 pred = outputs.argmax(dim=1) # 返回每一行中最大值元素索引
141 test_total += inputs.size(0)
142 test_correct += torch.eq(pred, labels).sum().item()
143
144 print("train_total = %d, Accuracy = %.5f %%, test_total= %d, Accuracy = %.5f %%" %(train_total, 100 * train_correct / train_total, test_total, 100 * test_correct / test_total))
145
146 train_acc_list.append(100 * train_correct / train_total)
147 test_acc_list.append(100 * test_correct / test_total)
148
149 # print("Accuracy of the network on the 10000 test images:%.5f %%" % (100 * test_correct / test_total))
150 # print("===============================================")
151
152 fig = plt.figure(figsize=(4, 4))
153
154 plt.plot(epochs_list, train_acc_list, label='train_acc_list')
155 plt.plot(epochs_list, test_acc_list, label='test_acc_list')
156 plt.legend()
157 plt.title("train_test_acc")
158 plt.savefig('resnet18_cc_epoch_{:04d}.png'.format(epochs))
159 plt.close()
160
161 if __name__ == "__main__":
162
163 main()
View Code
标签：02,Resnet18,nn,self,train,图像,test,ResNet18,channel From： https://www.cnblogs.com/zhaopengpeng/p/17040590.html