脚本之家,脚本语言编程技术及教程分享平台!
分类导航

Python|VBS|Ruby|Lua|perl|VBA|Golang|PowerShell|Erlang|autoit|Dos|bat|

服务器之家 - 脚本之家 - Python - 自己搭建resnet18网络并加载torchvision自带权重的操作

自己搭建resnet18网络并加载torchvision自带权重的操作

2021-11-01 10:08找不到服务器1703 Python

这篇文章主要介绍了自己搭建resnet18网络并加载torchvision自带权重的操作,具有很好的参考价值,希望对大家有所帮助。如有错误或未考虑完全的地方,望不吝赐教

直接搭建网络必须与torchvision自带的网络的权重也就是pth文件的结构、尺寸和变量命名完全一致,否则无法加载权重文件。

此时可比较2个字典逐一加载,详见

pytorch加载预训练模型与自己模型不匹配的解决方案

?
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
import torch
import torchvision
import cv2 as cv
from utils.utils import letter_box
from model.backbone import ResNet18
 
model1 = ResNet18(1)
model2 = torchvision.models.resnet18(progress=False)
fc = model2.fc
model2.fc = torch.nn.Linear(512, 1)
# print(model)
model_dict1 = model1.state_dict()
model_dict2 = torch.load('resnet18.pth')
model_list1 = list(model_dict1.keys())
model_list2 = list(model_dict2.keys())
len1 = len(model_list1)
len2 = len(model_list2)
minlen = min(len1, len2)
for n in range(minlen):
    if model_dict1[model_list1[n]].shape != model_dict2[model_list2[n]].shape:
        continue
    model_dict1[model_list1[n]] = model_dict2[model_list2[n]]
model1.load_state_dict(model_dict1)
missing, unspected = model2.load_state_dict(model_dict2)
image = cv.imread('zhn1.jpg')
image = letter_box(image, 224)
image = image[:, :, ::-1].transpose(2, 0, 1)
print('Network loading complete.')
model1.eval()
model2.eval()
with torch.no_grad():
    image = torch.tensor(image/256, dtype=torch.float32).unsqueeze(0)
    predict1 = model1(image)
    predict2 = model2(image)
print('finished')
# torch.save(model.state_dict(), 'resnet18.pth')

以上为全部程序,最终可测试原模型与加载了自带权重的自定义模型的输出是否相等。

补充:使用Pytorch搭建ResNet分类网络并基于迁移学习训练

如果stride=1,padding=1

卷积处理是不会改变特征矩阵的高和宽

使用BN层时

卷积中的参数bias置为False(有无偏置BN层的输出都相同),BN层放在conv层和relu层的中间

复习BN层:

Batch Norm 层是对每层数据归一化后再进行线性变换改善数据分布, 其中的线性变换是可学习的.

Batch Norm优点:减轻过拟合;改善梯度传播(权重不会过高或过低)容许较高的学习率,能够提高训练速度。减轻对初始化权重的强依赖,使得数据分布在激活函数的非饱和区域,一定程度上解决梯度消失问题。作为一种正则化的方式,在某种程度上减少对dropout的使用。

Batch Norm层摆放位置:在激活层(如 ReLU )之前还是之后,没有一个统一的定论。

BN层与 Dropout 合作:Batch Norm的提出使得dropout的使用减少,但是Batch Norm不能完全取代dropout,保留较小的dropout率,如0.2可能效果更佳。

为什么要先normalize再通过γ,β线性变换恢复接近原来的样子,这不是多此一举吗?

在一定条件下可以纠正原始数据的分布(方差,均值变为新值γ,β),当原始数据分布足够好时就是恒等映射,不改变分布。如果不做BN,方差和均值对前面网络的参数有复杂的关联依赖,具有复杂的非线性。在新参数 γH′ + β 中仅由 γ,β 确定,与前边网络的参数无关,因此新参数很容易通过梯度下降来学习,能够学习到较好的分布。

迁移学习导入权重和下载权重:

?
1
2
3
4
5
6
7
import torchvision.models.resnet#ctrl+鼠标左键点击即可下载权重
net = resnet34()#一开始不能设置全连接层的输出种类为自己想要的,必须先将模型参数载入,再修改全连接层
# 官方提供载入预训练模型的方法
model_weight_path = "./resnet34-pre.pth"#权重路径
missing_keys, unexpected_keys = net.load_state_dict(torch.load(model_weight_path), strict=False)#载入模型权重
inchannel = net.fc.in_features
net.fc = nn.Linear(inchannel, 5)#重新确定全连接层

完整代码:

model部分:

?
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
import torch.nn as nn
import torch
class BasicBlock(nn.Module):#对应18层和34层所对应的残差结构(既要有实线残差结构功能,也要有虚线残差结构功能)
    expansion = 1#残差结构主分支上的三个卷积层是否相同,相同为1,第三层是一二层四倍则为4
    def __init__(self, in_channel, out_channel, stride=1, downsample=None):#downsample代表虚线残差结构选项
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel,
                               kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channel)
        self.relu = nn.ReLU()
        self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel,
                               kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(out_channel)
        self.downsample = downsample
    def forward(self, x):
        identity = x
        if self.downsample is not None:
            identity = self.downsample(x)#得到捷径分支的输出
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        out += identity
        out = self.relu(out)
        return out#得到残差结构的最终输出
 
class Bottleneck(nn.Module):#对应50层、101层和152层所对应的残差结构
    expansion = 4#第三层卷积核个数是第一层和第二层的四倍
    def __init__(self, in_channel, out_channel, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel,
                               kernel_size=1, stride=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channel)
        self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel,
                               kernel_size=3, stride=stride, bias=False, padding=1)
        self.bn2 = nn.BatchNorm2d(out_channel)
        self.conv3 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel*self.expansion,
                               kernel_size=1, stride=1, bias=False)
        self.bn3 = nn.BatchNorm2d(out_channel*self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
    def forward(self, x):
        identity = x
        if self.downsample is not None:
            identity = self.downsample(x)
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)
        out = self.conv3(out)
        out = self.bn3(out)
        out += identity
        out = self.relu(out)
        return out
 
class ResNet(nn.Module):#定义整个网络的框架部分
#blocks_num是残差结构的数目,是一个列表参数,block对应哪个残差模块
    def __init__(self, block, blocks_num, num_classes=1000, include_top=True):
        super(ResNet, self).__init__()
        self.include_top = include_top
        self.in_channel = 64#通过第一个池化层后所得到的特征矩阵的深度
        self.conv1 = nn.Conv2d(3, self.in_channel, kernel_size=7, stride=2,
                               padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(self.in_channel)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, blocks_num[0])
        self.layer2 = self._make_layer(block, 128, blocks_num[1], stride=2)
        self.layer3 = self._make_layer(block, 256, blocks_num[2], stride=2)
        self.layer4 = self._make_layer(block, 512, blocks_num[3], stride=2)
        if self.include_top:
            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))  # output size = (1, 1)
            self.fc = nn.Linear(512 * block.expansion, num_classes)
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
    def _make_layer(self, block, channel, block_num, stride=1):#channel:残差结构中,第一个卷积层所使用的卷积核的个数
        downsample = None
        if stride != 1 or self.in_channel != channel * block.expansion:#18层和34层会直接跳过这个if语句
            downsample = nn.Sequential(
                nn.Conv2d(self.in_channel, channel * block.expansion, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(channel * block.expansion))
        layers = []
        layers.append(block(self.in_channel, channel, downsample=downsample, stride=stride))
        self.in_channel = channel * block.expansion
        for _ in range(1, block_num):
            layers.append(block(self.in_channel, channel))
        return nn.Sequential(*layers)
    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        if self.include_top:#默认是true
            x = self.avgpool(x)
            x = torch.flatten(x, 1)
            x = self.fc(x)
        return x
 
def resnet34(num_classes=1000, include_top=True):
    return ResNet(BasicBlock, [3, 4, 6, 3], num_classes=num_classes, include_top=include_top)
 
def resnet101(num_classes=1000, include_top=True):
    return ResNet(Bottleneck, [3, 4, 23, 3], num_classes=num_classes, include_top=include_top)

训练部分:

?
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
import torch
import torch.nn as nn
from torchvision import transforms, datasets
import json
import matplotlib.pyplot as plt
import os
import torch.optim as optim
from model import resnet34, resnet101
import torchvision.models.resnet#ctrl+鼠标左键点击即可下载权重
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
data_transform = {
    "train": transforms.Compose([transforms.RandomResizedCrop(224),
                                 transforms.RandomHorizontalFlip(),
                                 transforms.ToTensor(),
                                 transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]),#和官网初始化方法保持一致
    "val": transforms.Compose([transforms.Resize(256),
                               transforms.CenterCrop(224),
                               transforms.ToTensor(),
                               transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])}
 
data_root = os.path.abspath(os.path.join(os.getcwd(), "../.."))  # get data root path
image_path = data_root + "/data_set/flower_data/"  # flower data set path
train_dataset = datasets.ImageFolder(root=image_path+"train",
                                     transform=data_transform["train"])
train_num = len(train_dataset)
# {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}
flower_list = train_dataset.class_to_idx
cla_dict = dict((val, key) for key, val in flower_list.items())
# write dict into json file
json_str = json.dumps(cla_dict, indent=4)
with open('class_indices.json', 'w') as json_file:
    json_file.write(json_str)
batch_size = 16
train_loader = torch.utils.data.DataLoader(train_dataset,
                                           batch_size=batch_size, shuffle=True,
                                           num_workers=0)
validate_dataset = datasets.ImageFolder(root=image_path + "val",
                                        transform=data_transform["val"])
val_num = len(validate_dataset)
validate_loader = torch.utils.data.DataLoader(validate_dataset,
                                              batch_size=batch_size, shuffle=False,
                                              num_workers=0)
net = resnet34()#一开始不能设置全连接层的输出种类为自己想要的,必须先将模型参数载入,再修改全连接层
# 官方提供载入预训练模型的方法
model_weight_path = "./resnet34-pre.pth"#权重路径
missing_keys, unexpected_keys = net.load_state_dict(torch.load(model_weight_path), strict=False)#载入模型权重
inchannel = net.fc.in_features
net.fc = nn.Linear(inchannel, 5)#重新确定全连接层
net.to(device)
loss_function = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.0001)
best_acc = 0.0
save_path = './resNet34.pth'
for epoch in range(3):
    # train
    net.train()#控制BN层状态
    running_loss = 0.0
    for step, data in enumerate(train_loader, start=0):
        images, labels = data
        optimizer.zero_grad()
        logits = net(images.to(device))
        loss = loss_function(logits, labels.to(device))
        loss.backward()
        optimizer.step()
        # print statistics
        running_loss += loss.item()
        # print train process
        rate = (step+1)/len(train_loader)
        a = "*" * int(rate * 50)
        b = "." * int((1 - rate) * 50)
        print("\rtrain loss: {:^3.0f}%[{}->{}]{:.4f}".format(int(rate*100), a, b, loss), end="")
    print()
    # validate
    net.eval()#控制BN层状态
    acc = 0.0  # accumulate accurate number / epoch
    with torch.no_grad():
        for val_data in validate_loader:
            val_images, val_labels = val_data
            outputs = net(val_images.to(device))  # eval model only have last output layer
            # loss = loss_function(outputs, test_labels)
            predict_y = torch.max(outputs, dim=1)[1]
            acc += (predict_y == val_labels.to(device)).sum().item()
        val_accurate = acc / val_num
        if val_accurate > best_acc:
            best_acc = val_accurate
            torch.save(net.state_dict(), save_path)
        print('[epoch %d] train_loss: %.3f  test_accuracy: %.3f' %
              (epoch + 1, running_loss / step, val_accurate))
print('Finished Training')

预测部分:

?
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
import torch
from model import resnet34
from PIL import Image
from torchvision import transforms
import matplotlib.pyplot as plt
import json
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
data_transform = transforms.Compose(
    [transforms.Resize(256),
     transforms.CenterCrop(224),
     transforms.ToTensor(),
     transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])#采用和训练方法一样的标准化处理
# load image
img = Image.open("../aa.jpg")
plt.imshow(img)
# [N, C, H, W]
img = data_transform(img)
# expand batch dimension
img = torch.unsqueeze(img, dim=0)
# read class_indict
try:
    json_file = open('./class_indices.json', 'r')
    class_indict = json.load(json_file)
except Exception as e:
    print(e)
    exit(-1)
# create model
model = resnet34(num_classes=5)
# load model weights
model_weight_path = "./resNet34.pth"
model.load_state_dict(torch.load(model_weight_path, map_location=device))#载入训练好的模型参数
model.eval()#使用eval()模式
with torch.no_grad():#不跟踪损失梯度
    # predict class
    output = torch.squeeze(model(img))#压缩batch维度
    predict = torch.softmax(output, dim=0)#通过softmax得到概率分布
    predict_cla = torch.argmax(predict).numpy()#寻找最大值所对应的索引
print(class_indict[str(predict_cla)], predict[predict_cla].numpy())#打印类别信息和概率
plt.show()

以上为个人经验,希望能给大家一个参考,也希望大家多多支持服务器之家。

原文链接:https://blog.csdn.net/qq_34288751/article/details/114163057

延伸 · 阅读

精彩推荐