一、pytorch中的pre-train模型
卷积神经网络的训练是耗时的,很多场合不可能每次都从随机初始化参数开始训练网络。
pytorch中自带几种常用的深度学习网络预训练模型,如VGG、ResNet等。往往为了加快学习的进度,在训练的初期我们直接加载pre-train模型中预先训练好的参数,model的加载如下所示:
1. import torchvision.models as models
2.
3. #resnet
4. model = models.ResNet(pretrained=True)
5. model = models.resnet18(pretrained=True)
6. model = models.resnet34(pretrained=True)
7. model = models.resnet50(pretrained=True)
8.
9. #vgg
10. model = models.VGG(pretrained=True)
11. model = models.vgg11(pretrained=True)
12. model = models.vgg16(pretrained=True)
13. model = models.vgg16_bn(pretrained=True)
二、预训练模型的修改
1.参数修改
对于简单的参数修改,这里以resnet预训练模型举例,resnet源代码在Github点击打开链接。
resnet网络最后一层分类层fc是对1000种类型进行划分,对于自己的数据集,如果只有9类,修改的代码如下:
1. # coding=UTF-8
2. import torchvision.models as models
3.
4. #调用模型
5. model = models.resnet50(pretrained=True)
6. #提取fc层中固定的参数
7. fc_features = model.fc.in_features
8. #修改类别为9
9. model.fc = nn.Linear(fc_features, 9)
2.增减卷积层
前一种方法只适用于简单的参数修改,有的时候我们往往要修改网络中的层次结构,这时只能用参数覆盖的方法,即自己先定义一个类似的网络,再将预训练中的参数提取到自己的网络中来。这里以resnet预训练模型举例。
1. # coding=UTF-8
2. import torchvision.models as models
3. import torch
4. import torch.nn as nn
5. import math
6. import torch.utils.model_zoo as model_zoo
7.
8. class CNN(nn.Module):
9.
10. def __init__(self, block, layers, num_classes=9):
11. self.inplanes = 64
12. self).__init__()
13. self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
14. False)
15. self.bn1 = nn.BatchNorm2d(64)
16. self.relu = nn.ReLU(inplace=True)
17. self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
18. self.layer1 = self._make_layer(block, 64, layers[0])
19. self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
20. self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
21. self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
22. self.avgpool = nn.AvgPool2d(7, stride=1)
23. #新增一个反卷积层
24. self.convtranspose1 = nn.ConvTranspose2d(2048, 2048, kernel_size=3, stride=1, padding=1, output_padding=0, groups=1, bias=False, dilation=1)
25. #新增一个最大池化层
26. self.maxpool2 = nn.MaxPool2d(kernel_size=3, stride=1, padding=1)
27. #去掉原来的fc层,新增一个fclass层
28. self.fclass = nn.Linear(2048, num_classes)
29.
30. for m in self.modules():
31. if isinstance(m, nn.Conv2d):
32. 0] * m.kernel_size[1] * m.out_channels
33. 0, math.sqrt(2. / n))
34. elif isinstance(m, nn.BatchNorm2d):
35. 1)
36. m.bias.data.zero_()
37.
38. def _make_layer(self, block, planes, blocks, stride=1):
39. None
40. if stride != 1 or self.inplanes != planes * block.expansion:
41. downsample = nn.Sequential(
42. self.inplanes, planes * block.expansion,
43. 1, stride=stride, bias=False),
44. nn.BatchNorm2d(planes * block.expansion),
45. )
46.
47. layers = []
48. self.inplanes, planes, stride, downsample))
49. self.inplanes = planes * block.expansion
50. for i in range(1, blocks):
51. self.inplanes, planes))
52.
53. return nn.Sequential(*layers)
54.
55. def forward(self, x):
56. self.conv1(x)
57. self.bn1(x)
58. self.relu(x)
59. self.maxpool(x)
60.
61. self.layer1(x)
62. self.layer2(x)
63. self.layer3(x)
64. self.layer4(x)
65.
66. self.avgpool(x)
67. #新加层的forward
68. 0), -1)
69. self.convtranspose1(x)
70. self.maxpool2(x)
71. 0), -1)
72. self.fclass(x)
73.
74. return x
75.
76. #加载model
77. resnet50 = models.resnet50(pretrained=True)
78. cnn = CNN(Bottleneck, [3, 4, 6, 3])
79. #读取参数
80. pretrained_dict = resnet50.state_dict()
81. model_dict = cnn.state_dict()
82. # 将pretrained_dict里不属于model_dict的键剔除掉
83. pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
84. # 更新现有的model_dict
85. model_dict.update(pretrained_dict)
86. # 加载我们真正需要的state_dict
87. cnn.load_state_dict(model_dict)
88. # print(resnet50)
89. print(cnn)
以上就是相关的内容,本人刚入门的小白一枚,请轻喷~
微调FineTune:
import torchvision
- import torch.optim as optim
import torch.nn as nn
# 局部微调
# 有时候我们加载了训练模型后,只想调节最后的几层,
# 其他层不训练。其实不训练也就意味着不进行梯度计算,PyTorch中提供的requires_grad使得对训练的控制变得非常简单。
model = torchvision.models.resnet18(pretrained=True)
for param in model.parameters():
param.requires_grad = False
# 替换最后的全连接层, 改为训练100类
# 新构造的模块的参数默认requires_grad为True
model.fc = nn.Linear(512, 100)
# 只优化最后的分类层
optimizer = optim.SGD(model.fc.parameters(), lr=1e-2, momentum=0.9)
###########################################################################################
# 全局微调
# 有时候我们需要对全局都进行finetune,只不过我们希望改换过的层和其他层的学习速率不一样,
# 这时候我们可以把其他层和新层在optimizer中单独赋予不同的学习速率。比如:
ignored_params = list(map(id, model.fc.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params,model.parameters())
#this is the new way to use Optimd
optimizer = optim.SGD([
{'params': base_params},
{'params': model.fc.parameters(), 'lr': 1e-3}
], lr=1e-2, momentum=0.9)