Resnet 50 和VGG16 迁移学习fine tuning 的两种方法

最新推荐文章于 2024-07-24 18:05:59 发布

Zachary Zheng

最新推荐文章于 2024-07-24 18:05:59 发布

阅读量1.1k

点赞数

分类专栏： Pytorch 之computer vision 文章标签： python 深度学习

本文链接：https://blog.csdn.net/zachary_hao/article/details/104748866

版权

Pytorch 之computer vision 专栏收录该内容

2 篇文章 0 订阅

订阅专栏

修改Resnet 50 和VGG16 FC 层输出进行迁移学习方法仅供参考

Resnet50
方法一：

resnet50 = models.resnet50(pretrained=True)
print('Before:{%s}\n' % resnet50)
for param in resnet50.parameters(): param.requires_grad = False
fc_inputs = resnet50.fc.in_features
resnet50.fc = nn.Linear(fc_inputs,config.class_number)  # config.class_number 目标分类数
***#注意后面Loss用cross_entropy, cross_entropy 相当于logsoftmax()+NLLloss()***
loss = nn.CrossEntropyLoss()

方法二：

resnet50 = models.resnet50(pretrained=True)
print('Before:{%s}\n' % resnet50)
for param in resnet50.parameters(): param.requires_grad = False
num_ftrs = resnet50.fc.in_features
resnet50.fc = nn.Sequential(nn.Linear(num_ftrs, config.class_number),  # config.class_number 目标分类数,
                                   nn.LogSoftmax(dim=1))
***#注意后面Loss用NLLloss(), 因为已经定义了nn.LogSoftmax(dim=1)***
loss = nn.NLLLoss()

VGG16
方法一：

vgg = models.vgg16(pretrained=True)
print('Before:{%s}\n' % vgg)
for param in vgg.features.parameters():  param.requires_grad = False

vgg.classifier = nn.Sequential(torch.nn.Linear(25088, 4096),
                                torch.nn.ReLU(),
                                torch.nn.Dropout(p=0.5),
                                torch.nn.Linear(4096, 4096),
                                torch.nn.ReLU(),
                                torch.nn.Dropout(p=0.5),
                                torch.nn.Linear(4096, config.class_number))
print('========================================')
print('After:{%s}\n' % vgg)


loss_function = nn.CrossEntropyLoss()

方法二：

vgg = models.vgg16(pretrained=True)
print('Before:{%s}\n' % vgg)
for param in vgg.features.parameters():  param.requires_grad = False

vgg.classifier[6] = nn.Linear(4096, config.class_number)
print('========================================')
print('After:{%s}\n' % vgg)
loss_function = nn.CrossEntropyLoss()

~~方法三：~~ 方法三参考的是Vishnu Subramanian 的Deep Learning with PyTorch 书，但是不知道为什么修改后打印模型显示的输出out_features个数已经更改，Debug查看output.size 维度还是1000，不建议使用。

vgg = models.vgg16(pretrained=True)
print('Before:{%s}\n' % vgg)

for param in vgg.features.parameters():  param.requires_grad = False
# Freeze layers
vgg.classifier[6].out_features = config.class_number  # transfer learning, out_features for class number 这种修改方法还是有问题
print('========================================')
print('After:{%s}\n' % vgg)
loss_function = nn.CrossEntropyLoss()