Pytorch下微调网络模型进行图像分类

利用ImageNet下的预训练权重采用迁移学习策略，能够实现模型快速训练，提高图像分类性能。下面以vgg和resnet网络模型为例，微调最后的分类层进行分类。

 注意，微调只对分类层(也就是全连接层)的参数进行更新，前面的参数需要被冻结。

（1）微调VGG模型进行图像分类（以vgg16为例）

import torch
import torch.nn as nn
import torchvision.models as models

classes_num = 200 # 数据集的类别数

model = models.vgg16(pretrained=True)
for parameter in model.parameters():
    parameter.required_grad = False
model.classifier = nn.Sequential(nn.Linear(512*7*7, 4096),
                                 nn.ReLU(inplace=True),
                                 nn.Dropout(0.5),
                                 nn.Linear(4096, 4096),
                                 nn.ReLU(inplace=True),
                                 nn.Dropout(0.5),
                                 nn.Linear(4096, classes_num))
model = model.cuda()
print(model)

 

(2)微调ResNet模型进行图像分类(以ResNet-34为例)

import torch
import torch.nn as nn
import torchvision.models as models

classes_num = 200 # 数据集的类别数

model = models.resnet34(pretrained=True)
for parameter in model.parameters():
    parameter.required_grad = False
model.classifier = nn.Linear(512, classes_num)
model = model.cuda()
print(model)