Finetune，只更新fc层的参数

只更新最后一层的参数

1
2
3
4
5
6
7
8
9
10
11
model = models.vgg19(pretrained=True)
# 冻结所有层的参数，finetune过程中这些参数不更新
for param in model.parameters():
    param.requires_grad = False
# torch.nn.Linear(in_features, out_features, bias=True)，获得in_features，即num_ftrs
num_ftrs = model.fc.in_features
# 训练集只有两类
# nn.Linear()的weights和bias默认requires_grad = True
model.fc = nn.Linear(num_ftrs, 2)
# 只更新fc层的参数，model.fc.parameters()
optimizer = optim.SGD(model.fc.parameters(), lr=0.001, momentum=0.9)

也可不改变原有网络，在其后增加若干层,像下面这样操作

1
2
3
4
5
6
7
8
9
10
11
12
13
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.vgg = models.vgg19(pretrained=True)
        for para in self.vgg.parameters():
            para.requires_grad = False  
        self.relu = nn.ReLU(inplace=True)
        self.fc_1000_2 = nn.Linear(1000, 2)  # requires_grad默认为True
    def forward(self, x):
        vgg_out = self.vgg(x)
        out = self.relu(vgg_out)
        out = self.fc_1000_2(out)   # vgg的输出再加一层，1000-->2
        return out