不能替换Densenet121上的分类器

Question

我正在尝试使用这个github DenseNet121模型(https://github.com/gaetandi/cheXpert.git)进行一些转移学习。我遇到了一些问题，将分类层的大小从14个调整到2个。

github代码的相关部分是：

class DenseNet121(nn.Module):
    """Model modified.
    The architecture of our model is the same as standard DenseNet121
    except the classifier layer which has an additional sigmoid function.
    """
    def __init__(self, out_size):
        super(DenseNet121, self).__init__()
        self.densenet121 = torchvision.models.densenet121(pretrained=True)
        num_ftrs = self.densenet121.classifier.in_features
        self.densenet121.classifier = nn.Sequential(
            nn.Linear(num_ftrs, out_size),
            nn.Sigmoid()
        )
def forward(self, x):
    x = self.densenet121(x)
    return x

我装载并输入：

# initialize and load the model
model = DenseNet121(nnClassCount).cuda()
model = torch.nn.DataParallel(model).cuda()
modeldict = torch.load("model_ones_3epoch_densenet.tar")
model.load_state_dict(modeldict['state_dict'])

看起来，DenseNet不会将层划分成子层，因此model = nn.Sequential(*list(modelRes.children())[:-1])将无法工作。

model.classifier = nn.Linear(1024, 2)似乎在默认的DenseNets上工作，但是在这里使用修改后的分类器(附加的sigmoid函数)，它只是添加了一个额外的分类器层，而没有替换原来的分类器层。

我试过了

model.classifier = nn.Sequential(
    nn.Linear(1024, dset_classes_number), 
    nn.Sigmoid()
)

但我有相同的添加而不是替换的分类器问题：

...
      )
      (classifier): Sequential(
        (0): Linear(in_features=1024, out_features=14, bias=True)
        (1): Sigmoid()
      )
    )
  )
  (classifier): Sequential(
    (0): Linear(in_features=1024, out_features=2, bias=True)
    (1): Sigmoid()
  )
)

Answer 1

如果要替换作为classifier成员的densenet121内部的model，则需要分配

model.densenet121.classifier = nn.Sequential(...)

Answer 2

如果我理解您的问题，下面的代码将解决

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

import numpy as np

np.random.seed(0)
torch.manual_seed(0)




densenet121 = models.densenet121(pretrained=True)

for param in densenet121.parameters():
    param.requires_grad = False

densenet121.classifier = nn.Sequential(
    nn.Linear(1024, 14),
    nn.ReLU(),
    nn.Dropout(0.4),


    nn.Linear(14, 2),
)


densenet121.cuda()