pytorch_模型

1 、官方模型引入

官方的模型保存在torchvision.models中，可以去pytorch官网查看模型。eg.torchvision.models.vgg16(pretrained=True)

2、在模型后面添加module

（1）.add_module在最末尾追加

vgg16_true.add_module("add_linear",nn.Linear(1000,10))

vgg16_true = torchvision.models.vgg16(pretrained=True)
gg16_true.add_module("add_linear",nn.Linear(1000,10))

结果对比：

(classifier): Sequential(
    (0): Linear(in_features=25088, out_features=4096, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=4096, out_features=4096, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=4096, out_features=1000, bias=True)
  )
)

  (classifier): Sequential(
    (0): Linear(in_features=25088, out_features=4096, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=4096, out_features=4096, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=4096, out_features=1000, bias=True)
  )
  (add_linear): Linear(in_features=1000, out_features=10, bias=True)
)

（2）在指定的序列内部追加module

vgg16_true.classifier.add_module("add_linear",nn.Linear(1000,10))

追加后的，追加前的看上面代码。

  (classifier): Sequential(
    (0): Linear(in_features=25088, out_features=4096, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=4096, out_features=4096, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=4096, out_features=1000, bias=True)
    (add_linear): Linear(in_features=1000, out_features=10, bias=True)
  )
)

3、对模型进行修改

vgg16_true.classifier[6]=nn.Linear(4098,10)

对序列的操作是取数组下标
修改前后对比：

  (classifier): Sequential(
    (0): Linear(in_features=25088, out_features=4096, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=4096, out_features=4096, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=4096, out_features=1000, bias=True)
  )
)

  (classifier): Sequential(
    (0): Linear(in_features=25088, out_features=4096, bias=True)
    (1): ReLU(inplace=True)
    (2): Dropout(p=0.5, inplace=False)
    (3): Linear(in_features=4096, out_features=4096, bias=True)
    (4): ReLU(inplace=True)
    (5): Dropout(p=0.5, inplace=False)
    (6): Linear(in_features=4098, out_features=10, bias=True)
  )
)

4、模型保存

模型的保存主要是，模型经过训练之后参数是不断调优的，这个时候要把调优后的模型保存下来。
（1）保存模型的参数

torch.save(model.state_dict(), '\parameter.pkl')#要保持的模型参数，保存位置
model = TheModelClass(...)
model.load_state_dict(torch.load('\parameter.pkl'))

（2）保存整个模型

torch.save(model, '\model.pkl')
model = torch.load('\model.pkl')