如何在Pytorch中实现一个模型迁移功能-创新互联

这篇文章给大家介绍如何在Pytorch中实现一个模型迁移功能，内容非常详细，感兴趣的小伙伴们可以参考借鉴，希望对大家能有所帮助。

1. 利用resnet18做迁移学习

import torch
from torchvision import models 
if __name__ == "__main__":
  # device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  device = 'cpu'
  print("-----device:{}".format(device))
  print("-----Pytorch version:{}".format(torch.__version__))
 
  input_tensor = torch.zeros(1, 3, 100, 100)
  print('input_tensor:', input_tensor.shape)
  pretrained_file = "model/resnet18-5c106cde.pth"
  model = models.resnet18()
  model.load_state_dict(torch.load(pretrained_file))
  model.eval()
  out = model(input_tensor)
  print("out:", out.shape, out[0, 0:10])

结果输出:

input_tensor: torch.Size([1, 3, 100, 100])
out: torch.Size([1, 1000]) tensor([ 0.4010, 0.8436, 0.3072, 0.0627, 0.4446, 0.8470, 0.1882, 0.7012,0.2988, -0.7574], grad_fn=)

如果,我们修改了resnet18的网络结构,如何将原来预训练模型参数(resnet18-5c106cde.pth)迁移到新的resnet18网络中呢?

比如,这里将官方的resnet18的self.layer4 = self._make_layer(block, 512, layers[3], stride=2)改为:self.layer44 = self._make_layer(block, 512, layers[3], stride=2)

class ResNet(nn.Module): 
  def __init__(self, block, layers, num_classes=1000, zero_init_residual=False):
    super(ResNet, self).__init__()
    self.inplanes = 64
    self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                bias=False)
    self.bn1 = nn.BatchNorm2d(64)
    self.relu = nn.ReLU(inplace=True)
    self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
    self.layer1 = self._make_layer(block, 64, layers[0])
    self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
    self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
    self.layer44 = self._make_layer(block, 512, layers[3], stride=2)
    self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
    self.fc = nn.Linear(512 * block.expansion, num_classes)
 
    for m in self.modules():
      if isinstance(m, nn.Conv2d):
        nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
      elif isinstance(m, nn.BatchNorm2d):
        nn.init.constant_(m.weight, 1)
        nn.init.constant_(m.bias, 0)
 
    # Zero-initialize the last BN in each residual branch,
    # so that the residual branch starts with zeros, and each residual block behaves like an identity.
    # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
    if zero_init_residual:
      for m in self.modules():
        if isinstance(m, Bottleneck):
          nn.init.constant_(m.bn3.weight, 0)
        elif isinstance(m, BasicBlock):
          nn.init.constant_(m.bn2.weight, 0)
 
  def _make_layer(self, block, planes, blocks, stride=1):
    downsample = None
    if stride != 1 or self.inplanes != planes * block.expansion:
      downsample = nn.Sequential(
        conv1x1(self.inplanes, planes * block.expansion, stride),
        nn.BatchNorm2d(planes * block.expansion),
      )
 
    layers = []
    layers.append(block(self.inplanes, planes, stride, downsample))
    self.inplanes = planes * block.expansion
    for _ in range(1, blocks):
      layers.append(block(self.inplanes, planes))
 
    return nn.Sequential(*layers)
 
  def forward(self, x):
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.relu(x)
    x = self.maxpool(x)
 
    x = self.layer1(x)
    x = self.layer2(x)
    x = self.layer3(x)
    x = self.layer44(x)
 
    x = self.avgpool(x)
    x = x.view(x.size(0), -1)
    x = self.fc(x)
 
    return x

这时,直接加载模型:

  model = models.resnet18()
  model.load_state_dict(torch.load(pretrained_file))

这时,肯定会报错,类似：Missing key(s) in state_dict或者Unexpected key(s) in state_dict的错误：

RuntimeError: Error(s) in loading state_dict for ResNet:
Missing key(s) in state_dict: "layer44.0.conv1.weight", "layer44.0.bn1.weight", "layer44.0.bn1.bias", "layer44.0.bn1.running_mean", "layer44.0.bn1.running_var", "layer44.0.conv2.weight", "layer44.0.bn2.weight", "layer44.0.bn2.bias", "layer44.0.bn2.running_mean", "layer44.0.bn2.running_var", "layer44.0.downsample.0.weight", "layer44.0.downsample.1.weight", "layer44.0.downsample.1.bias", "layer44.0.downsample.1.running_mean", "layer44.0.downsample.1.running_var", "layer44.1.conv1.weight", "layer44.1.bn1.weight", "layer44.1.bn1.bias", "layer44.1.bn1.running_mean", "layer44.1.bn1.running_var", "layer44.1.conv2.weight", "layer44.1.bn2.weight", "layer44.1.bn2.bias", "layer44.1.bn2.running_mean", "layer44.1.bn2.running_var".
Unexpected key(s) in state_dict: "layer4.0.conv1.weight", "layer4.0.bn1.running_mean", "layer4.0.bn1.running_var", "layer4.0.bn1.weight", "layer4.0.bn1.bias", "layer4.0.conv2.weight", "layer4.0.bn2.running_mean", "layer4.0.bn2.running_var", "layer4.0.bn2.weight", "layer4.0.bn2.bias", "layer4.0.downsample.0.weight", "layer4.0.downsample.1.running_mean", "layer4.0.downsample.1.running_var", "layer4.0.downsample.1.weight", "layer4.0.downsample.1.bias", "layer4.1.conv1.weight", "layer4.1.bn1.running_mean", "layer4.1.bn1.running_var", "layer4.1.bn1.weight", "layer4.1.bn1.bias", "layer4.1.conv2.weight", "layer4.1.bn2.running_mean", "layer4.1.bn2.running_var", "layer4.1.bn2.weight", "layer4.1.bn2.bias".

Process finished with

RuntimeError: Error(s) in loading state_dict for ResNet:
Unexpected key(s) in state_dict: "layer4.0.conv1.weight", "layer4.0.bn1.running_mean", "layer4.0.bn1.running_var", "layer4.0.bn1.weight", "layer4.0.bn1.bias", "layer4.0.conv2.weight", "layer4.0.bn2.running_mean", "layer4.0.bn2.running_var", "layer4.0.bn2.weight", "layer4.0.bn2.bias", "layer4.0.downsample.0.weight", "layer4.0.downsample.1.running_mean", "layer4.0.downsample.1.running_var", "layer4.0.downsample.1.weight", "layer4.0.downsample.1.bias", "layer4.1.conv1.weight", "layer4.1.bn1.running_mean", "layer4.1.bn1.running_var", "layer4.1.bn1.weight", "layer4.1.bn1.bias", "layer4.1.conv2.weight", "layer4.1.bn2.running_mean", "layer4.1.bn2.running_var", "layer4.1.bn2.weight", "layer4.1.bn2.bias".

我们希望将原来预训练模型参数(resnet18-5c106cde.pth)迁移到新的resnet18网络,当然只能迁移二者相同的模型参数,不同的参数还是随机初始化的.

 def transfer_model(pretrained_file, model):
  '''
  只导入pretrained_file部分模型参数
  tensor([-0.7119, 0.0688, -1.7247, -1.7182, -1.2161, -0.7323, -2.1065, -0.5433,-1.5893, -0.5562]
  update:
    D.update([E, ]**F) -> None. Update D from dict/iterable E and F.
    If E is present and has a .keys() method, then does: for k in E: D[k] = E[k]
    If E is present and lacks a .keys() method, then does: for k, v in E: D[k] = v
    In either case, this is followed by: for k in F: D[k] = F[k]
  :param pretrained_file:
  :param model:
  :return:
  '''
  pretrained_dict = torch.load(pretrained_file) # get pretrained dict
  model_dict = model.state_dict() # get model dict
  # 在合并前(update),需要去除pretrained_dict一些不需要的参数
  pretrained_dict = transfer_state_dict(pretrained_dict, model_dict)
  model_dict.update(pretrained_dict) # 更新(合并)模型的参数
  model.load_state_dict(model_dict)
  return model
 
def transfer_state_dict(pretrained_dict, model_dict):
  '''
  根据model_dict,去除pretrained_dict一些不需要的参数,以便迁移到新的网络
  url: https://blog.csdn.net/qq_34914551/article/details/87871134
  :param pretrained_dict:
  :param model_dict:
  :return:
  '''
  # state_dict2 = {k: v for k, v in save_model.items() if k in model_dict.keys()}
  state_dict = {}
  for k, v in pretrained_dict.items():
    if k in model_dict.keys():
      # state_dict.setdefault(k, v)
      state_dict[k] = v
    else:
      print("Missing key(s) in state_dict :{}".format(k))
  return state_dict
 
if __name__ == "__main__":
 
  input_tensor = torch.zeros(1, 3, 100, 100)
  print('input_tensor:', input_tensor.shape)
  pretrained_file = "model/resnet18-5c106cde.pth"
  # model = resnet18()
  # model.load_state_dict(torch.load(pretrained_file))
  # model.eval()
  # out = model(input_tensor)
  # print("out:", out.shape, out[0, 0:10])
 
  model1 = resnet18()
  model1 = transfer_model(pretrained_file, model1)
  out1 = model1(input_tensor)
  print("out1:", out1.shape, out1[0, 0:10])

2. 修改网络名称并迁移学习

上面的例子,只是将官方的resnet18的self.layer4 = self._make_layer(block, 512, layers[3], stride=2)改为了:self.layer44 = self._make_layer(block, 512, layers[3], stride=2),我们仅仅是修改了一个网络名称而已,就导致 model.load_state_dict(torch.load(pretrained_file))出错,

那么,我们如何将预训练模型"model/resnet18-5c106cde.pth"转换成符合新的网络的模型参数呢?

方法很简单,只需要将resnet18-5c106cde.pth的模型参数中所有前缀为layer4的名称,改为layer44即可

本人已经定义好了方法:

modify_state_dict(pretrained_dict, model_dict, old_prefix, new_prefix)

def string_rename(old_string, new_string, start, end):
  new_string = old_string[:start] + new_string + old_string[end:]
  return new_string
 
def modify_model(pretrained_file, model, old_prefix, new_prefix):
  '''
  :param pretrained_file:
  :param model:
  :param old_prefix:
  :param new_prefix:
  :return:
  '''
  pretrained_dict = torch.load(pretrained_file)
  model_dict = model.state_dict()
  state_dict = modify_state_dict(pretrained_dict, model_dict, old_prefix, new_prefix)
  model.load_state_dict(state_dict)
  return model 
 
def modify_state_dict(pretrained_dict, model_dict, old_prefix, new_prefix):
  '''
  修改model dict
  :param pretrained_dict:
  :param model_dict:
  :param old_prefix:
  :param new_prefix:
  :return:
  '''
  state_dict = {}
  for k, v in pretrained_dict.items():
    if k in model_dict.keys():
      # state_dict.setdefault(k, v)
      state_dict[k] = v
    else:
      for o, n in zip(old_prefix, new_prefix):
        prefix = k[:len(o)]
        if prefix == o:
          kk = string_rename(old_string=k, new_string=n, start=0, end=len(o))
          print("rename layer modules:{}-->{}".format(k, kk))
          state_dict[kk] = v
  return state_dict

if __name__ == "__main__":
  input_tensor = torch.zeros(1, 3, 100, 100)
  print('input_tensor:', input_tensor.shape)
  pretrained_file = "model/resnet18-5c106cde.pth"
  # model = models.resnet18()
  # model.load_state_dict(torch.load(pretrained_file))
  # model.eval()
  # out = model(input_tensor)
  # print("out:", out.shape, out[0, 0:10])
  #
  # model1 = resnet18()
  # model1 = transfer_model(pretrained_file, model1)
  # out1 = model1(input_tensor)
  # print("out1:", out1.shape, out1[0, 0:10])
  #
  new_file = "new_model.pth"
  model = resnet18()
  new_model = modify_model(pretrained_file, model, old_prefix=["layer4"], new_prefix=["layer44"])
  torch.save(new_model.state_dict(), new_file)
 
  model2 = resnet18()
  model2.load_state_dict(torch.load(new_file))
  model2.eval()
  out2 = model2(input_tensor)
  print("out2:", out2.shape, out2[0, 0:10])

这时，输出，跟之前一模一样了。

out: torch.Size([1, 1000]) tensor([ 0.4010, 0.8436, 0.3072, 0.0627, 0.4446, 0.8470, 0.1882, 0.7012,0.2988, -0.7574], grad_fn=)

3.去除原模型的某些模块

下面是在不修改原模型代码的情况下,通过"resnet18.named_children()"和"resnet18.children()"的方法去除子模块"fc"和"avgpool"

import torch
import torchvision.models as models
from collections import OrderedDict
 
if __name__=="__main__":
  resnet18 = models.resnet18(False)
  print("resnet18",resnet18)
 
  # use named_children()
  resnet18_v1 = OrderedDict(resnet18.named_children())
  # remove avgpool,fc
  resnet18_v1.pop("avgpool")
  resnet18_v1.pop("fc")
  resnet18_v1 = torch.nn.Sequential(resnet18_v1)
  print("resnet18_v1",resnet18_v1)
  # use children
  resnet18_v2 = torch.nn.Sequential(*list(resnet18.children())[:-2])
  print(resnet18_v2,resnet18_v2)

补充：pytorch导入(部分)模型参数

背景介绍：

我的想法是把一个预训练的网络的参数导入到我的模型中，但是预训练模型的参数只是我模型参数的一小部分，怎样导进去不出差错了，请来听我说说。

解法

首先把你需要添加参数的那一小部分模型提取出来，并新建一个类进行重新定义，如图向Alexnet中添加前三层的参数，重新定义前三层。

接下来就是导入参数

checkpoint = torch.load(config.pretrained_model)
    # change name and load parameters
    model_dict = model.net1.state_dict()
    checkpoint = {k.replace('features.features', 'featureExtract1'): v for k, v in checkpoint.items()}
    checkpoint = {k:v for k,v in checkpoint.items() if k in model_dict.keys()}
 
    model_dict.update(checkpoint)
    model.net1.load_state_dict(model_dict)

关于如何在Pytorch中实现一个模型迁移功能就分享到这里了，希望以上内容可以对大家有一定的帮助，可以学到更多知识。如果觉得文章不错，可以把它分享出去让更多的人看到。