Pytorch模型遷移和遷移學習,導入部分模型參數的操作
1. 利用resnet18做遷移學習
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])
結果輸出:
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=<SliceBackward>)
如果,我們修改瞭resnet18的網絡結構,如何將原來預訓練模型參數(resnet18-5c106cde.pth)遷移到新的resnet18網絡中呢?
比如,這裡將官方的resnet18的self.layer4 = self._make_layer(block, 512, layers[3], stride=2)改為: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
這時,直接加載模型:
model = models.resnet18() model.load_state_dict(torch.load(pretrained_file))
這時,肯定會報錯,類似:Missing key(s) in state_dict或者Unexpected key(s) in state_dict的錯誤:
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”.
我們希望將原來預訓練模型參數(resnet18-5c106cde.pth)遷移到新的resnet18網絡,當然隻能遷移二者相同的模型參數,不同的參數還是隨機初始化的.
def transfer_model(pretrained_file, model): ''' 隻導入pretrained_file部分模型參數 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 # 在合並前(update),需要去除pretrained_dict一些不需要的參數 pretrained_dict = transfer_state_dict(pretrained_dict, model_dict) model_dict.update(pretrained_dict) # 更新(合並)模型的參數 model.load_state_dict(model_dict) return model def transfer_state_dict(pretrained_dict, model_dict): ''' 根據model_dict,去除pretrained_dict一些不需要的參數,以便遷移到新的網絡 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. 修改網絡名稱並遷移學習
上面的例子,隻是將官方的resnet18的self.layer4 = self._make_layer(block, 512, layers[3], stride=2)改為瞭:self.layer44 = self._make_layer(block, 512, layers[3], stride=2),我們僅僅是修改瞭一個網絡名稱而已,就導致 model.load_state_dict(torch.load(pretrained_file))出錯,
那麼,我們如何將預訓練模型”model/resnet18-5c106cde.pth”轉換成符合新的網絡的模型參數呢?
方法很簡單,隻需要將resnet18-5c106cde.pth的模型參數中所有前綴為layer4的名稱,改為layer44即可
本人已經定義好瞭方法:
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])
這時,輸出,跟之前一模一樣瞭。
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=<SliceBackward>)
3.去除原模型的某些模塊
下面是在不修改原模型代碼的情況下,通過”resnet18.named_children()”和”resnet18.children()”的方法去除子模塊”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)
補充:pytorch導入(部分)模型參數
背景介紹:
我的想法是把一個預訓練的網絡的參數導入到我的模型中,但是預訓練模型的參數隻是我模型參數的一小部分,怎樣導進去不出差錯瞭,請來聽我說說。
解法
首先把你需要添加參數的那一小部分模型提取出來,並新建一個類進行重新定義,如圖向Alexnet中添加前三層的參數,重新定義前三層。
接下來就是導入參數
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)
程序如上圖所示,主要是第三、四句,第三是替換,別人訓練的模型參數的鍵和自己的定義的會不一樣,所以需要替換成自己的;第四句有個if用於判斷導入需要的參數。其他語句都相當於是模板,套用即可。
以上為個人經驗,希望能給大傢一個參考,也希望大傢多多支持WalkonNet。如有錯誤或未考慮完全的地方,望不吝賜教。
推薦閱讀:
- pytorch 預訓練模型讀取修改相關參數的填坑問題
- pytorch 實現凍結部分參數訓練另一部分
- 解決pytorch 的state_dict()拷貝問題
- pytorch模型的保存和加載、checkpoint操作
- 解決Pytorch修改預訓練模型時遇到key不匹配的情況