python 运行时错误：加载state_dict时出错

bounty将在6天后过期。回答此问题可获得+500声望奖励。Paul Reiners希望引起更多人关注此问题。

我有以下PyTorch模型：

import math
from abc import abstractmethod

import torch.nn as nn

class AlexNet3D(nn.Module):
    @abstractmethod
    def get_head(self):
        pass

    def __init__(self, input_size):
        super().__init__()
        self.input_size = input_size
        self.features = nn.Sequential(
            nn.Conv3d(1, 64, kernel_size=(5, 5, 5), stride=(2, 2, 2), padding=0),
            nn.BatchNorm3d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=3, stride=3),

            nn.Conv3d(64, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=0),
            nn.BatchNorm3d(128),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=3, stride=3),

            nn.Conv3d(128, 192, kernel_size=(3, 3, 3), padding=1),
            nn.BatchNorm3d(192),
            nn.ReLU(inplace=True),

            nn.Conv3d(192, 192, kernel_size=(3, 3, 3), padding=1),
            nn.BatchNorm3d(192),
            nn.ReLU(inplace=True),

            nn.Conv3d(192, 128, kernel_size=(3, 3, 3), padding=1),
            nn.BatchNorm3d(128),
            nn.ReLU(inplace=True),
            nn.MaxPool3d(kernel_size=3, stride=3),
        )

        self.classifier = self.get_head()

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
            elif isinstance(m, nn.BatchNorm3d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()

    def forward(self, x):
        xp = self.features(x)
        x = xp.view(xp.size(0), -1)
        x = self.classifier(x)
        return [x, xp]

class AlexNet3DDropoutRegression(AlexNet3D):
    def get_head(self):
        return nn.Sequential(nn.Dropout(),
                             nn.Linear(self.input_size, 64),
                             nn.ReLU(inplace=True),
                             nn.Dropout(),
                             nn.Linear(64, 1),
                             )

我这样初始化模型：

def init_model(self):
    model = AlexNet3DDropoutRegression(4608)
    if self.use_cuda:
        log.info("Using CUDA; {} devices.".format(torch.cuda.device_count()))
        if torch.cuda.device_count() > 1:
            model = nn.DataParallel(model)
        model = model.to(self.device)
    return model

训练后，我将按如下方式保存模型：

torch.save(self.model.state_dict(), self.cli_args.model_save_location)

然后，我尝试加载保存的模型：

import torch
from reprex.models import AlexNet3DDropoutRegression

model_save_location = "/home/feczk001/shared/data/AlexNet/LoesScoring/loes_scoring_01.pt"

model = AlexNet3DDropoutRegression(4608)
model.load_state_dict(torch.load(model_save_location,
                                 map_location='cpu'))

但我得到了以下错误：

RuntimeError: Error(s) in loading state_dict for AlexNet3DDropoutRegression:
    Missing key(s) in state_dict: "features.0.weight", "features.0.bias", "features.1.weight", "features.1.bias", "features.1.running_mean", "features.1.running_var", "features.4.weight", "features.4.bias", "features.5.weight", "features.5.bias", "features.5.running_mean", "features.5.running_var", "features.8.weight", "features.8.bias", "features.9.weight", "features.9.bias", "features.9.running_mean", "features.9.running_var", "features.11.weight", "features.11.bias", "features.12.weight", "features.12.bias", "features.12.running_mean", "features.12.running_var", "features.14.weight", "features.14.bias", "features.15.weight", "features.15.bias", "features.15.running_mean", "features.15.running_var", "classifier.1.weight", "classifier.1.bias", "classifier.4.weight", "classifier.4.bias". 
    Unexpected key(s) in state_dict: "module.features.0.weight", "module.features.0.bias", "module.features.1.weight", "module.features.1.bias", "module.features.1.running_mean", "module.features.1.running_var", "module.features.1.num_batches_tracked", "module.features.4.weight", "module.features.4.bias", "module.features.5.weight", "module.features.5.bias", "module.features.5.running_mean", "module.features.5.running_var", "module.features.5.num_batches_tracked", "module.features.8.weight", "module.features.8.bias", "module.features.9.weight", "module.features.9.bias", "module.features.9.running_mean", "module.features.9.running_var", "module.features.9.num_batches_tracked", "module.features.11.weight", "module.features.11.bias", "module.features.12.weight", "module.features.12.bias", "module.features.12.running_mean", "module.features.12.running_var", "module.features.12.num_batches_tracked", "module.features.14.weight", "module.features.14.bias", "module.features.15.weight", "module.features.15.bias", "module.features.15.running_mean", "module.features.15.running_var", "module.features.15.num_batches_tracked", "module.classifier.1.weight", "module.classifier.1.bias", "module.classifier.4.weight", "module.classifier.4.bias".

这是怎么回事？

问题是您使用DataParallel训练模型，然后尝试在非并行网络中重新加载模型。DataParallel是一个 Package 器类，用于创建原始模型（torch.nn.module对象）名为module的DataParallel对象的类属性。此问题在pytorch discuss上解决，stack overflow和github，所以我也不会在这里重复细节，但是您可以通过以下任一方法修复此问题：
1.以DataParallel对象的形式专门保存和加载模型，当您希望使用模型进行推理时，该对象可能不再有效，或者
1.保存DataParallel对象的modulestate_dict，如下所示：

# save state dict of DataParallel object
torch.save(model.module.state_dict(), path)

 .... Later
# reload weights on non-parallel model
model.load_state_dict(torch.load(path)

下面是一个简单的例子：

model = AlexNet3DDropoutRegression(4608) # on cpu
model = nn.DataParallel(model)
model = model.to("cuda") # DataParallel object on GPU(s)

torch.save(model.module.state_dict(),"example_path.pt")

del model
model = AlexNet3DDropoutRegression(4608)

ret = model.load_state_dict(torch.load("example_path.pt")) 
print(ret)

输出：

>>> <All keys successfully matched>

1.或者，如果您已经有一个保存的state_dict需要重新加载，您也可以为DataParallel模型加载state_dict，重新Map键名以排除“module”，然后使用重新键入的state_dict，这可能更有用。

incompatible_state_dict = torch.load("DataParallel_save_file.pt")
state_dict = {}
for key in incompatible_state_dict():
    state_dict[key.split("module.")[-1]] = incompatible_state_dict[key]

 ret = model.load_state_dict(state_dict)
 print(ret)

输出：

>>> <All keys successfully matched>

python 运行时错误：加载state_dict时出错

1条答案

相关问题

热门标签

最新问答