加载预训练AlexNet模型时遇AttributeError：'dict'无'features'属性求助

Hey, let's fix this error right away! The problem here is that your alexnet_places365.pth.tar file doesn't contain a full PyTorch model instance—it's just a state_dict (a dictionary storing the model's trained parameters). When you call torch.load() on it, you get a dict, not a model with a features attribute, hence the AttributeError.

Here's how to fix it step by step:

• First, initialize the base AlexNet model structure (the Places365 AlexNet uses the same architecture as the standard AlexNet in torchvision).
• Load the pretrained weights from the .pth.tar file (which is a dict), then map those weights to your base model.
• Then you can safely extract the features layers you need.

Revised Full Code #

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

class AlexSal(nn.Module):
    def __init__(self):
        super(AlexSal, self).__init__()
        # Step 1: Initialize the base AlexNet structure
        base_alexnet = models.alexnet(pretrained=False)
        
        # Step 2: Load the pretrained weights dict and adjust for possible prefixes
        pretrained_weights = torch.load('alexnet_places365.pth.tar')
        # Many Places365 weights are saved with a 'module.' prefix (from distributed training)
        # We need to strip that to match the base model's parameter names
        if 'state_dict' in pretrained_weights:
            weight_dict = pretrained_weights['state_dict']
        else:
            weight_dict = pretrained_weights
        # Remove 'module.' prefix if present
        cleaned_weight_dict = {k.replace('module.', ''): v for k, v in weight_dict.items()}
        
        # Load the cleaned weights into the base model
        base_alexnet.load_state_dict(cleaned_weight_dict)
        
        # Step 3: Extract the first N layers of features (exclude last 2)
        self.features = nn.Sequential(*list(base_alexnet.features.children())[:-2])
        
        # Remaining layers stay the same
        self.relu = nn.ReLU()
        self.sigmoid = nn.Sigmoid()
        self.conv6 = nn.Conv2d(256, 1, kernel_size=(1, 1), stride=(1, 1))

    def forward(self, x):
        x = self.relu(self.features(x))
        x = self.sigmoid(self.conv6(x))
        x = x.squeeze(1)
        return x

model = AlexSal().cuda()

Quick Debug Tip #

If you're unsure about the structure of your pretrained file, print its keys first to confirm:

pretrained_weights = torch.load('alexnet_places365.pth.tar')
print(pretrained_weights.keys())

This will tell you if the weights are stored directly in the dict or under a state_dict key, so you can adjust the loading logic accordingly.

This should resolve the AttributeError and let you initialize your AlexSal model correctly.

内容的提问来源于stack exchange，提问作者Mr.Danish Mukhtar Zargar