PyTorch AutoEncoder报错：mat1与mat2无法相乘(1x512和12x64)

我通过Hook提取CNN输出的特征并转换为张量：extracted_features = torch.tensor(rn_output)，转换后张量尺寸为torch.Size([1014,512])。但AutoEncoder的decoder报出“无法相乘”错误，问题源于输入设置与形状逻辑错误。

原AutoEncoder代码 #

class AutoEncoder(nn.Module):
    def __init__(self):
        super(AutoEncoder, self).__init__()
        self.encoder = nn.Sequential(
            nn.Linear(in_features=512, out_features=256),
            nn.ReLU(),
            nn.Linear(in_features=256, out_features=128),
            nn.ReLU(),
            nn.Linear(in_features=128, out_features=64),
            nn.ReLU(),
            nn.Linear(in_features=64, out_features=12),
        )
        self.decoder = nn.Sequential(
            nn.Linear(in_features=12, out_features=64),
            nn.ReLU(),
            nn.Linear(in_features=64, out_features=128),
            nn.Linear(in_features=128, out_features=256),
            nn.ReLU(),
            nn.Linear(in_features=256, out_features=512),
            nn.Tanh()
        )
    
    def forward(self, x):
        encoded = self.encoder(x)
        decoded = self.decoder(x)
        return decoded

原调用代码 #

model = AutoEncoder()
criterion = nn.MSELoss()
optimiser = torch.optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-5)

num_epochs = 10
outputs = []
for epoch in range(num_epochs): 
    for (img) in extracted_features:
        recon = model(img)
        loss = criterion(recon, img)
        
        optimiser.zero_grad()
        loss.backward()
        optimiser.step()
        
    print(f'Epoch:{epoch+1}, Loss:{loss.item():.4f}')
    outputs.append((epoch, img, recon))

问题分析 #

1. forward逻辑错误：解码器接收的是原始输入x（维度512），但解码器第一层Linear(in_features=12)要求输入维度为12，直接导致矩阵维度不匹配，触发相乘错误。
2. 单样本输入维度错误：遍历extracted_features时，每个img是一维张量[512]，PyTorch的Linear层默认处理二维批量输入[batch_size, feature_dim]，单样本会被误判为[512, 1]，进一步引发维度冲突。

修复方案 #

1. 修正forward方法逻辑 #

将编码器的输出传给解码器，而不是原始输入：

def forward(self, x):
    encoded = self.encoder(x)
    decoded = self.decoder(encoded)  # 用编码后的特征作为解码器输入
    return decoded

2. 调整输入维度为批量格式 #

直接使用二维张量批量输入，或用DataLoader包装：

from torch.utils.data import TensorDataset, DataLoader

# 用DataLoader包装特征张量，按批次处理
dataset = TensorDataset(extracted_features)
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

num_epochs = 10
outputs = []
for epoch in range(num_epochs): 
    total_loss = 0.0
    for batch in dataloader:
        x = batch[0]  # 取出批次特征，维度为[32, 512]
        recon = model(x)
        loss = criterion(recon, x)
        
        optimiser.zero_grad()
        loss.backward()
        optimiser.step()
        
        total_loss += loss.item()
    
    avg_loss = total_loss / len(dataloader)
    print(f'Epoch:{epoch+1}, Loss:{avg_loss:.4f}')
    outputs.append((epoch, x, recon))

3. 额外检查 #

确保extracted_features的数据类型（如float32）与模型参数一致，避免类型不匹配问题。

内容的提问来源于stack exchange，提问作者Mark McAllister