PyTorch AttributeError排查：并行CNN+Transformer情感分类模型报错

错误原因 #

出现AttributeError: 'NoneType' object has no attribute 'size'的核心原因是torchsummary不支持模型返回多个输出张量。你的模型forward方法返回了output_logits和output_softmax两个结果，但torchsummary默认只能处理返回单个张量的模型，在追踪模型结构时会因无法解析多输出产生None值，进而触发该错误。

另外先确认卷积模块输出维度与全连接层输入维度的匹配性：单个卷积块输出flatten后为64*1*8=512，两个卷积块总维度是512*2=1024，加上Transformer的40维嵌入，总维度1024+40=1064，你的全连接层nn.Linear(512*2+40,num_emotions)维度设置是正确的。

解决方法 #

方法1：临时修改forward适配torchsummary #

运行summary前，临时让模型只返回一个输出张量：

def forward(self,x):
    # 原forward逻辑保持不变
    conv2d_embedding1 = self.conv2Dblock1(x)
    conv2d_embedding1 = torch.flatten(conv2d_embedding1, start_dim=1) 
    
    conv2d_embedding2 = self.conv2Dblock2(x)
    conv2d_embedding2 = torch.flatten(conv2d_embedding2, start_dim=1) 
    
    x_maxpool = self.transformer_maxpool(x)
    x_maxpool_reduced = torch.squeeze(x_maxpool,1)
    x = x_maxpool_reduced.permute(2,0,1) 
    
    transformer_output = self.transformer_encoder(x)
    transformer_embedding = torch.mean(transformer_output, dim=0)
    
    complete_embedding = torch.cat([conv2d_embedding1, conv2d_embedding2,transformer_embedding], dim=1)  
    output_logits = self.fc1_linear(complete_embedding)  
    output_softmax = self.softmax_out(output_logits)
    
    # 临时只返回一个张量
    return output_logits

完成summary查看后，再改回原forward方法返回两个值即可。

方法2：改用torchinfo替代torchsummary #

torchinfo对多输出模型的支持更完善，安装后直接使用：

from torchinfo import summary

device = 'cuda'
model = parallel_all_you_want(len(emotion_labels)).to(device)
summary(model, input_size=(1,40,282))  # 原生支持多输出模型的结构查看

方法3：排查forward中是否存在None张量（可选） #

可以在forward中添加打印语句，确认所有中间张量都不为None：

def forward(self,x):
    print("Input shape:", x.shape)
    
    conv2d_embedding1 = self.conv2Dblock1(x)
    print("Conv1 output shape:", conv2d_embedding1.shape)
    conv2d_embedding1 = torch.flatten(conv2d_embedding1, start_dim=1)
    print("Conv1 flattened shape:", conv2d_embedding1.shape)
    
    conv2d_embedding2 = self.conv2Dblock2(x)
    print("Conv2 output shape:", conv2d_embedding2.shape)
    conv2d_embedding2 = torch.flatten(conv2d_embedding2, start_dim=1)
    print("Conv2 flattened shape:", conv2d_embedding2.shape)
    
    x_maxpool = self.transformer_maxpool(x)
    print("Maxpool shape:", x_maxpool.shape)
    
    x_maxpool_reduced = torch.squeeze(x_maxpool,1)
    print("Squeezed shape:", x_maxpool_reduced.shape)
    
    x = x_maxpool_reduced.permute(2,0,1)
    print("Transformer input shape:", x.shape)
    
    transformer_output = self.transformer_encoder(x)
    print("Transformer output shape:", transformer_output.shape)
    
    transformer_embedding = torch.mean(transformer_output, dim=0)
    print("Transformer embedding shape:", transformer_embedding.shape)
    
    complete_embedding = torch.cat([conv2d_embedding1, conv2d_embedding2,transformer_embedding], dim=1)
    print("Complete embedding shape:", complete_embedding.shape)
    
    output_logits = self.fc1_linear(complete_embedding)
    print("Logits shape:", output_logits.shape)
    
    output_softmax = self.softmax_out(output_logits)
    print("Softmax shape:", output_softmax.shape)
    
    return output_logits, output_softmax

通过打印结果可以快速确认是否有张量意外变为None。

额外提示 #

你的Transformer模块输入维度处理是正确的：输入经过MaxPool2d后维度变为(batch,1,40,70)，squeeze后得到(batch,40,70)，permute为(70,batch,40)，完全符合TransformerEncoder要求的(seq_len, batch_size, d_model)格式。

内容的提问来源于stack exchange，提问作者Arianna