报错：UnboundLocalError: local variable 'encoder_extended_attention_mask' referenced before assignment

Question

这是我跑模型的代码

 def forward(self, image_feature, merge_feature):

        image_atts = torch.ones(image_feature.size()[:-1], dtype=torch.long)
        query_tokens = self.query_tokens.expand(merge_feature.shape[0], -1, -1)
        
        print("image:", image_feature.size())
        print("query:", query_tokens.size())
        print("merge:", merge_feature.size())
        # 检查和调整 image_feature 的 shape 以适应矩阵乘法
        # 我们需要确保 image_feature 的最后一个维度与 query_tokens 的最后一个维度匹配。
        image_feature = image_feature.unsqueeze(1).expand(-1, query_tokens.size(1), -1)
        image_feature = self.image_proj(image_feature)
        if image_feature.size(2) != query_tokens.size(2):
            raise ValueError("The last dimension of image_feature must match the last dimension of query_tokens.")




        print("image_upgrate:", image_feature.size())
        # print("image_att__upgrate:", image_atts.size())


        query_output = self.Qformer.bert(
            query_embeds=query_tokens,
            encoder_hidden_states=image_feature,
            encoder_attention_mask=image_atts,
            use_cache=True,
            return_dict=True,
        )

出现问题的地方在

def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
        """
        Invert an attention mask (e.g., switches 0. and 1.).

        Args:
            encoder_attention_mask (`torch.Tensor`): An attention mask.

        Returns:
            `torch.Tensor`: The inverted attention mask.
        """
        if encoder_attention_mask.dim() == 3:
            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
        if encoder_attention_mask.dim() == 2:
            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow
        # /transformer/transformer_layers.py#L270
        # encoder_extended_attention_mask = (encoder_extended_attention_mask ==
        # encoder_extended_attention_mask.transpose(-1, -2))
        encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * torch.finfo(self.dtype).min

        return encoder_extended_attention_mask

报错：UnboundLocalError: local variable 'encoder_extended_attention_mask' referenced before assignment

Answer 1

引自免费微信小程序：皆我百晓生

从你提供的代码片段来看，错误信息“UnboundLocalError: local variable 'encoder_extended_attention_mask' referenced before assignment”意味着你在尝试引用一个尚未被赋值的局部变量encoder_extended_attention_mask。这个问题通常发生在条件语句中，即根据不同的条件分别给变量赋值，但如果没有满足任何条件，那么该变量将保持未定义状态。

在invert_attention_mask方法中，你有两段代码用于根据encoder_attention_mask的维度来扩展它，但是这两段代码是互斥的，只有当encoder_attention_mask的维度为2或3时才会执行其中一个。如果encoder_attention_mask的维度既不是2也不是3，那么encoder_extended_attention_mask将不会被初始化，从而导致之后的代码引用未定义的变量。

为了解决这个问题，你需要确保无论输入的维度如何，encoder_extended_attention_mask都会被正确初始化。你可以在方法的开始处添加一个检查，以处理所有可能的维度情况，或者至少确保即使在非预期的维度下也能安全地初始化该变量。

下面是一个修改后的invert_attention_mask方法，它首先检查encoder_attention_mask的维度，并在不匹配的情况下抛出异常，确保了encoder_extended_attention_mask总是在后续操作前被正确初始化：

def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
    """
    Invert an attention mask (e.g., switches 0. and 1.).

    Args:
        encoder_attention_mask (`torch.Tensor`): An attention mask.

    Returns:
        `torch.Tensor`: The inverted attention mask.
    """
    # Check dimensions
    if encoder_attention_mask.dim() not in [2, 3]:
        raise ValueError("The attention mask should have 2 or 3 dimensions.")

    if encoder_attention_mask.dim() == 3:
        encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
    elif encoder_attention_mask.dim() == 2:
        encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]

    # Ensure the mask is of the correct data type for fp16 compatibility
    encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=self.dtype)

    # Invert the mask and apply the minimum value of the data type to non-masked elements
    encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * torch.finfo(self.dtype).min

    return encoder_extended_attention_mask

通过这种方式，我们确保了在调用invert_attention_mask方法时，不论输入的encoder_attention_mask具有何种维度，encoder_extended_attention_mask都将被正确初始化并返回。这样可以避免由于变量未定义而引发的运行时错误。