为什么Self-Attention要通过线性变换计算Q K V，背后的原理或直观解释是什么？ - 知乎回答题主问题题主的问题: 在attention中都经过一个映射，那么建模的相似度是否就没有意义了？个人感觉这…https://www.zhihu.com/question/592626839/answer/2965200007Cross-Attention in Transformer ArchitectureMerge two embedding sequences regardless of modality, e.g., image with text in Stable Diffusion U-Net.https://vaclavkosar.com/ml/cross-attention-in-transformer-architecture一言以蔽之，就是cross-attention是在sequence之间不同位置的输入做attention，self-attention是在sequence内部做attention。self和cross attention的区别仅在q和kv的来源上，self-attention Q（uery）K（ey）V（alue）均来源与一个sequence，而cross-attention中Q来源于另一个sequence，而且多为跨模态的sequence。

cross-attention混入了两种不同的序列。

class CrossAttention(nn.Module):
    r"""
    A cross attention layer.

    Parameters:
        query_dim (`int`): The number of channels in the query.
        cross_attention_dim (`int`, *optional*):
            The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
        heads (`int`,  *optional*, defaults to 8): The number of heads to use for multi-head attention.
        dim_head (`int`,  *optional*, defaults to 64): The number of channels in each head.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        bias (`bool`, *optional*, defaults to False):
            Set to `True` for the query, key, and value linear layers to contain a bias parameter.
    """
    query = attn.to_q(hidden_states)
    query = attn.head_to_batch_dim(query)

    encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
    key = attn.to_k(encoder_hidden_states)
    value = attn.to_v(encoder_hidden_states)
    key = attn.head_to_batch_dim(key)
    value = attn.head_to_batch_dim(value)

    attention_probs = attn.get_attention_scores(query, key, attention_mask)
    hidden_states = torch.bmm(attention_probs, value)

应用：

在transformer中有描述，但是是同模态的，还不叫cross-attention

stable diffusion中：

 perceiver io：