来自 Trax 的 AttentionQKV

Question

Trax实现的AttentionQKV层如下：AttentionQKV

def AttentionQKV(d_feature, n_heads=1, dropout=0.0, mode='train'):
  """Returns a layer that maps (q, k, v, mask) to (activations, mask).
  See `Attention` above for further context/details.
  Args:
    d_feature: Depth/dimensionality of feature embedding.
    n_heads: Number of attention heads.
    dropout: Probababilistic rate for internal dropout applied to attention
        activations (based on query-key pairs) before dotting them with values.
    mode: One of `'train'`, `'eval'`, or `'predict'`.
  """
  return cb.Serial(
      cb.Parallel(
          core.Dense(d_feature),
          core.Dense(d_feature),
          core.Dense(d_feature),
      ),
      PureAttention(  # pylint: disable=no-value-for-parameter
          n_heads=n_heads, dropout=dropout, mode=mode),
      core.Dense(d_feature),
  )

特别是三个平行的密集层的目的是什么？该层的输入是 q, k, v, mask。为什么q、k、v要经过一个dense layer？

Answer 1

这段代码 sn-p 是 Attention is all you need 论文第 5 页顶部的方程的实现，该论文在 2017 年介绍了 Transformer 模型。计算如图 2 所示：

隐藏状态被投影到 h 注意力头中，这些注意力头并行地进行缩放的点积注意力。投影可以解释为提取与头部相关的信息。然后每个头部根据不同的（学习的）标准进行概率检索。