Adding cross-attention to the decoder layer
To integrate the encoder and decoder stacks you've defined previously into an encoder-decoder transformer, you need to create a cross-attention mechanism to act as a bridge between the two.
The MultiHeadAttention class you defined previously is still available.
This exercise is part of the course
Transformer Models with PyTorch
Exercise instructions
- Define a cross-attention mechanism (using
MultiHeadAttention) and a third layer normalization (usingnn.LayerNorm) in the__init__method. - Complete the forward pass to add cross-attention to the decoder layer.
Hands-on interactive exercise
Have a go at this exercise by completing this sample code.
class DecoderLayer(nn.Module):
def __init__(self, d_model, num_heads, d_ff, dropout):
super().__init__()
self.self_attn = MultiHeadAttention(d_model, num_heads)
# Define cross-attention and a third layer normalization
self.cross_attn = ____
self.ff_sublayer = FeedForwardSubLayer(d_model, d_ff)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = ____
self.dropout = nn.Dropout(dropout)
def forward(self, x, y, tgt_mask, cross_mask):
self_attn_output = self.self_attn(x, x, x, tgt_mask)
x = self.norm1(x + self.dropout(self_attn_output))
# Complete the forward pass
cross_attn_output = self.____(____)
x = self.norm2(x + self.dropout(____))
ff_output = self.ff_sublayer(x)
x = self.norm3(x + self.dropout(ff_output))
return x