1
$\begingroup$

The Decoder mask, also called "look-ahead mask", is applied in the Decoder side to prevent it from attending future tokens. Something like this:

[0, 1, 1, 1, 1]
[0, 0, 1, 1, 1]
[0, 0, 0, 1, 1]
[0, 0, 0, 0, 1]
[0, 0, 0, 0, 0]

But is this mask applied only in the first Decoder block? Or to all its blocks?

$\endgroup$
1
$\begingroup$

The masking should be applied to all Decoder blocks, otherwise in some blocks, past words can attend to future words, which would be cheating during training.

This is reflected in The Annotated Transformer as well. Notice that in the Decoder class, the forward function applies the same mask to each layer of the decoder:

class Decoder(nn.Module):
    "Generic N layer decoder with masking."
    def __init__(self, layer, N):
        super(Decoder, self).__init__()
        self.layers = clones(layer, N)
        self.norm = LayerNorm(layer.size)
        
    def forward(self, x, memory, src_mask, tgt_mask):
        for layer in self.layers:
            x = layer(x, memory, src_mask, tgt_mask)
        return self.norm(x)
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.