7557794793295e3b89130daf27bf284206eda8e8,onmt/modules/GlobalAttention.py,GlobalAttention,forward,#GlobalAttention#,108

Before Change

            aeq(batch, batch_)
            aeq(sourceL, sourceL_)

        if self.mask is not None:
            beam_, batch_, sourceL_ = self.mask.size()
            aeq(batch, batch_*beam_)
            aeq(sourceL, sourceL_)

        if coverage is not None:
            cover = coverage.view(-1).unsqueeze(1)
            context += self.linear_cover(cover).view_as(context)
            context = self.tanh(context)

        // compute attention scores, as in Luong et al.
        align = self.score(input, context)

        if self.mask is not None:
            mask_ = self.mask.view(batch, 1, sourceL)  // make it broardcastable
            align.data.masked_fill_(mask_, -float("inf"))

        // Softmax to normalize attention weights
        align_vectors = self.sm(align.view(batch*targetL, sourceL))
        align_vectors = align_vectors.view(batch, targetL, sourceL)

After Change

        align = self.score(input, context)

        if context_lengths is not None:
            mask = sequence_mask(context_lengths)
            mask = mask.unsqueeze(1)  // Make it broadcastable.
            align.data.masked_fill_(1 - mask, -float("inf"))

        // Softmax to normalize attention weights
        align_vectors = self.sm(align.view(batch*targetL, sourceL))
        align_vectors = align_vectors.view(batch, targetL, sourceL)

Instances