72315ae500bd36c33835760005599fad7b6cb70b,onmt/modules/MultiHeadedAttn.py,MultiHeadedAttention,forward,#MultiHeadedAttention#Any#Any#Any#Any#,70

Before Change

        if mask is not None:
            scaled = scaled.view(b, self.head_count, l, dim_per_head)
            mask = mask.unsqueeze(1).expand_as(scaled)
            scaled = scaled.masked_fill(Variable(mask), -1e18) \
                           .view(bh, l, dim_per_head)

        // Return one attn

After Change

        
        if mask is not None:
            mask = mask.unsqueeze(1).expand_as(scaled_scores)
            scaled_scores = scaled_scores.masked_fill(Variable(mask), -1e18) 

        // 3) Apply attention dropout and compute context vectors.
        attn = self.sm(scaled_scores)
        drop_attn = self.dropout(attn)
        context = unshape(torch.matmul(drop_attn, value_up))
        
        output = self.final_linear(context)
        // CHECK
        batch_, q_len_, d_ = output.size()
        aeq(q_len, q_len_)
        aeq(batch, batch_)

Instances