91f834a5329d7f4c440811441095c9d574764dfe,thinc/neural/_classes/multiheaded_attention.py,MultiHeadedAttention,begin_update,#MultiHeadedAttention#,87

Before Change

            masks = [None for _ in Qs]
        assert len(Qs) == len(Ks) == len(Vs)
        Ys, get_dQs_dKs_dVs = self._attend_seqs(Qs, Ks, Vs, masks)
        lengths = self.ops.asarray([len(y) for y in Ys], dtype="i")
        assert sum(lengths) == sum([len(q) for q in Qs])
        Y = self.ops.flatten(Ys)
        
        def finish_update(dY, sgd=None):
            dYs = self.ops.unflatten(dY, lengths)
            return get_dQs_dKs_dVs(dYs, sgd=sgd)
        
        assert sum(lengths) == Y.shape[0], (sum(lengths, Y.shape[0]))
        assert sum([len(q) for q in Qs]) == sum(lengths)
        return (Y, lengths), finish_update

    def _attend_seqs(self, Qs, Ks, Vs, masks):
        outputs = []
        backprops = []

After Change

        if masks is None:
            masks = [None for _ in Qs]
        assert len(Qs) == len(Ks) == len(Vs)
        return self._attend_seqs(Qs, Ks, Vs, masks)
        
    def _attend_seqs(self, Qs, Ks, Vs, masks):
        outputs = []
        backprops = []

Instances