043d176e63960102db551b2b5739721c0467a43e,thinc/neural/_classes/multiheaded_attention.py,MultiHeadedAttention,begin_update,#MultiHeadedAttention#Any#Any#,232

Before Change

        assert Xattns[0].shape == (lengths[0], X.shape[1]), (Xattns[0].shape, X.shape[1])
        Xattn = self.ops.flatten(Xattns)
        assert Xattn.shape == X.shape
        Y, get_dXattn = self.get_output.begin_update(Xattn, drop=drop)

        def backprop_self_attn(dY, sgd=None):
            dXattn = get_dXattn(dY, sgd=sgd)
            dXattns = self.ops.unflatten(dXattn, lengths)

After Change

    def begin_update(self, inputs, drop=0.0):
        (Qs, Ks, Vs), get_d_inputs = self.handle_inputs(inputs, drop=drop)
        Y, get_dQs_dKs_dVs = self.attend((Qs, Ks, Vs), drop=drop)
        outputs, get_dY = self.handle_outputs(Y, inputs)

        def backprop_self_attn(d_outputs, sgd=None):

Instances