7ba65487591fb4795d6384540bd4f580e820ff61,keras_gat/graph_attention_layer.py,GraphAttention,call,#GraphAttention#,78

Before Change

            // Compute feature combinations
            repeated = K.reshape(K.tile(linear_transf_X, [1, N]), (N * N, self.F_))  // (N^2 x F")
            tiled = K.tile(linear_transf_X, [N, 1])  // (N^2 x F")
            combinations = K.concatenate([repeated, tiled])  // (N^2 x 2F")
            combination_slices = K.reshape(combinations, (N, -1, 2 * self.F_))  // (N x N x 2F")

            // Attention head
            dense = K.squeeze(K.dot(combination_slices, attention_kernel), -1)  // a(Wh_i, Wh_j) in the paper (N x N x 1)

After Change

            // Compute feature combinations
            // Note: [[a_1], [a_2]]^T [[Wh_i], [Wh_2]] = [a_1]^T [Wh_i] + [a_2]^T [Wh_j]
            attn_for_self = K.dot(
                linear_transf_X, attention_kernel[0])  // (N x 1), [a_1]^T [Wh_i]
            attn_for_neighs = K.dot(
                linear_transf_X, attention_kernel[1])  // (N x 1), [a_2]^T [Wh_j]

            // Attention head a(Wh_i, Wh_j) = a^T [[Wh_i], [Wh_j]]
            dense = attn_for_self + K.transpose(attn_for_neighs)  // (N x N) via broadcasting

            // add nonlinearty
            dense = LeakyReLU(alpha=0.2)(dense)

Instances