论文阅读04——《Attention-driven Graph Clustering Network》

class AGCN_H(nn.Module):
    def __init__(self, in_features, out_features, active=True):
        super(AGCN_H, self).__init__()
        self.active = active
        # 同一层输入特征维度一致，拼接后翻倍所以乘以2
        self.full_conn = Linear(in_features * 2, 2)
        self.gcn = GNNLayer(in_features, out_features)

    def forward(self, Z, H, adj):
        # 拼接
        N = torch.cat((Z, H), 1)
        # 全连接线性变换
        N = self.full_conn(N)
        # LeakyReLU激活
        N = F.leaky_relu(N, negative_slope=0.2)
        # softmax分类
        N = F.softmax(N, dim=1)
        # l2正则化
        M = F.normalize(N)
        # 切片并转置
        M_i_1 = M[:, 0].reshape(M.shape[0], 1)
        M_i_2 = M[:, 1].reshape(M.shape[0], 1)
        # 行向量1
        ones = torch.ones(1, Z.shape[1])
        ones = ones.cuda()
        # 权重系数与行向量1相乘得到权重矩阵
        w_1 = torch.mm(M_i_1, ones)
        w_2 = torch.mm(M_i_2, ones)
        # 加权融合
        Z = w_1 * Z + w_2 * H
        # 图卷积
        Z = self.gcn(Z, adj, active=self.active)
        return Z

'''
AGCN-H 的代码由于我一开始粗心，理解有偏差，特征融合仍然用了加权，而不是论文中的拼接，
这里注意一下!!!
'''
class AGCN_S(nn.Module):
    def __init__(self, z_dim, h1_dim, h2_dim, h3_dim, h4_dim, out_features):
        super(AGCN_S, self).__init__()
        self.transform_h1 = Linear(h1_dim, z_dim)
        self.transform_h2 = Linear(h2_dim, z_dim)
        self.transform_h3 = Linear(h3_dim, z_dim)
        self.transform_h4 = Linear(h4_dim, z_dim)
        self.full_conn = Linear(z_dim * 5, 5)
        self.gcn = GNNLayer(z_dim, out_features)

    def forward(self, z, h1, h2, h3, h4, adj):
        # 维度转换
        h1 = self.transform_h1(h1)
        h2 = self.transform_h2(h2)
        h3 = self.transform_h3(h3)
        h4 = self.transform_h4(h4)
        # 拼接
        N = torch.cat((z, h1, h2, h3, h4), 1)
        # 全连接线性变换
        N = self.full_conn(N)
        # LeakyReLU激活
        N = F.leaky_relu(N, negative_slope=0.2)
        # softmax分类
        N = F.softmax(N, dim=1)
        # l2正则化
        M = F.normalize(N)
        # 切片并转置
        M_i_1 = M[:, 0].reshape(M.shape[0], 1)
        M_i_2 = M[:, 1].reshape(M.shape[0], 1)
        M_i_3 = M[:, 2].reshape(M.shape[0], 1)
        M_i_4 = M[:, 3].reshape(M.shape[0], 1)
        M_i_5 = M[:, 4].reshape(M.shape[0], 1)
        # 行向量1
        ones = torch.ones(1, z.shape[1])
        ones = ones.cuda()
        # 权重矩阵
        w_1 = torch.mm(M_i_1, ones)
        w_2 = torch.mm(M_i_2, ones)
        w_3 = torch.mm(M_i_3, ones)
        w_4 = torch.mm(M_i_4, ones)
        w_5 = torch.mm(M_i_5, ones)
        # 融合
        Z = w_1 * z + w_2 * h1 + w_3 * h2 + w_4 * h3 + w_5 * h4
        # 图卷积
        Z = self.gcn(Z, adj, active=False)
        return Z