GraphSage模型cora数据集

刺骨的言语ヽ痛彻心扉 2022-09-11 09:23 13阅读 0赞

创建`GraphSage`模型：

input_size=1433
    out_size=128
    num_layers=2
    agg_func='MEAN'
    raw_features：原始特征，维度2708 * 1433
    adj_lists：所有边的连接，格式是：562: {60, 252, 370, 440, 671, 1117, 1183, 1401, 1889, 2014, 2018},
    sage_layer1：SageLayer
    sage_layer2：SageLayer(layer_size=128,out_size=128)

第一层`sage_layer1`:

input_size=1433
    out_size=128
    weight：权重参数，并初始化，维度128 * (1433*2)=128 * 2866

第二层`sage_layer2`：

input_size=128
    out_size=128
    weight：权重参数，并初始化，维度128 * (128*2)=128 * 256

创建`Classification`模型：

self.layer=全连接层 
    weight：权重参数，并初始化，维度128 * 7

进入`GraphSage`中的前向传播`forward`：

第一批20个节点数据
    nodes_batch={1, 16, 131, 245, 439, 501, 527, 658, 699, 716, 944, 963, 1003, 1081, 1439, 1582, 1660, 1681, 2007, 2577}

得到`nodes_batch_layers`，内容如下：  
![在这里插入图片描述][e1c4a268ff054c19b94ced7115ce262c.png]

# [1, 16, 131, 245, 439, 501, 527, 658, 699, 716, 944, 963, 1003, 1081, 1439, 1582, 1660, 1681, 2007, 2577}]
    nodes_batch_layers[2][0]：原始batch的20个数据节点list
    
    # [1, 131, 388, 2180, 2182, 1545, 521, 1020, 527, 16, 1681, 1682, 658, 2577, 791, 663, 157, 1439, 672, 1697, 803, 420, 163, 294, 2342, 552, 1577, 427, 172, 2606, 174, 1968, 1582, 944, 2483, 53, 566, 439, 952, 1081, 1082, 699, 1083, 1469, 1086, 2367, 192, 1337, 1335, 963, 68, 588, 76, 716, 2508, 466, 1364, 2007, 344, 97, 2146, 611, 1636, 2017, 2662, 1897, 747, 1003, 1004, 2028, 1519, 496, 632, 1259, 1779, 1780, 245, 501, 503, 1656, 1908, 2677, 1909, 1660]
    nodes_batch_layers[1][0]：与20个节点相连的84个邻居节点list，包含上一层原始的20个节点
    # [{527, 747, 1656, 1968, 2606, 2662},
    #  {1681, 1682},
    #  {97, 192, 388, 420, 439},
    #  {68, 1779, 1780, 2007},
    #  {53, 803, 1439},
    #  {344, 963, 1519},
    #  {294, 566, 588, 699, 747},
    #  {76, 131},
    #  {672, 791, 1003, 1004, 1364, 1469, 1897, 2028, 2146, 2182, 2483},
    #  {1, 344},
    #  {163, 427, 658, 1577},
    #  {952, 1083, 1660},
    #  {16, 466, 552, 566, 1697},
    #  {157, 611, 716},
    #  {172, 245, 1636},
    #  {1020, 2180, 2342, 2367, 2508, 2577},
    #  {174, 496, 501, 503, 663},
    #  {632, 1086, 1545, 1582, 1908, 1909, 2017, 2677},
    #  {521, 1081, 1082, 1259},
    #  {944, 1335, 1337}]
    nodes_batch_layers[1][1]：与20个节点相连邻居节点
    {1: 0,
    # 131: 1,
    # 388: 2,
    # 2180: 3,
    # 2182: 4,
    # 1545: 5,
    # 521: 6,
    # 1020: 7,
    # 527: 8,
    # 16: 9,
    # 1681: 10,
    # 1682: 11,
    # 658: 12,
    # 2577: 13,
    # 791: 14,
    # 663: 15,
    # 157: 16,
    # 1439: 17,
    # 672: 18,
    # 1697: 19,
    # 803: 20,
    # 420: 21,
    # 163: 22,
    # 294: 23,
    # 2342: 24,
    # 552: 25,
    # 1577: 26,
    # 427: 27,
    # 172: 28,
    # 2606: 29,
    # 174: 30,
    # 1968: 31,
    # 1582: 32,
    # 944: 33,
    # 2483: 34,
    # 53: 35,
    # 566: 36,
    # 439: 37,
    # 952: 38,
    # 1081: 39,
    # 1082: 40,
    # 699: 41,
    # 1083: 42,
    # 1469: 43,
    # 1086: 44,
    # 2367: 45,
    # 192: 46,
    # 1337: 47,
    # 1335: 48,
    # 963: 49,
    # 68: 50,
    # 588: 51,
    # 76: 52,
    # 716: 53,
    # 2508: 54,
    # 466: 55,
    # 1364: 56,
    # 2007: 57,
    # 344: 58,
    # 97: 59,
    # 2146: 60,
    # 611: 61,
    # 1636: 62,
    # 2017: 63,
    # 2662: 64,
    # 1897: 65,
    # 747: 66,
    # 1003: 67,
    # 1004: 68,
    # 2028: 69,
    # 1519: 70,
    # 496: 71,
    # 632: 72,
    # 1259: 73,
    # 1779: 74,
    # 1780: 75,
    # 245: 76,
    # 501: 77,
    # 503: 78,
    # 1656: 79,
    # 1908: 80,
    # 2677: 81,
    # 1909: 82,
    # 1660: 83}
    nodes_batch_layers[1][2]：84个节点进行dict编号
    
    nodes_batch_layers[0][0]：与84个节点相连的304个邻居节点list，包含84个自身节点
    nodes_batch_layers[0][1]：84个节点的邻居节点list
    nodes_batch_layers[0][2]：304个节点的dict自编编号

接下来进行aggregation操作：

index=1
    nb = 自己和邻居节点list 共84个
    pre_neighs = 这层节点的上层邻居的所有信息，包括：聚合自己和邻居节点，点的dict，涉及到的所有节点
    
    aggregation：
    参数：nodes，当前节点list 84个
    参数：pre_hidden_embs，数据特征，维度：2708 * 1433
    参数：pre_neighs，上层邻居的所有信息
    embed_matrix：当前304个节点的特征，304*1433
    mask：构建 84*304 的邻接矩阵，并且进行归一化mask=mask/mask.sum(dim=1)，
    aggregate_feats：聚合邻居特征，mask * embed_matrix ，生成维度84*1433
    总结：当前节点84个，聚合了周围304个邻居节点的特征，对特征进行归一化，生成特征维度84*1433
    
    sageLayer:
    参数：self_feats=原始特征  维度84*1433
    参数：aggregate_feats=聚合了304个邻居的特征 维度84*1433
    combined：将这两个特征进行拼接，生成维度84*2866
    F.relu(combined * W)，进行激活函数生成维度84*128的特征
    
    返回特征 84*128
    
    
    
    index=2
    nb=最初的20个节点
    pre_neighs=这层节点的上层邻居的所有信息，包括：聚合自己和邻居节点，点的dict，涉及到的所有节点
    
    aggregation：
    参数：nodes，当前节点list 20个
    参数：pre_hidden_embs，数据特征，维度：84*128
    参数：pre_neighs，上层邻居的所有信息
    embed_matrix：=pre_hidden_embs 利用上一层的节点特征，即84个节点的特征，84*128
    mask：构建 20*84 的邻接矩阵，并且进行归一化mask=mask/mask.sum(dim=1)，
    aggregate_feats：聚合邻居特征，mask * embed_matrix ，生成维度20*128
    总结：当前节点20个，聚合了周围84个邻居节点的特征，（这84个邻居节点特征是上一层计算得到的，因此维度是84*128，而不是84*1433），对特征进行归一化，最终生成特征维度20*128
    
    sageLayer:
    参数：self_feats=原始特征  维度20*128
    参数：aggregate_feats=聚合了20个邻居的特征 维度20*128
    combined：将这两个特征进行拼接，生成维度20*256
    F.relu(combined * W)，进行激活函数生成维度20*128的特征
    
    最终这20个节点返回特征 20*128

接下来就要通过`Classification`模型进行计算，得到维度`20 * 7`

logists=
    tensor([[-1.9839, -1.9866, -2.0804, -1.9014, -1.8841, -1.7800, -2.0367],
            [-1.9142, -1.9110, -2.0967, -1.8599, -2.0505, -1.8341, -1.9832],
            [-1.9683, -1.9229, -1.9720, -1.9572, -1.9711, -1.8678, -1.9667],
            [-1.9209, -2.0136, -2.0654, -1.8666, -1.9862, -1.8006, -1.9937],
            [-2.0788, -2.0250, -1.9542, -1.8534, -1.8923, -1.8745, -1.9634],
            [-2.0361, -1.8748, -2.0676, -1.8287, -1.9016, -1.8289, -2.1284],
            [-2.0782, -1.8997, -1.9607, -1.9720, -1.9491, -1.9568, -1.8227],
            [-2.0626, -1.8991, -2.0198, -1.8183, -1.9802, -1.8021, -2.0779],
            [-1.9372, -1.8791, -1.9737, -1.9121, -1.9777, -1.8976, -2.0547],
            [-2.0668, -1.9538, -1.9821, -1.8051, -2.0310, -1.8033, -2.0139],
            [-2.0350, -1.9644, -1.9913, -1.8626, -1.9652, -1.8815, -1.9324],
            [-2.0077, -1.8961, -2.0062, -2.0043, -1.8142, -1.8510, -2.0695],
            [-1.9880, -1.9198, -1.9179, -1.8831, -1.9339, -2.0266, -1.9590],
            [-2.0564, -1.9405, -1.9988, -1.8407, -1.8986, -1.9623, -1.9384],
            [-2.0603, -1.9230, -1.9647, -1.8786, -1.9230, -1.9342, -1.9470],
            [-2.0483, -1.9780, -2.0290, -1.9329, -1.8557, -1.8542, -1.9407],
            [-1.9123, -1.9745, -1.9420, -1.9195, -1.9199, -1.9526, -2.0038],
            [-2.0144, -1.9391, -1.9132, -1.8205, -1.9620, -1.9279, -2.0622],
            [-2.1094, -1.7795, -2.0029, -1.8137, -2.0519, -1.8241, -2.1027],
            [-2.0052, -1.9684, -2.0337, -1.8064, -1.9325, -1.8070, -2.1065]],
     loss_sup = -torch.sum(logists[range(logists.size(0)), labels_batch], 0) 根据label获取对应的loss值求和
    = tensor([-1.9839, -1.8599, -1.9683, -2.0654, -1.8745, -1.8748, -1.9607, -1.8183,
            -1.9777, -1.9538, -1.8815, -2.0043, -1.9179, -2.0564, -1.9470, -1.9329,
            -1.9123, -1.8205, -1.7795, -1.8064], grad_fn=<IndexBackward>)
      loss_sup /= len(nodes_batch)
    =1.9198

这样第一批数据的20个节点就计算完成了：

array([ 527, 1681,  439, 2007, 1439,  963,  699,  131, 1003,    1,  658,
           1660,   16,  716,  245, 2577,  501, 1582, 1081,  944])
    Step [1/68], Loss: 1.9198, Dealed Nodes [20/1355]

[e1c4a268ff054c19b94ced7115ce262c.png]: /images/20220828/30c41294e40d41778a55e99fb1eca4ad.png