2020-11-28

最新推荐文章于 2023-05-15 23:00:51 发布

papa is a cat

最新推荐文章于 2023-05-15 23:00:51 发布

阅读量302

点赞数 3

文章标签：神经网络

本文链接：https://blog.csdn.net/dongming_happy/article/details/110274677

版权

飞桨图神经网络七日打卡营----图采样学习心得

图采样算法详解
1. GraphSAGE(SAmple & aggreGatE)
1. PinSAGE(SAmple & aggreGatE)
图采样算法之后—邻居聚会
源码实现

GraphSage采样函数实现

 import numpy as np

def traverse(item):
    """traverse
    """
    if isinstance(item, list) or isinstance(item, np.ndarray):
        for i in iter(item):
            for j in traverse(i):
                yield j
    else:
        yield item


def flat_node_and_edge(nodes):
    """flat_node_and_edge
    """
    nodes = list(set(traverse(nodes)))
    return nodes


def my_graphsage_sample(graph, batch_train_samples, samples):
    """
    输入：graph - 图结构 Graph
         batch_train_samples - 中心节点 list (batch_size,)
         samples - 采样时的最大邻节点数列表 list 
    输出：被采样节点下标的集合 
         对当前节点进行k阶采样后得到的子图 
    """
    
    start_nodes = batch_train_samples
    nodes = start_nodes
    edges = []
    for max_deg in samples:
        #################################
        # 请在这里补充每阶邻居采样的代码：此部分课堂实践内容已详细讲解，加油~
        # 提示：graph.sample_predecessor(该 API用于获取目标节点对应的源节点，具体用法到 pgl.Graph 结构中查看)
        pred_nodes = graph.sample_predecessor(nodes,max_deg)
    for dst_node, src_nodes in zip(nodes, pred_nodes):
        for node in src_nodes:
            edges.append((node, dst_node))
        #################################

        # 合并已采样节点并找出新的节点作为start_nodes
        last_nodes = nodes
        nodes = [nodes, pred_nodes]
        nodes = flat_node_and_edge(nodes)
        start_nodes = list(set(nodes) - set(last_nodes))
        if len(start_nodes) == 0:
            break

    subgraph = graph.subgraph(
         nodes=nodes,
         edges=edges,
         with_node_feat=False,
         with_edge_feat=False)
         
    return nodes, subgraph

MaxPool Aggregator实现

import paddle.fluid as fluid

def my_graphsage_maxpool(gw,
                      feature,
                      hidden_size,
                      act,
                      name,
                      inner_hidden_size=512):
    """
    输入：gw - GraphWrapper对象
         feature - 当前节点表示 (num_nodes, embed_dim)
         hidden_size - 新的节点表示维数 int
         act - 激活函数名 str
         name - 聚合函数名 str
         inner_hidden_size - 消息传递过程中邻居信息的维数 int
    输出：新的节点表示
    """
    
    ####################################
    # 请在这里实现MaxPool Aggregator

    def copy_send(src_feat, dst_feat, edge_feat):
         return src_feat["h"]
    def maxpool_recv(feat):
         return fluid.layers.sequence_pool(feat, pool_type="max")

    # 补充消息传递机制触发代码
    neigh_feature = fluid.layers.fc(feature, inner_hidden_size, act="relu")
    msg = gw.send(copy_send, nfeat_list=[("h", neigh_feature)])
    neigh_feature = gw.recv(msg, maxpool_recv)
    ####################################
    
    # 自身表示和邻居表示的结合
    self_feature = feature
    self_feature = fluid.layers.fc(self_feature,
                                   hidden_size,
                                   act=act,
                                   name=name + '_l')
    neigh_feature = fluid.layers.fc(neigh_feature,
                                    hidden_size,
                                    act=act,
                                    name=name + '_r')
    output = fluid.layers.concat([self_feature, neigh_feature], axis=1)
    output = fluid.layers.l2_normalize(output, axis=1)
    return output