代码库：DGL_异构图神经网络的基本框架

import torch as th
import dgl

Using backend: pytorch

创建HGNN

graph_data = {
    ('a', 'B', 'b'): (th.tensor([0, 0]), th.tensor([1, 2])),
    ('b', 'C', 'b'): (th.tensor([1]), th.tensor([2]))
}
heter_g = dgl.heterograph(graph_data)

'''
(src_type, edge_type, dest_type): (src_idx_list, dst_idx_list)

It's should be: len(src_idx_list) == len(dst_idx_list)


heter_g:
    a(0) -> B -> b(0)
    a(0) -> B -> b(1)
    b(0) -> C -> b(1)
'''

"\n(src_type, edge_type, dest_type): (src_idx_list, dst_idx_list)\n\nIt's should be: len(src_idx_list) == len(dst_idx_list)\n\n\nheter_g:\n    a(0) -> B -> b(0)\n    a(0) -> B -> b(1)\n    b(0) -> C -> b(1)\n"

heter_g

Graph(num_nodes={'a': 1, 'b': 3},
      num_edges={('a', 'B', 'b'): 2, ('b', 'C', 'b'): 1},
      metagraph=[('a', 'b', 'B'), ('b', 'b', 'C')])

查看HGNN属性与编号

heter_g.srctypes

['a', 'b']

heter_g.dsttypes

['a', 'b']

heter_g.ntypes

['a', 'b']

heter_g.etypes

['B', 'C']

heter_g.num_nodes()

3

heter_g.num_nodes('a')

1

heter_g.num_edges()

3

heter_g.num_edges('B')

2

heter_g.nodes('b')

tensor([0, 1])

heter_g.edges(etype='B')

'''
form:
    'uv': return src nodes id and dst nodes id
    'eid': return edge id
    'all': return 'uv' and 'eid'
etype: edge type
'''

(tensor([0, 0]), tensor([0, 1]))

heter_g.edges(form='uv',  etype='B')

(tensor([0, 0]), tensor([1, 2]))

节点（边）特征

# 节点类别
heter_g.nodes['a']

NodeSpace(data={})

heter_g.nodes['a'].data['feat'] = th.randn((heter_g.num_nodes('a'), 3))
heter_g.nodes['a'].data['feat']

tensor([[ 1.1605,  0.6090, -0.2546]])

heter_g.ndata['feat']

{'a': tensor([[ 1.1605,  0.6090, -0.2546]])}

# 边类别
heter_g.edges['B']

EdgeSpace(data={})

metapath

heter_g.canonical_etypes

[('a', 'B', 'b'), ('b', 'C', 'b')]

heter_g.metagraph().edges()

OutMultiEdgeDataView([('a', 'b'), ('b', 'b')])

heter_g.metagraph().nodes()

NodeView(('a', 'b'))

根据metapath取子图

sub_g = dgl.edge_type_subgraph(heter_g, [('a', 'B', 'b')])
sub_g

Graph(num_nodes={'a': 1, 'b': 2},
      num_edges={('a', 'B', 'b'): 2},
      metagraph=[('a', 'b', 'B')])

sub_g.nodes['a'].data['feat']

tensor([[ 1.1605,  0.6090, -0.2546]])

HGNN 2 GNN

heter_g

Graph(num_nodes={'a': 1, 'b': 2},
      num_edges={('a', 'B', 'b'): 2, ('b', 'C', 'b'): 1},
      metagraph=[('a', 'b', 'B'), ('b', 'b', 'C')])

g = dgl.to_homogeneous(heter_g)
g

Graph(num_nodes=3, num_edges=3,
      ndata_schemes={'_ID': Scheme(shape=(), dtype=torch.int64), '_TYPE': Scheme(shape=(), dtype=torch.int64)}
      edata_schemes={'_ID': Scheme(shape=(), dtype=torch.int64), '_TYPE': Scheme(shape=(), dtype=torch.int64)})

# '_' 表示当前数据在原数据的表示形式
# 默认情况下不会保留节点和边上的特征
g.ndata

{'_ID': tensor([0, 0, 1]), '_TYPE': tensor([0, 1, 1])}

heter_g.ntypes

['a', 'b']

# 复制粘贴特征矩阵
# 转换成同构图时 须假定每个节点都具有相同维度大小的特征矩阵 不能没有特征 否则无法合并或保留
heter_g.nodes['b'].data['feat'] = th.randn(heter_g.num_nodes('b'), 3)
heter_g.ndata['feat']

{'a': tensor([[ 1.1605,  0.6090, -0.2546]]),
 'b': tensor([[ 0.6973,  0.0062, -1.3152],
         [-2.1732, -1.2436,  0.0826]])}

g = dgl.to_homogeneous(heter_g, ndata=['feat'], edata=None)
g

Graph(num_nodes=3, num_edges=3,
      ndata_schemes={'feat': Scheme(shape=(3,), dtype=torch.float32), '_ID': Scheme(shape=(), dtype=torch.int64), '_TYPE': Scheme(shape=(), dtype=torch.int64)}
      edata_schemes={'_ID': Scheme(shape=(), dtype=torch.int64), '_TYPE': Scheme(shape=(), dtype=torch.int64)})

g.ndata['feat']

tensor([[ 1.1605,  0.6090, -0.2546],
        [ 0.6973,  0.0062, -1.3152],
        [-2.1732, -1.2436,  0.0826]])

消息传递

• multi_update_all(etype_dict, cross_reducer, apply_node_func=None):

  1. 对每一类metapath 中的所有 metapath instances 进行update_all，得到每一类metapath latent vector。这个过程相当于GNN中的广义消息传递的(1) and (2)，即尚未考虑目标节点自身的特征
  2. 对所有metapath (latent vector) 进行 聚合

• etype_dict: {(src_type, edge_type, dst_type): (message_func, reduce_func, apply_node_func=None}

  • etype_dict[(src_type, edge_type, dst_type)] 表明对此类metapath instances进行update_all的参数

• cross_reducer: 对所有metapath进行聚合，取值包括 ‘mean’, ‘min’, ‘max’, ‘sum’, ‘stack’

import dgl.function as fn

# metapath types
heter_g.canonical_etypes

[('a', 'B', 'b'), ('b', 'C', 'b')]

for metapath in heter_g.canonical_etypes:
    print(metapath)

('a', 'B', 'b')
('b', 'C', 'b')

metapath_update = {}
cross_reducer = 'mean'

for metapath in heter_g.canonical_etypes:
    metapath_update[metapath] = (fn.copy_u('feat', 'm'), fn.sum('m', 'feat_'))
    
heter_g.multi_update_all(metapath_update, cross_reducer)

# a类型节点没有消息传递后的新特征矩阵feat_ ，因为a类型节点没有出度，无法被更新
heter_g.nodes['a'].data.keys()

dict_keys(['feat'])

heter_g.nodes['b'].data.keys()

dict_keys(['feat', 'feat_'])

# nodes features
heter_g.ndata['feat']

{'a': tensor([[ 1.1605,  0.6090, -0.2546]]),
 'b': tensor([[ 0.6973,  0.0062, -1.3152],
         [-2.1732, -1.2436,  0.0826]])}

# ([ 1.1605,  0.6090, -0.2546] + [ 0.6973,  0.0062, -1.3152]) / 2 = [ 0.9289,  0.3076, -0.7849]
# ([ 1.1605,  0.6090, -0.2546] + 0) / 2 = [ 0.5802,  0.3045, -0.1273]
# mean 根据全图的metapath种类数进行平均

heter_g.nodes['b'].data['feat_']

tensor([[ 0.5802,  0.3045, -0.1273],
        [ 0.9289,  0.3076, -0.7849]])

metapath_update = {}
cross_reducer = 'stack'

for metapath in heter_g.canonical_etypes:
    metapath_update[metapath] = (fn.copy_u('feat', 'm'), fn.sum('m', 'feat1'))
    
heter_g.multi_update_all(metapath_update, cross_reducer)

heter_g.ndata['feat']

{'a': tensor([[ 1.1605,  0.6090, -0.2546]]),
 'b': tensor([[ 0.6973,  0.0062, -1.3152],
         [-2.1732, -1.2436,  0.0826]])}

# 由于b.0 节点 没有种类为C的边的入度 故stack的结果会有一个为全0 的向量
heter_g.ndata['feat1']

{'b': tensor([[[ 1.1605,  0.6090, -0.2546],
          [ 0.0000,  0.0000,  0.0000]],
 
         [[ 1.1605,  0.6090, -0.2546],
          [ 0.6973,  0.0062, -1.3152]]])}

heter_g.is_block?