参考来源：
    https://pytorch-geometric.readthedocs.io/en/latest/modules/utils.html#torch_geometric.utils.dropout_adj
    https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/utils/dropout.html

    1. dropout_adj(
    2.     edge_index, 
    3.     edge_attr=None, 
    4.     p=0.5, 
    5.     force_undirected=False, 
    6.     num_nodes=None, 
    7.     training=True
    8. )

    功能：
    使用来自伯努利分布的样本，以概率 p 从邻接矩阵 (edge_index, edge_attr) 中随机删除边。
    参数：

    • **edge_index** (LongTensor)：边的索引矩阵。
    • **edge_attr** (Tensor, optional)：边的权重或边的多维特征。 （默认：None）
    • **p** (float, optional)：Dropout 概率。 （默认值：0.5）
    • **force_undirected** (bool, optional)：如果设置为 True，将删除或保留无向边的两个边。 （默认：False）
    • **num_nodes** (int, optional)：节点数，即 edge_index 的 max_val + 1。 （默认：None）
    • **training** (bool, optional)：如果设置为 False，则此操作为空操作。 （默认：True）
    1. def dropout_adj(edge_index, edge_attr=None, p=0.5, force_undirected=False,
    2.                 num_nodes=None, training=True):
    3.     r"""Randomly drops edges from the adjacency matrix
    4.     :obj:`(edge_index, edge_attr)` with probability :obj:`p` using samples from
    5.     a Bernoulli distribution.
    7.     Args:
    8.         edge_index (LongTensor): The edge indices.
    9.         edge_attr (Tensor, optional): Edge weights or multi-dimensional
    10.             edge features. (default: :obj:`None`)
    11.         p (float, optional): Dropout probability. (default: :obj:`0.5`)
    12.         force_undirected (bool, optional): If set to :obj:`True`, will either
    13.             drop or keep both edges of an undirected edge.
    14.             (default: :obj:`False`)
    15.         num_nodes (int, optional): The number of nodes, *i.e.*
    16.             :obj:`max_val + 1` of :attr:`edge_index`. (default: :obj:`None`)
    17.         training (bool, optional): If set to :obj:`False`, this operation is a
    18.             no-op. (default: :obj:`True`)
    19.     """
    21.     if p < 0. or p > 1.:
    22.         raise ValueError('Dropout probability has to be between 0 and 1, '
    23.                          'but got {}'.format(p))
    25.     if not training or p == 0.0:
    26.         return edge_index, edge_attr
    28.     N = maybe_num_nodes(edge_index, num_nodes)
    29.     row, col = edge_index
    31.     if force_undirected:
    32.         row, col, edge_attr = filter_adj(row, col, edge_attr, row < col)
    34.     mask = edge_index.new_full((row.size(0), ), 1 - p, dtype=torch.float)
    35.     mask = torch.bernoulli(mask).to(torch.bool)
    37.     row, col, edge_attr = filter_adj(row, col, edge_attr, mask)
    39.     if force_undirected:
    40.         edge_index = torch.stack(
    41.             [torch.cat([row, col], dim=0),
    42.              torch.cat([col, row], dim=0)], dim=0)
    43.         if edge_attr is not None:
    44.             edge_attr = torch.cat([edge_attr, edge_attr], dim=0)
    45.         edge_index, edge_attr = coalesce(edge_index, edge_attr, N, N)
    46.     else:
    47.         edge_index = torch.stack([row, col], dim=0)
    49.     return edge_index, edge_attr