Source code for

from typing import List, Optional, Tuple, NamedTuple

import torch
from torch_sparse import SparseTensor

class Adj(NamedTuple):
    edge_index: torch.Tensor
    e_id: torch.Tensor
    size: Tuple[int, int]

    def to(self, *args, **kwargs):
        return Adj(*args, **kwargs),
         *args, **kwargs), self.size)

[docs]class NeighborSampler( r"""The neighbor sampler from the `"Inductive Representation Learning on Large Graphs" <>`_ paper, which allows for mini-batch training of GNNs on large-scale graphs where full-batch training is not feasible. Given a GNN with :math:`L` layers and a specific mini-batch of nodes :obj:`node_idx` for which we want to compute embeddings, this module iteratively samples neighbors and constructs bipartite graphs that simulate the actual computation flow of GNNs. More specifically, :obj:`sizes` denotes how much neighbors we want to sample for each node in each layer. This module then takes in these :obj:`sizes` and iteratively samples :obj:`sizes[l]` for each node involved in layer :obj:`l`. In the next layer, sampling is repeated for the union of nodes that were already encountered. The actual computation graphs are then returned in reverse-mode, meaning that we pass messages from a larger set of nodes to a smaller one, until we reach the nodes for which we originally wanted to compute embeddings. Hence, an item returned by :class:`NeighborSampler` holds the current :obj:`batch_size`, the IDs :obj:`n_id` of all nodes involved in the computation, and a list of bipartite graph objects via the tuple :obj:`(edge_index, e_id, size)`, where :obj:`edge_index` represents the bipartite edges between source and target nodes, :obj:`e_id` denotes the IDs of original edges in the full graph, and :obj:`size` holds the shape of the bipartite graph. For each bipartite graph, target nodes are also included at the beginning of the list of source nodes so that one can easily apply skip-connections or add self-loops. .. note:: For an example of using :obj:`NeighborSampler`, see `examples/ <>`_ or `examples/ <>`_. Args: edge_index (LongTensor): The edge indices of the full-graph. size ([int]): The number of neighbors to sample for each node in each layer. If set to :obj:`sizes[i] = -1`, all neighbors are included in layer :obj:`l`. node_idx (LongTensor, optional): The nodes that should be considered for creating mini-batches. If set to :obj:`None`, all nodes will be considered. flow (string, optional): The flow direction of message passing (:obj:`"source_to_target"` or :obj:`"target_to_source"`). (default: :obj:`"source_to_target"`) **kwargs (optional): Additional arguments of :class:``, such as :obj:`batch_size`, :obj:`shuffle`, :obj:`drop_last` or :obj:`num_workers`. """ def __init__(self, edge_index: torch.Tensor, sizes: List[int], node_idx: Optional[torch.Tensor] = None, num_nodes: Optional[int] = None, flow: str = "source_to_target", **kwargs): N = int(edge_index.max() + 1) if num_nodes is None else num_nodes edge_attr = torch.arange(edge_index.size(1)) adj = SparseTensor(row=edge_index[0], col=edge_index[1], value=edge_attr, sparse_sizes=(N, N), is_sorted=False) adj = adj.t() if flow == 'source_to_target' else adj self.adj ='cpu') if node_idx is None: node_idx = torch.arange(N) elif node_idx.dtype == torch.bool: node_idx = node_idx.nonzero().view(-1) self.sizes = sizes self.flow = flow assert self.flow in ['source_to_target', 'target_to_source'] super(NeighborSampler, self).__init__(node_idx.tolist(), collate_fn=self.sample, **kwargs) def sample(self, batch): if not isinstance(batch, torch.Tensor): batch = torch.tensor(batch) batch_size: int = len(batch) adjs: List[Adj] = [] n_id = batch for size in self.sizes: adj, n_id = self.adj.sample_adj(n_id, size, replace=False) if self.flow == 'source_to_target': adj = adj.t() row, col, e_id = adj.coo() size = adj.sparse_sizes() edge_index = torch.stack([row, col], dim=0) adjs.append(Adj(edge_index, e_id, size)) if len(adjs) > 1: return batch_size, n_id, adjs[::-1] else: return batch_size, n_id, adjs[0] def __repr__(self): return '{}(sizes={})'.format(self.__class__.__name__, self.sizes)