Source code for torch_geometric.nn.models.re_net

import math

import torch
from torch.nn import Parameter, GRU, Linear
import torch.nn.functional as F
from torch_scatter import scatter_mean


[docs]class RENet(torch.nn.Module): r"""The Recurrent Event Network model from the `"Recurrent Event Network for Reasoning over Temporal Knowledge Graphs" <https://arxiv.org/abs/1904.05530>`_ paper .. math:: f_{\mathbf{\Theta}}(\mathbf{e}_s, \mathbf{e}_r, \mathbf{h}^{(t-1)}(s, r)) based on a RNN encoder .. math:: \mathbf{h}^{(t)}(s, r) = \textrm{RNN}(\mathbf{e}_s, \mathbf{e}_r, g(\mathcal{O}^{(t)}_r(s)), \mathbf{h}^{(t-1)}(s, r)) where :math:`\mathbf{e}_s` and :math:`\mathbf{e}_r` denote entity and relation embeddings, and :math:`\mathcal{O}^{(t)}_r(s)` represents the set of objects interacted with subject :math:`s` under relation :math:`r` at timestamp :math:`t`. This model implements :math:`g` as the **Mean Aggregator** and :math:`f_{\mathbf{\Theta}}` as a linear projection. Args: num_nodes (int): The number of nodes in the knowledge graph. num_rels (int): The number of relations in the knowledge graph. hidden_channels (int): Hidden size of node and relation embeddings. seq_len (int): The sequence length of past events. num_layers (int, optional): The number of recurrent layers. (default: :obj:`1`) dropout (float): If non-zero, introduces a dropout layer before the final prediction. (default: :obj:`0.`) bias (bool, optional): If set to :obj:`False`, all layers will not learn an additive bias. (default: :obj:`True`) """ def __init__(self, num_nodes, num_rels, hidden_channels, seq_len, num_layers=1, dropout=0., bias=True): super(RENet, self).__init__() self.num_nodes = num_nodes self.hidden_channels = hidden_channels self.num_rels = num_rels self.seq_len = seq_len self.dropout = dropout self.ent = Parameter(torch.Tensor(num_nodes, hidden_channels)) self.rel = Parameter(torch.Tensor(num_rels, hidden_channels)) self.sub_gru = GRU(3 * hidden_channels, hidden_channels, num_layers, batch_first=True, bias=bias) self.obj_gru = GRU(3 * hidden_channels, hidden_channels, num_layers, batch_first=True, bias=bias) self.sub_lin = Linear(3 * hidden_channels, num_nodes, bias=bias) self.obj_lin = Linear(3 * hidden_channels, num_nodes, bias=bias) self.reset_parameters()
[docs] def reset_parameters(self): torch.nn.init.xavier_uniform_(self.ent, gain=math.sqrt(2.0)) torch.nn.init.xavier_uniform_(self.rel, gain=math.sqrt(2.0)) self.sub_gru.reset_parameters() self.obj_gru.reset_parameters() self.sub_lin.reset_parameters() self.obj_lin.reset_parameters()
[docs] @staticmethod def pre_transform(seq_len): r"""Precomputes history objects .. math:: \{ \mathcal{O}^{(t-k-1)}_r(s), \ldots, \mathcal{O}^{(t-1)}_r(s) \} of a :class:`torch_geometric.datasets.icews.EventDataset` with :math:`k` denoting the sequence length :obj:`seq_len`. """ class PreTransform(object): def __init__(self, seq_len): self.seq_len = seq_len self.inc = 5000 self.t_last = 0 self.sub_hist = self.increase_hist_node_size([]) self.obj_hist = self.increase_hist_node_size([]) def increase_hist_node_size(self, hist): hist_inc = torch.zeros((self.inc, self.seq_len + 1, 0)) return hist + hist_inc.tolist() def get_history(self, hist, node, rel): hists, ts = [], [] for s in range(seq_len): h = hist[node][s] hists += h ts.append(torch.full((len(h), ), s, dtype=torch.long)) node, r = torch.tensor(hists, dtype=torch.long).view( -1, 2).t().contiguous() node = node[r == rel] t = torch.cat(ts, dim=0)[r == rel] return node, t def step(self, hist): for i in range(len(hist)): hist[i] = hist[i][1:] hist[i].append([]) return hist def __call__(self, data): sub, rel, obj, t = data.sub, data.rel, data.obj, data.t if max(sub, obj) + 1 > len(self.sub_hist): # pragma: no cover self.sub_hist = self.increase_hist_node_size(self.sub_hist) self.obj_hist = self.increase_hist_node_size(self.obj_hist) # Delete last timestamp in history. if t > self.t_last: self.sub_hist = self.step(self.sub_hist) self.obj_hist = self.step(self.obj_hist) self.t_last = t # Save history in data object. data.h_sub, data.h_sub_t = self.get_history( self.sub_hist, sub, rel) data.h_obj, data.h_obj_t = self.get_history( self.obj_hist, obj, rel) # Add new event to history. self.sub_hist[sub][-1].append([obj, rel]) self.obj_hist[obj][-1].append([sub, rel]) return data def __repr__(self): # pragma: no cover return '{}(seq_len={})'.format(self.__class__.__name__, self.seq_len) return PreTransform(seq_len)
[docs] def forward(self, data): """Given a :obj:`data` batch, computes the forward pass. Args: data (torch_geometric.data.Data): The input data, holding subject :obj:`sub`, relation :obj:`rel` and object :obj:`obj` information with shape :obj:`[batch_size]`. In addition, :obj:`data` needs to hold history information for subjects, given by a vector of node indices :obj:`h_sub` and their relative timestamps :obj:`h_sub_t` and batch assignments :obj:`h_sub_batch`. The same information must be given for objects (:obj:`h_obj`, :obj:`h_obj_t`, :obj:`h_obj_batch`). """ assert 'h_sub_batch' in data and 'h_obj_batch' in data batch_size, seq_len = data.sub.size(0), self.seq_len h_sub_t = data.h_sub_t + data.h_sub_batch * seq_len h_obj_t = data.h_obj_t + data.h_obj_batch * seq_len h_sub = scatter_mean(self.ent[data.h_sub], h_sub_t, dim=0, dim_size=batch_size * seq_len).view( batch_size, seq_len, -1) h_obj = scatter_mean(self.ent[data.h_obj], h_obj_t, dim=0, dim_size=batch_size * seq_len).view( batch_size, seq_len, -1) sub = self.ent[data.sub].unsqueeze(1).repeat(1, seq_len, 1) rel = self.rel[data.rel].unsqueeze(1).repeat(1, seq_len, 1) obj = self.ent[data.obj].unsqueeze(1).repeat(1, seq_len, 1) _, h_sub = self.sub_gru(torch.cat([sub, h_sub, rel], dim=-1)) _, h_obj = self.obj_gru(torch.cat([obj, h_obj, rel], dim=-1)) h_sub, h_obj = h_sub.squeeze(0), h_obj.squeeze(0) h_sub = torch.cat([self.ent[data.sub], h_sub, self.rel[data.rel]], dim=-1) h_obj = torch.cat([self.ent[data.obj], h_obj, self.rel[data.rel]], dim=-1) h_sub = F.dropout(h_sub, p=self.dropout, training=self.training) h_obj = F.dropout(h_obj, p=self.dropout, training=self.training) log_prob_obj = F.log_softmax(self.sub_lin(h_sub), dim=1) log_prob_sub = F.log_softmax(self.obj_lin(h_obj), dim=1) return log_prob_obj, log_prob_sub
[docs] def test(self, logits, y): """Given ground-truth :obj:`y`, computes Mean Reciprocal Rank (MRR) and Hits at 1/3/10.""" _, perm = logits.sort(dim=1, descending=True) mask = (y.view(-1, 1) == perm) mrr = (1 / (mask.nonzero()[:, -1] + 1).to(torch.float)).mean().item() hits1 = mask[:, :1].sum().item() / y.size(0) hits3 = mask[:, :3].sum().item() / y.size(0) hits10 = mask[:, :10].sum().item() / y.size(0) return torch.tensor([mrr, hits1, hits3, hits10])