TransD.py

import numpy
import os
import json
import torch as th
import torch.sparse as sparse
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from . import BaseModel, register_model
import dgl

@register_model('TransD')
class TransD(BaseModel):
    @classmethod
    def build_model_from_args(cls, args, hg):
        return cls(args, hg)

    def __init__(self, args, hg):
        super(TransD, self).__init__()
        self.device = args.device
        self.ent_num = hg.num_nodes()
        self.rel_num = len(hg.canonical_etypes)
        self.ent_dim = args.ent_dim
        self.rel_dim = args.rel_dim
        self.margin = args.margin
        self.dis_norm = args.dis_norm

        self.n_emb = nn.Embedding(self.ent_num, self.ent_dim)
        self.r_emb = nn.Embedding(self.rel_num, self.rel_dim)
        self.n_emb_p = nn.Embedding(self.ent_num, self.ent_dim)
        self.r_emb_p = nn.Embedding(self.rel_num, self.rel_dim)
        
        nn.init.xavier_uniform_(self.n_emb.weight.data)
        nn.init.xavier_uniform_(self.r_emb.weight.data)
        nn.init.xavier_uniform_(self.n_emb_p.weight.data)
        nn.init.xavier_uniform_(self.r_emb_p.weight.data)

    def _resize(self, tensor, axis, size):
        shape = tensor.size()
        axis = len(shape)+axis
        osize = shape[axis]
        if osize == size:
            return tensor
        elif osize > size:
            return th.narrow(tensor, axis, 0, size)
        else:
            for i in range(len(shape)):
                if i == axis:
                    paddings = [0, size - osize] + paddings
                else:
                    paddings = [0, 0] + paddings
            return F.pad(tensor, pad = paddings, mode = "constant", value = 0)

    def _transfer(self, n, n_emb_p, r_emb_p):
        if n.shape[0] != r_emb_p.shape[0]:
            n = n.view(-1, r_emb_p.shape[0], n.shape[-1])
            n_emb_p = n_emb_p.view(-1, r_emb_p.shape[0], n_emb_p.shape[-1])
            r_emb_p = r_emb_p.view(-1, r_emb_p.shape[0], r_emb_p.shape[-1])
            n = F.normalize(
                self._resize(n, -1, r_emb_p.size()[-1]) + th.sum(n * n_emb_p, -1, True) * r_emb_p,
                p = 2, 
                dim = -1
            )            
            return n.view(-1, n.shape[-1])
        else:
            return F.normalize(
                self._resize(n, -1, r_emb_p.shape[-1]) + th.sum(n * n_emb_p, -1, True) * r_emb_p,
                p = 2, dim = -1
            )

    def forward(self, h, r, t):
        if self.training:
            self.n_emb.weight.data = F.normalize(self.n_emb.weight.data, p=2.0, dim=-1)
            self.r_emb.weight.data = F.normalize(self.r_emb.weight.data, p=2.0, dim=-1)
            self.n_emb_p.weight.data = F.normalize(self.n_emb_p.weight.data, p=2.0, dim=-1)
            self.r_emb_p.weight.data = F.normalize(self.r_emb_p.weight.data, p=2.0, dim=-1)

        if h.shape == th.Size([]):
            h = h.view(1)
        if r.shape == th.Size([]):
            r = r.view(1)
        if t.shape == th.Size([]):
            t = t.view(1)
        h = h.to(self.device)
        r = r.to(self.device)
        t = t.to(self.device)
        h_emb = self.n_emb(h)
        r_emb = self.r_emb(r)
        t_emb = self.n_emb(t)
        h_emb_p = self.n_emb_p(h)
        r_emb_p = self.r_emb_p(r)
        t_emb_p = self.n_emb_p(t)
        h_emb = self._transfer(h_emb, h_emb_p, r_emb_p)
        t_emb = self._transfer(t_emb, t_emb_p, r_emb_p)
        h_emb = F.normalize(h_emb, 2.0, -1)
        r_emb = F.normalize(r_emb, 2.0, -1)
        t_emb = F.normalize(t_emb, 2.0, -1)
        score = th.norm(h_emb+r_emb-t_emb, self.dis_norm, dim=-1)
        return score