深度学习 · 4月 21, 2022 0

TransE模型 PyTorch版本实现

内容纲要

版本1

出自 https://github.com/mklimasz/TransE-PyTorch/blob/master/model.py

import numpy as np
import torch
import torch.nn as nn

class TransE(nn.Module):

    def __init__(self, entity_count, relation_count, device, norm=1, dim=100, margin=1.0):
        super(TransE, self).__init__()
        self.entity_count = entity_count
        self.relation_count = relation_count
        self.device = device
        self.norm = norm
        self.dim = dim
        self.entities_emb = self._init_enitity_emb()
        self.relations_emb = self._init_relation_emb()
        self.criterion = nn.MarginRankingLoss(margin=margin, reduction='none')

    def _init_enitity_emb(self):
        entities_emb = nn.Embedding(num_embeddings=self.entity_count + 1,
                                    embedding_dim=self.dim,
                                    padding_idx=self.entity_count)
        uniform_range = 6 / np.sqrt(self.dim)
        entities_emb.weight.data.uniform_(-uniform_range, uniform_range)
        return entities_emb

    def _init_relation_emb(self):
        relations_emb = nn.Embedding(num_embeddings=self.relation_count + 1,
                                     embedding_dim=self.dim,
                                     padding_idx=self.relation_count)
        uniform_range = 6 / np.sqrt(self.dim)
        relations_emb.weight.data.uniform_(-uniform_range, uniform_range)
        # -1 to avoid nan for OOV vector
        relations_emb.weight.data[:-1, :].div_(relations_emb.weight.data[:-1, :].norm(p=1, dim=1, keepdim=True))
        return relations_emb

    def forward(self, positive_triplets: torch.LongTensor, negative_triplets: torch.LongTensor):
        """Return model losses based on the input.
        :param positive_triplets: triplets of positives in Bx3 shape (B - batch, 3 - head, relation and tail)
        :param negative_triplets: triplets of negatives in Bx3 shape (B - batch, 3 - head, relation and tail)
        :return: tuple of the model loss, positive triplets loss component, negative triples loss component
        """
        # -1 to avoid nan for OOV vector
        self.entities_emb.weight.data[:-1, :].div_(self.entities_emb.weight.data[:-1, :].norm(p=2, dim=1, keepdim=True))

        assert positive_triplets.size()[1] == 3
        positive_distances = self._distance(positive_triplets)

        assert negative_triplets.size()[1] == 3
        negative_distances = self._distance(negative_triplets)

        return self.loss(positive_distances, negative_distances), positive_distances, negative_distances

    def predict(self, triplets: torch.LongTensor):
        """Calculated dissimilarity score for given triplets.
        :param triplets: triplets in Bx3 shape (B - batch, 3 - head, relation and tail)
        :return: dissimilarity score for given triplets
        """
        return self._distance(triplets)

    def loss(self, positive_distances, negative_distances):
        target = torch.tensor([-1], dtype=torch.long, device=self.device)
        return self.criterion(positive_distances, negative_distances, target)

    def _distance(self, triplets):
        """Triplets should have shape Bx3 where dim 3 are head id, relation id, tail id."""
        assert triplets.size()[1] == 3
        heads = triplets[:, 0]
        relations = triplets[:, 1]
        tails = triplets[:, 2]
        return (self.entities_emb(heads) + self.relations_emb(relations) - self.entities_emb(tails)).norm(p=self.norm,
                                                                                                          dim=1)

版本2实现

https://github.com/toooooodo/pytorch-TransE/blob/master/model.py

import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import torch.nn.functional as F
from prepare_data import TrainSet, TestSet
import math

class TranE(nn.Module):
    def __init__(self, entity_num, relation_num, device, dim=50, d_norm=2, gamma=1):
        """
        :param entity_num: number of entities
        :param relation_num: number of relations
        :param dim: embedding dim
        :param device:
        :param d_norm: measure d(h+l, t), either L1-norm or L2-norm
        :param gamma: margin hyperparameter
        """
        super(TranE, self).__init__()
        self.dim = dim
        self.d_norm = d_norm
        self.device = device
        self.gamma = torch.FloatTensor([gamma]).to(self.device)
        self.entity_num = entity_num
        self.relation_num = relation_num
        self.entity_embedding = nn.Embedding.from_pretrained(
            torch.empty(entity_num, self.dim).uniform_(-6 / math.sqrt(self.dim), 6 / math.sqrt(self.dim)), freeze=False)
        self.relation_embedding = nn.Embedding.from_pretrained(
            torch.empty(relation_num, self.dim).uniform_(-6 / math.sqrt(self.dim), 6 / math.sqrt(self.dim)),
            freeze=False)
        # l <= l / ||l||
        relation_norm = torch.norm(self.relation_embedding.weight.data, dim=1, keepdim=True)
        self.relation_embedding.weight.data = self.relation_embedding.weight.data / relation_norm

    def forward(self, pos_head, pos_relation, pos_tail, neg_head, neg_relation, neg_tail):
        """
        :param pos_head: [batch_size]
        :param pos_relation: [batch_size]
        :param pos_tail: [batch_size]
        :param neg_head: [batch_size]
        :param neg_relation: [batch_size]
        :param neg_tail: [batch_size]
        :return: triples loss
        """
        pos_dis = self.entity_embedding(pos_head) + self.relation_embedding(pos_relation) - self.entity_embedding(
            pos_tail)
        neg_dis = self.entity_embedding(neg_head) + self.relation_embedding(neg_relation) - self.entity_embedding(
            neg_tail)
        # return pos_head_and_relation, pos_tail, neg_head_and_relation, neg_tail
        return self.calculate_loss(pos_dis, neg_dis).requires_grad_()

    def calculate_loss(self, pos_dis, neg_dis):
        """
        :param pos_dis: [batch_size, embed_dim]
        :param neg_dis: [batch_size, embed_dim]
        :return: triples loss: [batch_size]
        """
        distance_diff = self.gamma + torch.norm(pos_dis, p=self.d_norm, dim=1) - torch.norm(neg_dis, p=self.d_norm,
                                                                                            dim=1)
        return torch.sum(F.relu(distance_diff))

    def tail_predict(self, head, relation, tail, k=10):
        """
        to do tail prediction hits@k
        :param head: [batch_size]
        :param relation: [batch_size]
        :param tail: [batch_size]
        :param k: hits@k
        :return:
        """
        # head: [batch_size]
        # h_and_r: [batch_size, embed_size] => [batch_size, 1, embed_size] => [batch_size, N, embed_size]
        h_and_r = self.entity_embedding(head) + self.relation_embedding(relation)
        h_and_r = torch.unsqueeze(h_and_r, dim=1)
        h_and_r = h_and_r.expand(h_and_r.shape[0], self.entity_num, self.dim)
        # embed_tail: [batch_size, N, embed_size]
        embed_tail = self.entity_embedding.weight.data.expand(h_and_r.shape[0], self.entity_num, self.dim)
        # indices: [batch_size, k]
        values, indices = torch.topk(torch.norm(h_and_r - embed_tail, dim=2), k, dim=1, largest=False)
        # tail: [batch_size] => [batch_size, 1]
        tail = tail.view(-1, 1)
        return torch.sum(torch.eq(indices, tail)).item()

if __name__ == '__main__':
    train_data_set = TrainSet()
    test_data_set = TestSet()
    test_data_set.convert_word_to_index(train_data_set.entity_to_index, train_data_set.relation_to_index,
                                        test_data_set.raw_data)
    train_loader = DataLoader(train_data_set, batch_size=32, shuffle=True)
    test_loader = DataLoader(test_data_set, batch_size=32, shuffle=True)
    for batch_idx, data in enumerate(test_loader):
        print(data.shape)
        break
%d 博主赞过: