# coding:utf-8
#
# unike/module/model/RotatE.py
#
# git pull from OpenKE-PyTorch by LuYF-Lemon-love <luyanfeng_nlp@qq.com> on May 7, 2023
# updated by LuYF-Lemon-love <luyanfeng_nlp@qq.com> on Jan 11, 2023
#
# 该头文件定义了 RotatE.
"""
RotatE - 将实体表示成复数向量，关系建模为复数向量空间的旋转。
"""
import torch
import typing
import torch.nn as nn
from .Model import Model
from typing_extensions import override
[文档]class RotatE(Model):
"""
``RotatE`` :cite:`RotatE` 提出于 2019 年，将实体表示成复数向量，关系建模为复数向量空间的旋转。
评分函数为:
.. math::
\gamma - \parallel \mathbf{h} \circ \mathbf{r} - \mathbf{t} \parallel_{L_2}
:math:`\circ` 表示哈达玛积（Hadamard product），正三元组的评分函数的值越大越好，如果想获得更详细的信息请访问 :ref:`RotatE <rotate>`。
例子::
from unike.data import KGEDataLoader, UniSampler, TradTestSampler
from unike.module.model import RotatE
from unike.module.loss import SigmoidLoss
from unike.module.strategy import NegativeSampling
from unike.config import Trainer, Tester
# dataloader for training
dataloader = KGEDataLoader(
in_path = '../../benchmarks/WN18RR/',
batch_size = 2000,
neg_ent = 64,
test = True,
test_batch_size = 10,
num_workers = 16,
train_sampler = UniSampler,
test_sampler = TradTestSampler
)
# define the model
rotate = RotatE(
ent_tol = dataloader.get_ent_tol(),
rel_tol = dataloader.get_rel_tol(),
dim = 1024,
margin = 6.0,
epsilon = 2.0,
)
# define the loss function
model = NegativeSampling(
model = rotate,
loss = SigmoidLoss(adv_temperature = 2),
regul_rate = 0.0,
)
# test the model
tester = Tester(model = rotate, data_loader = dataloader, use_gpu = True, device = 'cuda:1')
# train the model
trainer = Trainer(model = model, data_loader = dataloader.train_dataloader(), epochs = 6000,
lr = 2e-5, opt_method = 'adam', use_gpu = True, device = 'cuda:1',
tester = tester, test = True, valid_interval = 100,
log_interval = 100, save_interval = 100,
save_path = '../../checkpoint/rotate.pth', use_wandb = False)
trainer.run()
"""
[文档] def __init__(
self,
ent_tol: int,
rel_tol: int,
dim: int = 100,
margin: float = 6.0,
epsilon: float = 2.0):
"""创建 RotatE 对象。
:param ent_tol: 实体的个数
:type ent_tol: int
:param rel_tol: 关系的个数
:type rel_tol: int
:param dim: 实体和关系嵌入向量的维度
:type dim: int
:param margin: 原论文中损失函数的 gamma。
:type margin: float
:param epsilon: RotatE 原论文对应的源代码固定为 2.0。
:type epsilon: float
"""
super(RotatE, self).__init__(ent_tol, rel_tol)
#: RotatE 原论文对应的源代码固定为 2.0。
self.epsilon: int = epsilon
#: RotatE 原论文的实现中将实体嵌入向量的维度指定为 ``dim`` 的 2 倍。
#: 因为实体嵌入向量需要划分为实部和虚部。
self.dim_e: int = dim * 2
#: 关系嵌入向量的维度，为 ``dim``。
self.dim_r: int = dim
#: 根据实体个数，创建的实体嵌入。
self.ent_embeddings: torch.nn.Embedding = nn.Embedding(self.ent_tol, self.dim_e)
#: 根据关系个数，创建的关系嵌入。
self.rel_embeddings: torch.nn.Embedding = nn.Embedding(self.rel_tol, self.dim_r)
self.ent_embedding_range = nn.Parameter(
torch.Tensor([(margin + self.epsilon) / self.dim_e]),
requires_grad=False
)
nn.init.uniform_(
tensor = self.ent_embeddings.weight.data,
a=-self.ent_embedding_range.item(),
b=self.ent_embedding_range.item()
)
self.rel_embedding_range = nn.Parameter(
torch.Tensor([(margin + self.epsilon) / self.dim_r]),
requires_grad=False
)
nn.init.uniform_(
tensor = self.rel_embeddings.weight.data,
a=-self.rel_embedding_range.item(),
b=self.rel_embedding_range.item()
)
#: 原论文中损失函数的 gamma。
self.margin: torch.nn.parameter.Parameter = nn.Parameter(torch.Tensor([margin]))
self.margin.requires_grad = False
[文档] @override
def forward(
self,
triples: torch.Tensor,
negs: torch.Tensor = None,
mode: str = 'single') -> torch.Tensor:
"""
定义每次调用时执行的计算。
:py:class:`torch.nn.Module` 子类必须重写 :py:meth:`torch.nn.Module.forward`。
:param triples: 正确的三元组
:type triples: torch.Tensor
:param negs: 负三元组类别
:type negs: torch.Tensor
:param mode: 模式
:type triples: str
:returns: 三元组的得分
:rtype: torch.Tensor
"""
head_emb, relation_emb, tail_emb = self.tri2emb(triples, negs, mode)
score = self.margin - self._calc(head_emb, relation_emb, tail_emb)
return score
[文档] def _calc(
self,
h: torch.Tensor,
r: torch.Tensor,
t: torch.Tensor) -> torch.Tensor:
"""计算 RotatE 的评分函数。
利用 :py:func:`torch.chunk` 拆分实体嵌入向量获得复数的实部和虚部。
原论文使用 L1-norm 作为距离函数，而这里使用的 L2-norm 作为距离函数。
:param h: 头实体的向量。
:type h: torch.Tensor
:param r: 关系的向量。
:type r: torch.Tensor
:param t: 尾实体的向量。
:type t: torch.Tensor
:returns: 三元组的得分
:rtype: torch.Tensor
"""
pi = self.pi_const
re_head, im_head = torch.chunk(h, 2, dim=-1)
re_tail, im_tail = torch.chunk(t, 2, dim=-1)
# Make phases of relations uniformly distributed in [-pi, pi]
phase_relation = r / (self.rel_embedding_range.item() / pi)
re_relation = torch.cos(phase_relation)
im_relation = torch.sin(phase_relation)
re_score = re_head * re_relation - im_head * im_relation
im_score = re_head * im_relation + im_head * re_relation
re_score = re_score - re_tail
im_score = im_score - im_tail
score = torch.stack([re_score, im_score], dim = 0)
score = score.norm(dim = 0).sum(dim = -1)
return score
[文档] @override
def predict(
self,
data: dict[str, typing.Union[torch.Tensor,str]],
mode) -> torch.Tensor:
"""RotatE 的推理方法。
:param data: 数据。
:type data: dict[str, typing.Union[torch.Tensor,str]]
:returns: 三元组的得分
:rtype: torch.Tensor
"""
triples = data["positive_sample"]
head_emb, relation_emb, tail_emb = self.tri2emb(triples, mode=mode)
score = self.margin - self._calc(head_emb, relation_emb, tail_emb)
return score
[文档] def regularization(
self,
data: dict[str, typing.Union[torch.Tensor, str]]) -> torch.Tensor:
"""L2 正则化函数（又称权重衰减），在损失函数中用到。
:param data: 数据。
:type data: dict[str, typing.Union[torch.Tensor, str]]
:returns: 模型参数的正则损失
:rtype: torch.Tensor
"""
pos_sample = data["positive_sample"]
neg_sample = data["negative_sample"]
mode = data["mode"]
pos_head_emb, pos_relation_emb, pos_tail_emb = self.tri2emb(pos_sample)
if mode == "bern":
neg_head_emb, neg_relation_emb, neg_tail_emb = self.tri2emb(neg_sample)
else:
neg_head_emb, neg_relation_emb, neg_tail_emb = self.tri2emb(pos_sample, neg_sample, mode)
pos_regul = (torch.mean(pos_head_emb ** 2) +
torch.mean(pos_relation_emb ** 2) +
torch.mean(pos_tail_emb ** 2)) / 3
neg_regul = (torch.mean(neg_head_emb ** 2) +
torch.mean(neg_relation_emb ** 2) +
torch.mean(neg_tail_emb ** 2)) / 3
regul = (pos_regul + neg_regul) / 2
return regul
[文档]def get_rotate_hpo_config() -> dict[str, dict[str, typing.Any]]:
"""返回 :py:class:`RotatE` 的默认超参数优化配置。
默认配置为::
parameters_dict = {
'model': {
'value': 'RotatE'
},
'dim': {
'values': [256, 512, 1024]
},
'margin': {
'values': [1.0, 3.0, 6.0]
},
'epsilon': {
'value': 2.0
}
}
:returns: :py:class:`RotatE` 的默认超参数优化配置
:rtype: dict[str, dict[str, typing.Any]]
"""
parameters_dict = {
'model': {
'value': 'RotatE'
},
'dim': {
'values': [256, 512, 1024]
},
'margin': {
'values': [1.0, 3.0, 6.0]
},
'epsilon': {
'value': 2.0
}
}
return parameters_dict