# coding:utf-8
#
# unike/module/model/RESCAL.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 31, 2023
#
# 该头文件定义了 RESCAL.
"""
RESCAL - 一个张量分解模型。
"""
import torch
import typing
import torch.nn as nn
from .Model import Model
from typing_extensions import override
[文档]class RESCAL(Model):
"""
``RESCAL`` :cite:`RESCAL` 提出于 2011 年，是很多张量分解模型的基石，模型较复杂。
评分函数为:
.. math::
-\mathbf{h}^T \mathbf{M}_r \mathbf{t}
正三元组的评分函数的值越小越好，如果想获得更详细的信息请访问 :ref:`RESCAL <rescal>`。
例子::
from unike.utils import WandbLogger
from unike.data import KGEDataLoader, BernSampler, TradTestSampler
from unike.module.model import RESCAL
from unike.module.loss import MarginLoss
from unike.module.strategy import NegativeSampling
from unike.config import Trainer, Tester
wandb_logger = WandbLogger(
project="unike",
name="RESCAL-FB15K237",
config=dict(
in_path = '../../benchmarks/FB15K237/',
batch_size = 2048,
neg_ent = 25,
test = True,
test_batch_size = 10,
num_workers = 16,
dim = 50,
margin = 1.0,
use_gpu = True,
device = 'cuda:0',
epochs = 1000,
lr = 0.1,
opt_method = 'adagrad',
valid_interval = 100,
log_interval = 100,
save_interval = 100,
save_path = '../../checkpoint/rescal.pth'
)
)
config = wandb_logger.config
# dataloader for training
dataloader = KGEDataLoader(
in_path = config.in_path,
batch_size = config.batch_size,
neg_ent = config.neg_ent,
test = config.test,
test_batch_size = config.test_batch_size,
num_workers = config.num_workers,
train_sampler = BernSampler,
test_sampler = TradTestSampler
)
# define the model
rescal = RESCAL(
ent_tol = dataloader.get_ent_tol(),
rel_tol = dataloader.get_rel_tol(),
dim = config.dim
)
# define the loss function
model = NegativeSampling(
model = rescal,
loss = MarginLoss(margin = config.margin)
)
# test the model
tester = Tester(model = rescal, data_loader = dataloader, use_gpu = config.use_gpu, device = config.device)
# train the model
trainer = Trainer(model = model, data_loader = dataloader.train_dataloader(), epochs = config.epochs,
lr = config.lr, opt_method = config.opt_method, use_gpu = config.use_gpu, device = config.device,
tester = tester, test = config.test, valid_interval = config.valid_interval,
log_interval = config.log_interval, save_interval = config.save_interval,
save_path = config.save_path, wandb_logger = wandb_logger)
trainer.run()
"""
[文档] def __init__(
self,
ent_tol: int,
rel_tol: int,
dim: int = 100):
"""创建 RESCAL 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
"""
super(RESCAL, self).__init__(ent_tol, rel_tol)
#: 实体和关系嵌入向量的维度
self.dim: 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)
#: 根据关系个数，创建的关系矩阵
self.rel_matrices: torch.nn.Embedding = nn.Embedding(self.rel_tol, self.dim * self.dim)
nn.init.xavier_uniform_(self.ent_embeddings.weight.data)
nn.init.xavier_uniform_(self.rel_matrices.weight.data)
[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, tail_emb = self.tri2emb(triples, negs, mode)
rel_matric = self.rel_matrices(triples[:, 1])
score = self._calc(head_emb, rel_matric, tail_emb)
return score
[文档] @override
def tri2emb(
self,
triples: torch.Tensor,
negs: torch.Tensor = None,
mode: str = 'single') -> tuple[torch.Tensor, torch.Tensor]:
"""
返回三元组对应的嵌入向量。
: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: 头实体和尾实体的嵌入向量
:rtype: tuple[torch.Tensor, torch.Tensor]
"""
if mode == "single":
head_emb = self.ent_embeddings(triples[:, 0]).unsqueeze(1)
tail_emb = self.ent_embeddings(triples[:, 2]).unsqueeze(1)
elif mode == "head-batch" or mode == "head_predict":
if negs is None:
head_emb = self.ent_embeddings.weight.data.unsqueeze(0)
else:
head_emb = self.ent_embeddings(negs)
tail_emb = self.ent_embeddings(triples[:, 2]).unsqueeze(1)
elif mode == "tail-batch" or mode == "tail_predict":
head_emb = self.ent_embeddings(triples[:, 0]).unsqueeze(1)
if negs is None:
tail_emb = self.ent_embeddings.weight.data.unsqueeze(0)
else:
tail_emb = self.ent_embeddings(negs)
return head_emb, tail_emb
[文档] def _calc(
self,
h: torch.Tensor,
r: torch.Tensor,
t: torch.Tensor) -> torch.Tensor:
"""计算 RESCAL 的评分函数。
:param h: 头实体的向量。
:type h: torch.Tensor
:param r: 关系矩阵。
: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
"""
r = r.view(-1, self.dim, self.dim)
r = r.unsqueeze(dim=1)
h = h.unsqueeze(dim=-2)
hr = torch.matmul(h, r)
hr = hr.squeeze(dim=-2)
return -torch.sum(hr * t, -1)
[文档] @override
def predict(
self,
data: dict[str, typing.Union[torch.Tensor,str]],
mode) -> torch.Tensor:
"""RESCAL 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, tail_emb = self.tri2emb(triples, mode=mode)
rel_matric = self.rel_matrices(triples[:, 1])
score = self._calc(head_emb, rel_matric, 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_tail_emb = self.tri2emb(pos_sample)
pos_rel_transfer = self.rel_matrices(pos_sample[:, 1])
if mode == "bern":
neg_head_emb, neg_tail_emb = self.tri2emb(neg_sample)
else:
neg_head_emb, neg_tail_emb = self.tri2emb(pos_sample, neg_sample, mode)
neg_rel_transfer = self.rel_matrices(pos_sample[:, 1])
pos_regul = (torch.mean(pos_head_emb ** 2) +
torch.mean(pos_tail_emb ** 2) +
torch.mean(pos_rel_transfer ** 2)) / 3
neg_regul = (torch.mean(neg_head_emb ** 2) +
torch.mean(neg_tail_emb ** 2) +
torch.mean(neg_rel_transfer ** 2)) / 3
regul = (pos_regul + neg_regul) / 2
return regul
[文档]def get_rescal_hpo_config() -> dict[str, dict[str, typing.Any]]:
"""返回 :py:class:`RESCAL` 的默认超参数优化配置。
默认配置为::
parameters_dict = {
'model': {
'value': 'RESCAL'
},
'dim': {
'values': [50, 100, 200]
}
}
:returns: :py:class:`RESCAL` 的默认超参数优化配置
:rtype: dict[str, dict[str, typing.Any]]
"""
parameters_dict = {
'model': {
'value': 'RESCAL'
},
'dim': {
'values': [50, 100, 200]
}
}
return parameters_dict