备注
Go to the end to download the full example code
TransR-FB15K237-single-gpu || TransR-FB15K237-single-gpu-wandb || TransR-FB15K237-single-gpu-hpo || TransR-FB15K237-accelerate
TransR-FB15K237-single-gpu-wandb¶
备注
created by LuYF-Lemon-love <luyanfeng_nlp@qq.com> on May 7, 2023
备注
updated by LuYF-Lemon-love <luyanfeng_nlp@qq.com> on May 13, 2024
备注
last run by LuYF-Lemon-love <luyanfeng_nlp@qq.com> on May 13, 2024
这一部分介绍如何用一个 GPU 在 FB15K237 知识图谱上训练 TransR [LLS+15],使用 wandb 记录实验结果。
导入数据¶
pybind11-OpenKE 有两个工具用于导入数据: pybind11_ke.data.KGEDataLoader。
from pybind11_ke.utils import WandbLogger
from pybind11_ke.data import KGEDataLoader, BernSampler, TradTestSampler
from pybind11_ke.module.model import TransE, TransR
from pybind11_ke.module.loss import MarginLoss
from pybind11_ke.module.strategy import NegativeSampling
from pybind11_ke.config import Trainer, Tester
首先初始化 pybind11_ke.utils.WandbLogger 日志记录器,它是对 wandb 初始化操作的一层简单封装。
wandb_logger = WandbLogger(
project="pybind11-ke",
name="TransR-FB15K237",
config=dict(
in_path = '../../benchmarks/FB15K237/',
batch_size = 2048,
neg_ent = 25,
test = True,
test_batch_size = 10,
num_workers = 16,
dim = 100,
dim_e = 100,
dim_r = 100,
p_norm = 1,
norm_flag = True,
rand_init = False,
margin_e = 5.0,
margin_r = 4.0,
epochs_e = 1,
lr_e = 0.5,
opt_method = "sgd",
use_gpu = True,
device = 'cuda:0',
epochs_r = 1000,
lr_r = 1.0,
valid_interval = 100,
log_interval = 100,
save_interval = 100,
save_path = '../../checkpoint/transr.pth'
)
)
config = wandb_logger.config
pybind11-OpenKE 提供了很多数据集,它们很多都是 KGE 原论文发表时附带的数据集。
pybind11_ke.data.KGEDataLoader 包含 in_path 用于传递数据集目录。
# 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
)
导入模型¶
pybind11-OpenKE 提供了很多 KGE 模型,它们都是目前最常用的基线模型。我们首先导入
pybind11_ke.module.model.TransE,它是最简单的平移模型,
因为为了避免过拟合,pybind11_ke.module.model.TransR 实体和关系的嵌入向量初始化为
pybind11_ke.module.model.TransE 的结果。
# define the transe
transe = TransE(
ent_tol = dataloader.get_ent_tol(),
rel_tol = dataloader.get_rel_tol(),
dim = config.dim,
p_norm = config.p_norm,
norm_flag = config.norm_flag)
下面导入 pybind11_ke.module.model.TransR 模型,
是一个为实体和关系嵌入向量分别构建了独立的向量空间,将实体向量投影到特定的关系向量空间进行平移操作的模型。
transr = TransR(
ent_tol = dataloader.get_ent_tol(),
rel_tol = dataloader.get_rel_tol(),
dim_e = config.dim_e,
dim_r = config.dim_r,
p_norm = config.p_norm,
norm_flag = config.norm_flag,
rand_init = config.rand_init)
损失函数¶
我们这里使用了 TransE [BUGD+13] 原论文使用的损失函数:pybind11_ke.module.loss.MarginLoss,
pybind11_ke.module.strategy.NegativeSampling 对
pybind11_ke.module.loss.MarginLoss 进行了封装,加入权重衰减等额外项。
model_e = NegativeSampling(
model = transe,
loss = MarginLoss(margin = config.margin_e)
)
model_r = NegativeSampling(
model = transr,
loss = MarginLoss(margin = config.margin_r)
)
训练模型¶
pybind11-OpenKE 将训练循环包装成了 pybind11_ke.config.Trainer,
可以运行它的 pybind11_ke.config.Trainer.run() 函数进行模型学习;
也可以通过传入 pybind11_ke.config.Tester,
使得训练器能够在训练过程中评估模型。
# pretrain transe
trainer = Trainer(model = model_e, data_loader = dataloader.train_dataloader(),
epochs = config.epochs_e, lr = config.lr_e, opt_method = config.opt_method,
use_gpu = config.use_gpu, device = config.device)
trainer.run()
parameters = transe.get_parameters()
transe.save_parameters("../../checkpoint/transr_transe.json")
# test the transr
tester = Tester(model = transr, data_loader = dataloader, use_tqdm = False,
use_gpu = config.use_gpu, device = config.device)
# train transr
transr.set_parameters(parameters)
trainer = Trainer(model = model_r, data_loader = dataloader.train_dataloader(),
epochs = config.epochs_r, lr = config.lr_r, opt_method = config.opt_method,
use_gpu = config.use_gpu, device = config.device,
tester = tester, test = True, valid_interval = config.valid_interval,
log_interval = config.log_interval, save_interval = config.save_interval,
save_path = config.save_path, use_wandb = True)
trainer.run()
备注
上述代码的运行日志可以从 此处 下载。
备注
上述代码的运行报告可以从 此处 下载。
Total running time of the script: ( 0 minutes 0.000 seconds)