Shortcuts

ComplEx-WN18RR-single-gpu-wandb || ComplEx-WN18RR-single-gpu-hpo

ComplEx-WN18RR-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 19, 2024

备注

last run by LuYF-Lemon-love <luyanfeng_nlp@qq.com> on May 19, 2024

这一部分介绍如何用一个 GPU 在 WN18RR 知识图谱上训练 ComplEx [TWR+16],使用 wandb 记录实验结果。

导入数据

UniKE 有两个工具用于导入数据: unike.data.KGEDataLoader

import os
from unike.utils import WandbLogger
from unike.data import KGEDataLoader, BernSampler, TradTestSampler
from unike.module.model import ComplEx
from unike.module.loss import SoftplusLoss
from unike.module.strategy import NegativeSampling
from unike.config import Trainer, Tester

首先初始化 unike.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().set_config(
    project="unike",
    name="ComplEx-WN18RR",
    config=dict(
            in_path = os.path.join(os.path.dirname(__file__), '../../benchmarks/WN18RR/'),
            batch_size = 4096,
            neg_ent = 25,
            test = True,
            test_batch_size = 10,
            num_workers = 16,
            dim = 200,
            regul_rate = 1.0,
        use_tqdm = False,
            use_gpu = True,
            device = 'cuda:1',
            epochs = 2000,
            lr = 0.5,
            opt_method = 'adagrad',
            valid_interval = 100,
            log_interval = 100,
            save_interval = 100,
            save_path = '../../checkpoint/complex.pth'
    )
)

config = wandb_logger.config

UniKE 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。 unike.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

UniKE 提供了很多 KGE 模型,它们都是目前最常用的基线模型。我们下面将要导入 unike.module.model.ComplEx,它是第一个真正意义上的复数域模型。

# define the model
complEx = ComplEx(
    ent_tol = dataloader.get_ent_tol(),
    rel_tol = dataloader.get_rel_tol(),
    dim = config.dim
)

损失函数

我们这里使用了逻辑损失函数:unike.module.loss.SoftplusLossunike.module.strategy.NegativeSamplingunike.module.loss.SoftplusLoss 进行了封装,加入权重衰减等额外项。

# define the loss function
model = NegativeSampling(
    model = complEx,
    loss = SoftplusLoss(),
    regul_rate = config.regul_rate
)

训练模型

UniKE 将训练循环包装成了 unike.config.Trainer, 可以运行它的 unike.config.Trainer.run() 函数进行模型学习; 也可以通过传入 unike.config.Tester, 使得训练器能够在训练过程中评估模型。

# test the model
tester = Tester(model = complEx, data_loader = dataloader, use_tqdm = config.use_tqdm,
                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()

# close your wandb run
wandb_logger.finish()

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)

Gallery generated by Sphinx-Gallery

Docs

Access comprehensive developer documentation for UniKE

View Docs