TransE-FB15K-single-gpu || TransE-FB15K-single-gpu-wandb || TransE-FB15K-single-gpu-hpo || TransE-FB15K-multigpu || TransE-FB15K-multigpu-wandb || TransE-FB15K237-single-gpu-wandb || TransE-WN18RR-single-gpu-adv-wandb

TransE-FB15K-multigpu-wandb

这一部分介绍如何用多个 GPU 在 FB15k 知识图谱上训练 TransE [BUGD+13]，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.TrainDataLoader 和 pybind11_ke.data.TestDataLoader。

from pybind11_ke.utils import WandbLogger
from pybind11_ke.config import trainer_distributed_data_parallel
from pybind11_ke.module.model import TransE
from pybind11_ke.module.loss import MarginLoss
from pybind11_ke.module.strategy import NegativeSampling
from pybind11_ke.data import TrainDataLoader

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="transe",
    config=dict(
            in_path = "../../benchmarks/FB15K/",
            nbatches = 200,
            threads = 8,
            sampling_mode = "normal",
            bern = True,
            neg_ent = 25,
            neg_rel = 0,
            dim = 50,
            p_norm = 1,
            norm_flag = True,
            margin = 1.0,
            use_gpu = True,
            device = 'cuda:1',
            epochs = 1000,
            lr = 0.01,
            opt_method = "adam",
            test = True,
            valid_interval = 100,
            log_interval = 100,
            save_interval = 100,
            save_path = '../../checkpoint/transe.pth',
            type_constrain = True
    )
)

config = wandb_logger.config

pybind11-KE 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.TrainDataLoader 包含 in_path 用于传递数据集目录。

# dataloader for training
train_dataloader = TrainDataLoader(
    in_path = config.in_path,
    nbatches = config.nbatches,
    threads = config.threads,
    sampling_mode = config.sampling_mode,
    bern = config.bern,
    neg_ent = config.neg_ent,
    neg_rel = config.neg_rel)

---

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。我们下面将要导入 pybind11_ke.module.model.TransE，它是最简单的平移模型。

# define the model
transe = TransE(
    ent_tol = train_dataloader.get_ent_tol(),
    rel_tol = train_dataloader.get_rel_tol(),
    dim = config.dim,
    p_norm = config.p_norm,
    norm_flag = config.norm_flag)

---

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。

# define the loss function
model = NegativeSampling(
    model = transe,
    loss = MarginLoss(margin = config.margin),
    batch_size = train_dataloader.get_batch_size()
)

---

训练模型

pybind11-OpenKE 将训练循环包装成了 pybind11_ke.config.trainer_distributed_data_parallel() 函数， 进行并行训练，该函数必须由 if __name__ == '__main__' 保护。

if __name__ == "__main__":

    print("Start parallel training...")

    trainer_distributed_data_parallel(model = model, data_loader = train_dataloader,
            epochs = config.epochs, lr = config.lr, opt_method = config.opt_method,
            test = config.test, valid_interval = config.valid_interval, log_interval = config.log_interval,
            save_interval = config.save_interval, save_path = config.save_path,
            type_constrain = True, use_wandb=True)

    # close your wandb run
    wandb_logger.finish()