run_adv_wgan_gp_4modal.py

import torch
from mmkgc.config import Tester, WCGTrainerKuai16KGP
from mmkgc.module.model import AdvRelRotatEKuai16K
from mmkgc.module.loss import SigmoidLoss
from mmkgc.module.strategy import NegativeSamplingGP
from mmkgc.data import TrainDataLoader, TestDataLoader
from mmkgc.adv.modules import CombinedGenerator2

from args import get_args

if __name__ == "__main__":
    args = get_args()
    print(args)
    # set the seed
    torch.manual_seed(args.seed)
    torch.cuda.manual_seed_all(args.seed)
    # dataloader for training
    train_dataloader = TrainDataLoader(
        in_path="./benchmarks/" + args.dataset + '/',
        batch_size=args.batch_size,
        threads=8,
        sampling_mode="normal",
        bern_flag=1,
        filter_flag=1,
        neg_ent=args.neg_num,
        neg_rel=0
    )
    # dataloader for test
    test_dataloader = TestDataLoader(
        "./benchmarks/" + args.dataset + '/', "link")
    img_emb = torch.load('./embeddings/' + args.dataset + '-visual.pth')
    text_emb = torch.load('./embeddings/' + args.dataset + '-textual.pth')
    audio_emb = torch.load('./embeddings/' + args.dataset + '-audio.pth')
    video_emb = torch.load('./embeddings/' + args.dataset + '-video.pth')
    # define the model
    kge_score = AdvRelRotatEKuai16K(
        ent_tot=train_dataloader.get_ent_tot(),
        rel_tot=train_dataloader.get_rel_tot(),
        dim=args.dim,
        margin=args.margin,
        epsilon=2.0,
        img_emb=img_emb,
        text_emb=text_emb,
        audio_emb=audio_emb,
        video_emb=video_emb
    )
    print(kge_score)
    # define the loss function
    model = NegativeSamplingGP(
        model=kge_score,
        loss=SigmoidLoss(adv_temperature=args.adv_temp),
        batch_size=train_dataloader.get_batch_size(),
        regul_rate=0.00001
    )
    
    adv_generator = CombinedGenerator2(
        noise_dim=64,
        structure_dim=2*args.dim,
        img_dim=4*args.dim
    )
    tester = Tester(model=kge_score, data_loader=test_dataloader, use_gpu=True)
    # train the model
    trainer = WCGTrainerKuai16KGP(
        model=model,
        data_loader=train_dataloader,
        train_times=args.epoch,
        alpha=args.learning_rate,
        use_gpu=True,
        opt_method='Adam',
        generator=adv_generator,
        lrg=args.lrg,
        mu=args.mu,
        tester=tester
    )

    trainer.run()
    kge_score.save_checkpoint(args.save)

    # test the model
    kge_score.load_checkpoint(args.save)
    
    tester.run_link_prediction(type_constrain=False)