def run():
# load data
ent2id = read_elements(os.path.join(config.data_path, "entities.dict"))
rel2id = read_elements(os.path.join(config.data_path, "relations.dict"))
ent_num = len(ent2id)
rel_num = len(rel2id)
train_triples = read_triples(os.path.join(config.data_path, "train.txt"),
ent2id, rel2id)
valid_triples = read_triples(os.path.join(config.data_path, "valid.txt"),
ent2id, rel2id)
test_triples = read_triples(os.path.join(config.data_path, "test.txt"),
ent2id, rel2id)
symmetry_test = read_triples(
os.path.join(config.data_path, "symmetry_test.txt"), ent2id, rel2id)
inversion_test = read_triples(
os.path.join(config.data_path, "inversion_test.txt"), ent2id, rel2id)
composition_test = read_triples(
os.path.join(config.data_path, "composition_test.txt"), ent2id, rel2id)
others_test = read_triples(
os.path.join(config.data_path, "other_test.txt"), ent2id, rel2id)
logging.info("#ent_num: %d" % ent_num)
logging.info("#rel_num: %d" % rel_num)
logging.info("#train triple num: %d" % len(train_triples))
logging.info("#valid triple num: %d" % len(valid_triples))
logging.info("#test triple num: %d" % len(test_triples))
logging.info("#Model: %s" % config.model)
# 创建模型
kge_model = TransE(ent_num, rel_num)
if config.model == "TransH":
kge_model = TransH(ent_num, rel_num)
elif config.model == "TransR":
kge_model = SimpleTransR(ent_num, rel_num)
elif config.model == "TransD":
kge_model = TransD(ent_num, rel_num)
elif config.model == "STransE":
kge_model = STransE(ent_num, rel_num)
elif config.model == "LineaRE":
kge_model = LineaRE(ent_num, rel_num)
elif config.model == "DistMult":
kge_model = DistMult(ent_num, rel_num)
elif config.model == "ComplEx":
kge_model = ComplEx(ent_num, rel_num)
elif config.model == "RotatE":
kge_model = RotatE(ent_num, rel_num)
elif config.model == "TransIJ":
kge_model = TransIJ(ent_num, rel_num)
kge_model = kge_model.cuda(torch.device("cuda:0"))
logging.info("Model Parameter Configuration:")
for name, param in kge_model.named_parameters():
logging.info("Parameter %s: %s, require_grad = %s" %
(name, str(param.size()), str(param.requires_grad)))
# 训练
train(model=kge_model,
triples=(train_triples, valid_triples, test_triples, symmetry_test,
inversion_test, composition_test, others_test),
ent_num=ent_num)
def run():
set_logger()
# load data
ent_path = os.path.join(config.data_path, "entities.dict")
rel_path = os.path.join(config.data_path, "relations.dict")
ent2id = read_elements(ent_path)
rel2id = read_elements(rel_path)
ent_num = len(ent2id)
rel_num = len(rel2id)
train_triples = read_triples(os.path.join(config.data_path, "train.txt"),
ent2id, rel2id)
valid_triples = read_triples(os.path.join(config.data_path, "valid.txt"),
ent2id, rel2id)
test_triples = read_triples(os.path.join(config.data_path, "test.txt"),
ent2id, rel2id)
logging.info("#ent_num: %d" % ent_num)
logging.info("#rel_num: %d" % rel_num)
logging.info("#train triple num: %d" % len(train_triples))
logging.info("#valid triple num: %d" % len(valid_triples))
logging.info("#test triple num: %d" % len(test_triples))
logging.info("#Model: %s" % config.model)
# 创建模型
kge_model = TransE(ent_num, rel_num)
if config.model == "TransH":
kge_model = TransH(ent_num, rel_num)
elif config.model == "TransR":
kge_model = TransR(ent_num, rel_num)
elif config.model == "TransD":
kge_model = TransD(ent_num, rel_num)
elif config.model == "STransE":
kge_model = STransE(ent_num, rel_num)
elif config.model == "LineaRE":
kge_model = LineaRE(ent_num, rel_num)
elif config.model == "DistMult":
kge_model = DistMult(ent_num, rel_num)
elif config.model == "ComplEx":
kge_model = ComplEx(ent_num, rel_num)
elif config.model == "RotatE":
kge_model = RotatE(ent_num, rel_num)
if config.cuda:
kge_model = kge_model.cuda()
logging.info("Model Parameter Configuration:")
for name, param in kge_model.named_parameters():
logging.info("Parameter %s: %s, require_grad = %s" %
(name, str(param.size()), str(param.requires_grad)))
# 训练
train(model=kge_model,
triples=(train_triples, valid_triples, test_triples),
ent_num=ent_num)
class Train:
def __init__(self, data_name):
self.dataset = get_dataset(data_name)
self.n_entities = self.dataset.n_entities
self.n_relations = self.dataset.n_relations
def prepareData(self):
print("Perpare dataloader")
self.train = TrainDataset(self.dataset)
self.trainloader = None
self.valid = EvalDataset(self.dataset)
self.validloader = DataLoader(self.valid,
batch_size=self.valid.n_triples,
shuffle=False)
def prepareModel(self):
print("Perpare model")
self.model = TransE(self.n_entities, self.n_relations, embDim=100)
if GPU:
self.model.cuda()
def saveModel(self):
pickle.dump(self.model.get_emb_weights(), open('emb_weight.pkl', 'wb'))
def fit(self):
optim = torch.optim.Adam(self.model.parameters(), lr=LR)
minLoss = float("inf")
bestMR = float("inf")
GlobalEpoch = 0
for seed in range(100):
print(f"# Using seed: {seed}")
self.train.regenerate_neg_samples(seed=seed)
self.trainloader = DataLoader(self.train,
batch_size=1024,
shuffle=True,
num_workers=4)
for epoch in range(EPOCHS_PER_SEED):
GlobalEpoch += 1
for sample in self.trainloader:
if GPU:
pos_triples = torch.LongTensor(
sample['pos_triples']).cuda()
neg_triples = torch.LongTensor(
sample['neg_triples']).cuda()
else:
pos_triples = torch.LongTensor(sample['pos_triples'])
neg_triples = torch.LongTensor(sample['neg_triples'])
self.model.normal_emb()
loss = self.model(pos_triples, neg_triples)
if GPU:
lossVal = loss.cpu().item()
else:
lossVal = loss.item()
optim.zero_grad()
loss.backward()
optim.step()
if minLoss > lossVal:
minLoss = lossVal
MR = Eval_MR(self.validloader, "L2",
**self.model.get_emb_weights())
if MR < bestMR:
bestMR = MR
print('save embedding weight')
self.saveModel()
print(
f"Epoch: {epoch + 1}, Total_Train: {GlobalEpoch}, Loss: {lossVal}, minLoss: {minLoss},"
f"MR: {MR}, bestMR: {bestMR}")
if GlobalEpoch % LR_DECAY_EPOCH == 0:
adjust_learning_rate(optim, 0.96)
def main():
opts = get_train_args()
print("load data ...")
data = DataSet('data/modified_triples.txt')
dataloader = DataLoader(data, shuffle=True, batch_size=opts.batch_size)
print("load model ...")
if opts.model_type == 'transe':
model = TransE(opts, data.ent_tot, data.rel_tot)
elif opts.model_type == "distmult":
model = DistMult(opts, data.ent_tot, data.rel_tot)
if opts.optimizer == 'Adam':
optimizer = optim.Adam(model.parameters(), lr=opts.lr)
elif opts.optimizer == 'SGD':
optimizer = optim.SGD(model.parameters(), lr=opts.lr)
model.cuda()
model.relation_normalize()
loss = torch.nn.MarginRankingLoss(margin=opts.margin)
print("start training")
for epoch in range(1, opts.epochs + 1):
print("epoch : " + str(epoch))
model.train()
epoch_start = time.time()
epoch_loss = 0
tot = 0
cnt = 0
for i, batch_data in enumerate(dataloader):
optimizer.zero_grad()
batch_h, batch_r, batch_t, batch_n = batch_data
batch_h = torch.LongTensor(batch_h).cuda()
batch_r = torch.LongTensor(batch_r).cuda()
batch_t = torch.LongTensor(batch_t).cuda()
batch_n = torch.LongTensor(batch_n).cuda()
pos_score, neg_score, dist = model.forward(batch_h, batch_r,
batch_t, batch_n)
pos_score = pos_score.cpu()
neg_score = neg_score.cpu()
dist = dist.cpu()
train_loss = loss(pos_score, neg_score,
torch.ones(pos_score.size(-1))) + dist
train_loss.backward()
optimizer.step()
batch_loss = torch.sum(train_loss)
epoch_loss += batch_loss
batch_size = batch_h.size(0)
tot += batch_size
cnt += 1
print('\r{:>10} epoch {} progress {} loss: {}\n'.format(
'', epoch, tot / data.__len__(), train_loss),
end='')
end = time.time()
time_used = end - epoch_start
epoch_loss /= cnt
print('one epoch time: {} minutes'.format(time_used / 60))
print('{} epochs'.format(epoch))
print('epoch {} loss: {}'.format(epoch, epoch_loss))
if epoch % opts.save_step == 0:
print("save model...")
model.entity_normalize()
torch.save(model.state_dict(), 'model.pt')
print("save model...")
model.entity_normalize()
torch.save(model.state_dict(), 'model.pt')
print("[Saving embeddings of whole entities & relations...]")
save_embeddings(model, opts, data.id2ent, data.id2rel)
print("[Embedding results are saved successfully.]")