def read_data(path, data_type, args):
X = []
y = []
if data_type == 'deepwalk':
word_vector = KeyedVectors.load_word2vec_format(path, binary=False)
for word in word_vector.vocab:
vector = word_vector[word]
X.append(vector)
y.append(int(word))
if data_type == "transe":
from models.transe import TransE as Model
model = Model.load_model(path)
ent_vocab = Vocab.load(args.ent)
rel_vocab = Vocab.load(args.rel)
valid_rels = np.random.randint(0, len(rel_vocab), args.n_cls)
with open(args.kb) as f:
for line in f:
sub, rel, obj = line.strip().split('\t')
sub_emb = model.pick_ent(ent_vocab[sub])
obj_emb = model.pick_ent(ent_vocab[obj])
rel_id = rel_vocab[rel]
if (args.n_cls >= len(rel_vocab)) or (rel_id in valid_rels):
X.append(sub_emb - obj_emb)
y.append(rel_id)
else:
word_vector = KeyedVectors.load_word2vec_format(path, binary=False)
ent_vocab = Vocab.load(args.ent)
rel_vocab = Vocab.load(args.rel)
def pick_emb(eid):
if str(eid) in word_vector:
return word_vector[str(eid)]
else:
return np.array([0.0 for i in range(len(word_vector[0]))])
with open(args.kb) as f:
for line in f:
sub, rel, obj = line.strip().split('\t')
sub_emb = pick_emb(ent_vocab[sub])
obj_emb = pick_emb(ent_vocab[obj])
rel_id = rel_vocab[rel]
valid_rels = np.random.randint(0, len(rel_vocab), args.n_cls)
if (args.n_cls >= len(rel_vocab)) or (rel_id in valid_rels):
X.append(sub_emb - obj_emb)
y.append(rel_id)
return np.array(X), np.array(y)
def test(args):
ent_vocab = Vocab.load(args.ent)
rel_vocab = Vocab.load(args.rel)
# preparing data
if args.task == 'kbc':
test_dat = TripletDataset.load(args.data, ent_vocab, rel_vocab)
elif args.task == 'tc':
test_dat = LabeledTripletDataset.load(args.data, ent_vocab, rel_vocab)
else:
raise ValueError('Invalid task: {}'.format(args.task))
print('loading model...')
if args.method == 'transe':
from models.transe import TransE as Model
elif args.method == 'complex':
from models.complex import ComplEx as Model
elif args.method == 'analogy':
from models.analogy import ANALOGY as Model
else:
raise NotImplementedError
if args.filtered:
print('loading whole graph...')
from utils.graph import TensorTypeGraph
graphall = TensorTypeGraph.load_from_raw(args.graphall, ent_vocab,
rel_vocab)
# graphall = TensorTypeGraph.load(args.graphall)
else:
graphall = None
model = Model.load_model(args.model)
if args.metric == 'all':
evaluator = Evaluator('all', None, args.filtered, False, graphall)
if args.filtered:
evaluator.prepare_valid(test_dat)
all_res = evaluator.run_all_matric(model, test_dat)
for metric in sorted(all_res.keys()):
print('{:20s}: {}'.format(metric, all_res[metric]))
else:
evaluator = Evaluator(args.metric, None, False, True, None)
res = evaluator.run(model, test_dat)
print('{:20s}: {}'.format(args.metric, res))
def train(args):
# setting for logging
if not os.path.exists(args.log):
os.mkdir(args.log)
logger = logging.getLogger()
logging.basicConfig(level=logging.INFO)
log_path = os.path.join(args.log, 'log')
file_handler = logging.FileHandler(log_path)
fmt = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
file_handler.setFormatter(fmt)
logger.addHandler(file_handler)
# TODO: develop the recording of arguments in logging
logger.info('Arguments...')
for arg, val in sorted(vars(args).items()):
logger.info('{:>10} -----> {}'.format(arg, val))
ent_vocab = Vocab.load(args.ent)
rel_vocab = Vocab.load(args.rel)
n_entity, n_relation = len(ent_vocab), len(rel_vocab)
# preparing data
logger.info('preparing data...')
train_dat = TripletDataset.load(args.train, ent_vocab, rel_vocab)
valid_dat = TripletDataset.load(args.valid, ent_vocab, rel_vocab) if args.valid else None
if args.filtered:
logger.info('loading whole graph...')
from utils.graph import TensorTypeGraph
whole_graph = TensorTypeGraph.load_from_raw(args.graphall, ent_vocab, rel_vocab)
else:
whole_graph = None
if args.opt == 'sgd':
opt = SGD(args.lr)
elif args.opt == 'adagrad':
opt = Adagrad(args.lr)
elif args.opt == 'dsgd':
opt = DecaySGD(args.lr)
else:
raise NotImplementedError
if args.l2_reg > 0:
opt.set_l2_reg(args.l2_reg)
if args.gradclip > 0:
opt.set_gradclip(args.gradclip)
logger.info('building model...')
if args.method == 'complex':
from models.complex import ComplEx
model = ComplEx(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'distmult':
from models.distmult import DistMult
model = DistMult(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'transe':
from models.transe import TransE
model = TransE(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'hole':
from models.hole import HolE
model = HolE(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'rescal':
from models.rescal import RESCAL
model = RESCAL(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'analogy':
from models.analogy import ANALOGY
model = ANALOGY(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
cp_ratio=args.cp_ratio,
mode=args.mode)
elif args.method == 'transe_set':
from models.transe_set import TransE_set
model = TransE_set(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'line':
from models.line_model import LineModel
model = LineModel(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
else:
raise NotImplementedError
# evaluator = Evaluator(args.metric, args.nbest, args.filtered, whole_graph) if args.valid or args.synthetic else None
evaluator = Evaluator(args.metric, args.nbest, args.filtered, whole_graph) if args.valid else None
# delete args.synthetic to run
if args.filtered and args.valid:
evaluator.prepare_valid(valid_dat)
if args.mode == 'pairwise':
trainer = PairwiseTrainer(model=model, opt=opt, save_step=args.save_step,
batchsize=args.batch, logger=logger,
evaluator=evaluator, valid_dat=valid_dat,
n_negative=args.negative, epoch=args.epoch,
model_dir=args.log)
elif args.mode == 'single':
trainer = SingleTrainer(model=model, opt=opt, save_step=args.save_step,
batchsize=args.batch, logger=logger,
evaluator=evaluator, valid_dat=valid_dat,
n_negative=args.negative, epoch=args.epoch,
model_dir=args.log)
else:
raise NotImplementedError
trainer.fit(train_dat)
logger.info('done all')
def train(args):
# setting for logging
if not os.path.exists(args.log):
os.mkdir(args.log)
logger = logging.getLogger()
logging.basicConfig(level=logging.INFO)
log_path = os.path.join(args.log, 'log')
file_handler = logging.FileHandler(log_path)
fmt = logging.Formatter('%(asctime)s %(levelname)s %(message)s')
file_handler.setFormatter(fmt)
logger.addHandler(file_handler)
# TODO: develop the recording of arguments in logging
logger.info('Arguments...')
for arg, val in vars(args).items():
logger.info('{:>10} -----> {}'.format(arg, val))
ent_vocab = Vocab.load(args.ent)
rel_vocab = Vocab.load(args.rel)
n_entity, n_relation = len(ent_vocab), len(rel_vocab)
# preparing data
logger.info('preparing data...')
train_dat = TripletDataset.load(args.train, ent_vocab, rel_vocab)
valid_dat = TripletDataset.load(args.valid, ent_vocab,
rel_vocab) if args.valid else None
if args.filtered:
logger.info('loading whole graph...')
from utils.graph import TensorTypeGraph
whole_graph = TensorTypeGraph.load_from_raw(args.graphall, ent_vocab,
rel_vocab)
else:
whole_graph = None
if args.opt == 'sgd':
opt = SGD(args.lr)
elif args.opt == 'adagrad':
opt = Adagrad(args.lr)
else:
raise NotImplementedError
if args.l2_reg > 0:
opt.set_l2_reg(args.l2_reg)
if args.gradclip > 0:
opt.set_gradclip(args.gradclip)
logger.info('building model...')
if args.method == 'complex':
from models.complex import ComplEx
model = ComplEx(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'distmult':
from models.distmult import DistMult
model = DistMult(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'transe':
from models.transe import TransE
model = TransE(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'hole':
from models.hole import HolE
model = HolE(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'rescal':
from models.rescal import RESCAL
model = RESCAL(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
mode=args.mode)
elif args.method == 'analogy':
from models.analogy import ANALOGY
model = ANALOGY(n_entity=n_entity,
n_relation=n_relation,
margin=args.margin,
dim=args.dim,
cp_ratio=args.cp_ratio,
mode=args.mode)
elif args.method == 'randwalk':
from models.randwalk import RandWalk
logger.info(
'using random walk model to learning embedding unsupervisedly.')
model = RandWalk(n_entity=n_entity,
n_relation=n_relation,
knowledge_path=args.train,
ent_vocab=ent_vocab,
rel_vocab=rel_vocab,
dim=args.dim,
output=args.log)
model.train()
model.save_model(os.path.join(args.log, model.__class__.__name__))
return
elif args.method == "lr":
from models.lr import LogisticReg
model = LogisticReg(n_entity=n_entity,
n_relation=n_relation,
train_path=args.train,
ent_vocab=ent_vocab,
rel_vocab=rel_vocab,
dim=args.dim,
output=args.log,
wv_model_path=args.wv_model,
negative=args.negative,
feat_type=args.feat_type)
starttime = time()
if args.mode == "triplet_cls":
logger.info("Training a triple classifer")
model.train_triple_classifer()
else:
model.train()
endtime = time()
logger.info("lr model train time {:.6f}".format(endtime - starttime))
model.save_model(os.path.join(args.log, model.__class__.__name__))
return
else:
raise NotImplementedError
evaluator = Evaluator(
args.metric, args.nbest, args.filtered,
whole_graph) if args.valid or args.synthetic else None
if args.filtered and args.valid:
evaluator.prepare_valid(valid_dat)
if args.mode == 'pairwise':
trainer = PairwiseTrainer(model=model,
opt=opt,
save_step=args.save_step,
batchsize=args.batch,
logger=logger,
evaluator=evaluator,
valid_dat=valid_dat,
n_negative=args.negative,
epoch=args.epoch,
model_dir=args.log)
elif args.mode == 'single':
trainer = SingleTrainer(model=model,
opt=opt,
save_step=args.save_step,
batchsize=args.batch,
logger=logger,
evaluator=evaluator,
valid_dat=valid_dat,
n_negative=args.negative,
epoch=args.epoch,
model_dir=args.log)
else:
raise NotImplementedError
trainer.fit(train_dat)
logger.info('done all')