def __init__(self, chpt_path, max_length=128):
"""
Initializer
"""
self.checkpoint_path = chpt_path
self.bert_pretrained_checkpoint = 'bert-base-uncased'
self.max_length = max_length
self.sentence_encoder = BERTPAIRSentenceEncoder(
self.bert_pretrained_checkpoint, self.max_length)
self.model = Pair(self.sentence_encoder, hidden_size=768)
if torch.cuda.is_available():
self.model = self.model.cuda()
self.model.eval()
# self.nlp_coref = spacy.load("en_core_web_sm")
# neuralcoref.add_to_pipe(self.nlp_coref)
self.nlp_no_coref = spacy.load("en_core_web_sm")
self.load_model()
def main():
opt = {'train': 'train_wiki',
'val': 'val_wiki',
'test': 'val_semeval',
'adv': None,
'trainN': 10,
'N': 5,
'K': 1,
'Q': 1,
'batch_size': 1,
'train_iter': 5,
'val_iter': 1000,
'test_iter': 10,
'val_step': 2000,
'model': 'pair',
'encoder': 'bert',
'max_length': 64,
'lr': -1,
'weight_decay': 1e-5,
'dropout': 0.0,
'na_rate':0,
'optim': 'adam',
'load_ckpt': './src/FewRel-master/Checkpoints/ckpt_5-Way-1-Shot_FewRel.pth',
'save_ckpt': './src/FewRel-master/Checkpoints/post_ckpt_5-Way-1-Shot_FewRel.pth',
'fp16':False,
'only_test': False,
'ckpt_name': 'Checkpoints/ckpt_5-Way-1-Shot_FewRel.pth',
'pair': True,
'pretrain_ckpt': '',
'cat_entity_rep': False,
'dot': False,
'no_dropout': False,
'mask_entity': False,
'use_sgd_for_bert': False
}
opt = DotMap(opt)
trainN = opt.trainN
N = opt.N
K = opt.K
Q = opt.Q
batch_size = opt.batch_size
model_name = opt.model
encoder_name = opt.encoder
max_length = opt.max_length
print("{}-way-{}-shot Few-Shot Relation Classification".format(N, K))
print("model: {}".format(model_name))
print("encoder: {}".format(encoder_name))
print("max_length: {}".format(max_length))
pretrain_ckpt = opt.pretrain_ckpt or 'bert-base-uncased'
sentence_encoder = BERTPAIRSentenceEncoder(
pretrain_ckpt,
max_length)
train_data_loader = get_loader_pair(opt.train, sentence_encoder,
N=trainN, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size, encoder_name=encoder_name)
val_data_loader = get_loader_pair(opt.val, sentence_encoder,
N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size, encoder_name=encoder_name)
test_data_loader = get_loader_pair(opt.test, sentence_encoder,
N=N, K=K, Q=Q, na_rate=opt.na_rate, batch_size=batch_size, encoder_name=encoder_name)
if opt.optim == 'sgd':
pytorch_optim = optim.SGD
elif opt.optim == 'adam':
pytorch_optim = optim.Adam
elif opt.optim == 'adamw':
pytorch_optim = AdamW
else:
raise NotImplementedError
framework = FewShotREFramework(train_data_loader, val_data_loader, test_data_loader)
prefix = '-'.join([model_name, encoder_name, opt.train, opt.val, str(N), str(K)])
if opt.na_rate != 0:
prefix += '-na{}'.format(opt.na_rate)
if opt.dot:
prefix += '-dot'
if opt.cat_entity_rep:
prefix += '-catentity'
if len(opt.ckpt_name) > 0:
prefix += '-' + opt.ckpt_name
model = Pair(sentence_encoder, hidden_size=opt.hidden_size)
if not os.path.exists('checkpoint'):
os.mkdir('checkpoint')
ckpt = 'checkpoint/{}.pth.tar'.format(prefix)
if opt.save_ckpt:
ckpt = opt.save_ckpt
if torch.cuda.is_available():
model.cuda()
if not opt.only_test:
if encoder_name in ['bert', 'roberta']:
bert_optim = True
else:
bert_optim = False
if opt.lr == -1:
if bert_optim:
opt.lr = 2e-5
else:
opt.lr = 1e-1
framework.train(model, prefix, batch_size, trainN, N, K, Q,
pytorch_optim=pytorch_optim, load_ckpt=opt.load_ckpt, save_ckpt=ckpt,
na_rate=opt.na_rate, val_step=opt.val_step, fp16=opt.fp16, pair=opt.pair,
train_iter=opt.train_iter, val_iter=opt.val_iter, bert_optim=bert_optim,
learning_rate=opt.lr, use_sgd_for_bert=opt.use_sgd_for_bert)
else:
ckpt = opt.load_ckpt
if ckpt is None:
print("Warning: --load_ckpt is not specified. Will load Huggingface pre-trained checkpoint.")
ckpt = 'none'
acc = framework.eval(model, batch_size, N, K, Q, opt.test_iter, na_rate=opt.na_rate, ckpt=ckpt, pair=opt.pair)
print("RESULT: %.2f" % (acc * 100))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--train', default='train_wiki', help='train file')
parser.add_argument('--val', default='val_wiki', help='val file')
parser.add_argument('--test', default='test_wiki', help='test file')
parser.add_argument('--adv', default=None, help='adv file')
parser.add_argument('--trainN', default=10, type=int, help='N in train')
parser.add_argument('--N', default=5, type=int, help='N way')
parser.add_argument('--K', default=5, type=int, help='K shot')
parser.add_argument('--Q',
default=5,
type=int,
help='Num of query per class')
parser.add_argument('--batch_size', default=4, type=int, help='batch size')
parser.add_argument('--train_iter',
default=30000,
type=int,
help='num of iters in training')
parser.add_argument('--val_iter',
default=1000,
type=int,
help='num of iters in validation')
parser.add_argument('--test_iter',
default=10000,
type=int,
help='num of iters in testing')
parser.add_argument('--val_step',
default=2000,
type=int,
help='val after training how many iters')
parser.add_argument('--model', default='proto', help='model name')
parser.add_argument('--encoder',
default='cnn',
help='encoder: cnn or bert or roberta')
parser.add_argument('--max_length',
default=128,
type=int,
help='max length')
parser.add_argument('--lr', default=1e-1, type=float, help='learning rate')
parser.add_argument('--weight_decay',
default=1e-5,
type=float,
help='weight decay')
parser.add_argument('--dropout',
default=0.0,
type=float,
help='dropout rate')
parser.add_argument('--na_rate',
default=0,
type=int,
help='NA rate (NA = Q * na_rate)')
parser.add_argument('--grad_iter',
default=1,
type=int,
help='accumulate gradient every x iterations')
parser.add_argument('--optim', default='sgd', help='sgd / adam / adamw')
parser.add_argument('--hidden_size',
default=230,
type=int,
help='hidden size')
parser.add_argument('--load_ckpt', default=None, help='load ckpt')
parser.add_argument('--save_ckpt', default=None, help='save ckpt')
parser.add_argument('--fp16',
action='store_true',
help='use nvidia apex fp16')
parser.add_argument('--only_test', action='store_true', help='only test')
# only for bert / roberta
parser.add_argument('--pair', action='store_true', help='use pair model')
parser.add_argument('--pretrain_ckpt',
default=None,
help='bert / roberta pre-trained checkpoint')
parser.add_argument(
'--cat_entity_rep',
action='store_true',
help='concatenate entity representation as sentence rep')
# only for prototypical networks
parser.add_argument('--dot',
action='store_true',
help='use dot instead of L2 distance for proto')
opt = parser.parse_args()
trainN = opt.trainN
N = opt.N
K = opt.K
Q = opt.Q
batch_size = opt.batch_size
model_name = opt.model
encoder_name = opt.encoder
max_length = opt.max_length
print("{}-way-{}-shot Few-Shot Relation Classification".format(N, K))
print("model: {}".format(model_name))
print("encoder: {}".format(encoder_name))
print("max_length: {}".format(max_length))
if encoder_name == 'cnn':
try:
glove_mat = np.load('./pretrain/glove/glove_mat.npy')
glove_word2id = json.load(
open('./pretrain/glove/glove_word2id.json'))
except:
raise Exception(
"Cannot find glove files. Run glove/download_glove.sh to download glove files."
)
sentence_encoder = CNNSentenceEncoder(glove_mat, glove_word2id,
max_length)
elif encoder_name == 'bert':
pretrain_ckpt = opt.pretrain_ckpt or 'bert-base-uncased'
if opt.pair:
sentence_encoder = BERTPAIRSentenceEncoder(pretrain_ckpt,
max_length)
else:
sentence_encoder = BERTSentenceEncoder(
pretrain_ckpt, max_length, cat_entity_rep=opt.cat_entity_rep)
elif encoder_name == 'roberta':
pretrain_ckpt = opt.pretrain_ckpt or 'roberta-base'
if opt.pair:
sentence_encoder = RobertaPAIRSentenceEncoder(
pretrain_ckpt, max_length)
else:
sentence_encoder = RobertaSentenceEncoder(
pretrain_ckpt, max_length, cat_entity_rep=opt.cat_entity_rep)
else:
raise NotImplementedError
if opt.pair:
train_data_loader = get_loader_pair(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size,
encoder_name=encoder_name)
val_data_loader = get_loader_pair(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size,
encoder_name=encoder_name)
test_data_loader = get_loader_pair(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size,
encoder_name=encoder_name)
else:
train_data_loader = get_loader(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
val_data_loader = get_loader(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
test_data_loader = get_loader(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.adv:
adv_data_loader = get_loader_unsupervised(opt.adv,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.optim == 'sgd':
pytorch_optim = optim.SGD
elif opt.optim == 'adam':
pytorch_optim = optim.Adam
elif opt.optim == 'adamw':
from transformers import AdamW
pytorch_optim = AdamW
else:
raise NotImplementedError
if opt.adv:
d = Discriminator(opt.hidden_size)
framework = FewShotREFramework(train_data_loader,
val_data_loader,
test_data_loader,
adv_data_loader,
adv=opt.adv,
d=d)
else:
framework = FewShotREFramework(train_data_loader, val_data_loader,
test_data_loader)
prefix = '-'.join(
[model_name, encoder_name, opt.train, opt.val,
str(N), str(K)])
if opt.adv is not None:
prefix += '-adv_' + opt.adv
if opt.na_rate != 0:
prefix += '-na{}'.format(opt.na_rate)
if opt.dot:
prefix += '-dot'
if opt.cat_entity_rep:
prefix += '-catentity'
if model_name == 'proto':
model = Proto(sentence_encoder, dot=opt.dot)
elif model_name == 'gnn':
model = GNN(sentence_encoder, N, hidden_size=opt.hidden_size)
elif model_name == 'snail':
model = SNAIL(sentence_encoder, N, K, hidden_size=opt.hidden_size)
elif model_name == 'metanet':
model = MetaNet(N, K, sentence_encoder.embedding, max_length)
elif model_name == 'siamese':
model = Siamese(sentence_encoder,
hidden_size=opt.hidden_size,
dropout=opt.dropout)
elif model_name == 'pair':
model = Pair(sentence_encoder, hidden_size=opt.hidden_size)
else:
raise NotImplementedError
if not os.path.exists('checkpoint'):
os.mkdir('checkpoint')
ckpt = 'checkpoint/{}.pth.tar'.format(prefix)
if opt.save_ckpt:
ckpt = opt.save_ckpt
if torch.cuda.is_available():
model.cuda()
if not opt.only_test:
if encoder_name in ['bert', 'roberta']:
bert_optim = True
else:
bert_optim = False
framework.train(model,
prefix,
batch_size,
trainN,
N,
K,
Q,
pytorch_optim=pytorch_optim,
load_ckpt=opt.load_ckpt,
save_ckpt=ckpt,
na_rate=opt.na_rate,
val_step=opt.val_step,
fp16=opt.fp16,
pair=opt.pair,
train_iter=opt.train_iter,
val_iter=opt.val_iter,
bert_optim=bert_optim)
else:
ckpt = opt.load_ckpt
acc = framework.eval(model,
batch_size,
N,
K,
Q,
opt.test_iter,
na_rate=opt.na_rate,
ckpt=ckpt,
pair=opt.pair)
print("RESULT: %.2f" % (acc * 100))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--train', default='train_wiki', help='train file')
parser.add_argument('--val', default='val_wiki', help='val file')
parser.add_argument('--test', default='test_wiki', help='test file')
parser.add_argument('--adv', default=None, help='adv file')
parser.add_argument('--trainN', default=10, type=int, help='N in train')
parser.add_argument('--N', default=5, type=int, help='N way')
parser.add_argument('--K', default=5, type=int, help='K shot')
parser.add_argument('--Q',
default=5,
type=int,
help='Num of query per class')
parser.add_argument('--batch_size', default=4, type=int, help='batch size')
parser.add_argument('--train_iter',
default=20000,
type=int,
help='num of iters in training')
parser.add_argument('--val_iter',
default=1000,
type=int,
help='num of iters in validation')
parser.add_argument('--test_iter',
default=2000,
type=int,
help='num of iters in testing')
parser.add_argument('--val_step',
default=2000,
type=int,
help='val after training how many iters')
parser.add_argument('--model', default='proto', help='model name')
parser.add_argument('--encoder',
default='cnn',
help='encoder: cnn or bert')
parser.add_argument('--max_length',
default=128,
type=int,
help='max length')
parser.add_argument('--lr', default=1e-1, type=float, help='learning rate')
parser.add_argument('--weight_decay',
default=1e-5,
type=float,
help='weight decay')
parser.add_argument('--dropout',
default=0.0,
type=float,
help='dropout rate')
parser.add_argument('--na_rate',
default=0,
type=int,
help='NA rate (NA = Q * na_rate)')
parser.add_argument('--grad_iter',
default=1,
type=int,
help='accumulate gradient every x iterations')
parser.add_argument('--optim',
default='sgd',
help='sgd / adam / bert_adam')
parser.add_argument('--hidden_size',
default=230,
type=int,
help='hidden size')
parser.add_argument('--load_ckpt', default=None, help='load ckpt')
parser.add_argument('--save_ckpt', default=None, help='save ckpt')
parser.add_argument('--fp16',
action='store_true',
help='use nvidia apex fp16')
parser.add_argument('--only_test', action='store_true', help='only test')
parser.add_argument('--pair', action='store_true', help='use pair model')
parser.add_argument('--language', type=str, default='eng', help='language')
parser.add_argument('--sup_cost',
type=int,
default=0,
help='use sup classifier')
opt = parser.parse_args()
trainN = opt.trainN
N = opt.N
K = opt.K
Q = opt.Q
batch_size = opt.batch_size
model_name = opt.model
encoder_name = opt.encoder
max_length = opt.max_length
sup_cost = bool(opt.sup_cost)
print(sup_cost)
print("{}-way-{}-shot Few-Shot Relation Classification".format(N, K))
print("model: {}".format(model_name))
print("encoder: {}".format(encoder_name))
print("max_length: {}".format(max_length))
embsize = 50
if opt.language == 'chn':
embsize = 100
if encoder_name == 'cnn':
try:
if opt.language == 'chn':
glove_mat = np.load('./pretrain/chinese_emb/emb.npy')
glove_word2id = json.load(
open('./pretrain/chinese_emb/word2id.json'))
else:
glove_mat = np.load('./pretrain/glove/glove_mat.npy')
glove_word2id = json.load(
open('./pretrain/glove/glove_word2id.json'))
except:
raise Exception(
"Cannot find glove files. Run glove/download_glove.sh to download glove files."
)
sentence_encoder = CNNSentenceEncoder(glove_mat,
glove_word2id,
max_length,
word_embedding_dim=embsize)
elif encoder_name == 'bert':
if opt.pair:
if opt.language == 'chn':
sentence_encoder = BERTPAIRSentenceEncoder(
'bert-base-chinese', #'./pretrain/bert-base-uncased',
max_length)
else:
sentence_encoder = BERTPAIRSentenceEncoder(
'bert-base-uncased', max_length)
else:
if opt.language == 'chn':
sentence_encoder = BERTSentenceEncoder(
'bert-base-chinese', #'./pretrain/bert-base-uncased',
max_length)
else:
sentence_encoder = BERTSentenceEncoder('bert-base-uncased',
max_length)
else:
raise NotImplementedError
if opt.pair:
train_data_loader = get_loader_pair(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
val_data_loader = get_loader_pair(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
test_data_loader = get_loader_pair(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
else:
train_data_loader = get_loader(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
val_data_loader = get_loader(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
test_data_loader = get_loader(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.adv:
adv_data_loader = get_loader_unsupervised(opt.adv,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.optim == 'sgd':
pytorch_optim = optim.SGD
elif opt.optim == 'adam':
pytorch_optim = optim.Adam
elif opt.optim == 'bert_adam':
from transformers import AdamW
pytorch_optim = AdamW
else:
raise NotImplementedError
if opt.adv:
d = Discriminator(opt.hidden_size)
framework = FewShotREFramework(train_data_loader,
val_data_loader,
test_data_loader,
adv_data_loader,
adv=opt.adv,
d=d)
else:
framework = FewShotREFramework(train_data_loader, val_data_loader,
test_data_loader)
prefix = '-'.join(
[model_name, encoder_name, opt.train, opt.val,
str(N), str(K)])
if opt.adv is not None:
prefix += '-adv_' + opt.adv
if opt.na_rate != 0:
prefix += '-na{}'.format(opt.na_rate)
if model_name == 'proto':
model = Proto(sentence_encoder, hidden_size=opt.hidden_size)
elif model_name == 'gnn':
model = GNN(sentence_encoder, N, use_sup_cost=sup_cost)
elif model_name == 'snail':
print("HINT: SNAIL works only in PyTorch 0.3.1")
model = SNAIL(sentence_encoder, N, K)
elif model_name == 'metanet':
model = MetaNet(N,
K,
sentence_encoder.embedding,
max_length,
use_sup_cost=sup_cost)
elif model_name == 'siamese':
model = Siamese(sentence_encoder,
hidden_size=opt.hidden_size,
dropout=opt.dropout)
elif model_name == 'pair':
model = Pair(sentence_encoder, hidden_size=opt.hidden_size)
else:
raise NotImplementedError
if not os.path.exists('checkpoint'):
os.mkdir('checkpoint')
ckpt = 'checkpoint/{}.pth.tar'.format(prefix)
if opt.save_ckpt:
ckpt = opt.save_ckpt
if torch.cuda.is_available():
model.cuda()
if not opt.only_test:
if encoder_name == 'bert':
bert_optim = True
else:
bert_optim = False
framework.train(model,
prefix,
batch_size,
trainN,
N,
K,
Q,
pytorch_optim=pytorch_optim,
load_ckpt=opt.load_ckpt,
save_ckpt=ckpt,
na_rate=opt.na_rate,
val_step=opt.val_step,
fp16=opt.fp16,
pair=opt.pair,
train_iter=opt.train_iter,
val_iter=opt.val_iter,
bert_optim=bert_optim,
sup_cls=sup_cost)
else:
ckpt = opt.load_ckpt
acc = framework.eval(model,
batch_size,
N,
K,
Q,
opt.test_iter,
na_rate=opt.na_rate,
ckpt=ckpt,
pair=opt.pair)
wfile = open('logs/' + ckpt.replace('checkpoint/', '') + '.txt', 'a')
wfile.write(str(N) + '\t' + str(K) + '\t' + str(acc * 100) + '\n')
wfile.close()
print("RESULT: %.2f" % (acc * 100))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--train', default='train_wiki', help='train file')
parser.add_argument('--val', default='val_wiki', help='val file')
parser.add_argument('--test', default='test_wiki', help='test file')
parser.add_argument('--adv', default=None, help='adv file')
parser.add_argument('--trainN', default=10, type=int, help='N in train')
parser.add_argument('--N', default=5, type=int, help='N way')
parser.add_argument('--K', default=5, type=int, help='K shot')
parser.add_argument('--Q',
default=5,
type=int,
help='Num of query per class')
parser.add_argument('--batch_size', default=4, type=int, help='batch size')
parser.add_argument('--train_iter',
default=30000,
type=int,
help='num of iters in training')
parser.add_argument('--val_iter',
default=1000,
type=int,
help='num of iters in validation')
parser.add_argument('--test_iter',
default=10000,
type=int,
help='num of iters in testing')
parser.add_argument('--val_step',
default=2000,
type=int,
help='val after training how many iters')
parser.add_argument('--model', default='regrab', help='model name')
parser.add_argument('--encoder',
default='bert',
help='encoder: cnn or bert')
parser.add_argument('--max_length',
default=128,
type=int,
help='max length')
parser.add_argument('--lr', default=1e-1, type=float, help='learning rate')
parser.add_argument('--weight_decay',
default=1e-5,
type=float,
help='weight decay')
parser.add_argument('--dropout',
default=0.0,
type=float,
help='dropout rate')
parser.add_argument('--na_rate',
default=0,
type=int,
help='NA rate (NA = Q * na_rate)')
parser.add_argument('--grad_iter',
default=1,
type=int,
help='accumulate gradient every x iterations')
parser.add_argument('--optim',
default='sgd',
help='sgd / adam / bert_adam')
parser.add_argument('--hidden_size',
default=230,
type=int,
help='hidden size')
parser.add_argument('--load_ckpt', default=None, help='load ckpt')
parser.add_argument('--save_ckpt', default=None, help='save ckpt')
parser.add_argument('--fp16',
action='store_true',
help='use nvidia apex fp16')
parser.add_argument('--only_test', action='store_true', help='only test')
parser.add_argument('--pair', action='store_true', help='use pair model')
parser.add_argument('--eps',
default=0.1,
type=float,
help='step size for SG-MCMC')
parser.add_argument('--temp',
default=10.0,
type=float,
help='temperature for softmax')
parser.add_argument('--step',
default=5,
type=int,
help='steps for SG-MCMC')
parser.add_argument('--smp',
default=10,
type=int,
help='samples for SG-MCMC')
parser.add_argument('--ratio',
default=0.01,
type=float,
help='decay ratio of step size for SG-MCMC')
parser.add_argument('--wtp',
default=0.1,
type=float,
help='weight of the prior term')
parser.add_argument('--wtn',
default=1.0,
type=float,
help='weight of the noise term')
parser.add_argument('--wtb',
default=0.0,
type=float,
help='weight of the background term')
parser.add_argument('--metric',
default='dot',
help='similarity metric (dot or l2)')
parser.add_argument('--seed', default=1234, type=int, help='random seed')
opt = parser.parse_args()
trainN = opt.trainN
N = opt.N
K = opt.K
Q = opt.Q
batch_size = opt.batch_size
model_name = opt.model
encoder_name = opt.encoder
max_length = opt.max_length
torch.manual_seed(opt.seed)
np.random.seed(opt.seed)
random.seed(opt.seed)
torch.cuda.manual_seed(opt.seed)
print("{}-way-{}-shot Few-Shot Relation Classification".format(N, K))
print("model: {}".format(model_name))
print("encoder: {}".format(encoder_name))
print("max_length: {}".format(max_length))
if encoder_name == 'cnn':
try:
glove_mat = np.load('./pretrain/glove/glove_mat.npy')
glove_word2id = json.load(
open('./pretrain/glove/glove_word2id.json'))
except:
raise Exception(
"Cannot find glove files. Run glove/download_glove.sh to download glove files."
)
sentence_encoder = CNNSentenceEncoder(glove_mat, glove_word2id,
max_length)
elif encoder_name == 'bert':
if opt.pair:
sentence_encoder = BERTPAIRSentenceEncoder(
'./pretrain/bert-base-uncased', max_length)
else:
sentence_encoder = BERTSentenceEncoder(
'./pretrain/bert-base-uncased', max_length)
else:
raise NotImplementedError
if opt.pair:
train_data_loader = get_loader_pair(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
val_data_loader = get_loader_pair(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
test_data_loader = get_loader_pair(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
else:
train_data_loader = get_loader(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
val_data_loader = get_loader(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
test_data_loader = get_loader(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.adv:
adv_data_loader = get_loader_unsupervised(opt.adv,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.optim == 'sgd':
pytorch_optim = optim.SGD
elif opt.optim == 'adam':
pytorch_optim = optim.Adam
elif opt.optim == 'bert_adam':
from transformers import AdamW
pytorch_optim = AdamW
else:
raise NotImplementedError
if opt.adv:
d = Discriminator(opt.hidden_size)
framework = FewShotREFramework(train_data_loader,
val_data_loader,
test_data_loader,
adv_data_loader,
adv=opt.adv,
d=d)
else:
framework = FewShotREFramework(train_data_loader, val_data_loader,
test_data_loader)
prefix = '-'.join(
[model_name, encoder_name, opt.train, opt.val,
str(N), str(K)])
if opt.adv is not None:
prefix += '-adv_' + opt.adv
if opt.na_rate != 0:
prefix += '-na{}'.format(opt.na_rate)
# Loading relation embeddings.
data = pickle.load(open('./data/embeddings.pkl', 'rb'))
rellist = data['relations']
relemb = data['embeddings']
array0 = np.zeros((1, relemb.shape[1]), dtype=relemb.dtype)
relemb = np.concatenate([array0, relemb], axis=0)
rel2id = dict([(rel, k + 1) for k, rel in enumerate(rellist)])
# Loading relation graphs.
with open('./data/graph.txt', 'r') as fi:
us, vs, ws = [], [], []
for line in fi:
items = line.strip().split('\t')
us += [rel2id[items[0]]]
vs += [rel2id[items[1]]]
ws += [float(items[2])]
index = torch.LongTensor([us, vs])
value = torch.Tensor(ws)
shape = torch.Size([len(rel2id) + 1, len(rel2id) + 1])
reladj = torch.sparse.FloatTensor(index, value, shape).cuda()
# End
model = REGRAB(sentence_encoder,
hidden_size=opt.hidden_size,
eps=opt.eps,
temp=opt.temp,
step=opt.step,
smp=opt.smp,
ratio=opt.ratio,
wtp=opt.wtp,
wtn=opt.wtn,
wtb=opt.wtb,
metric=opt.metric)
model.set_relemb(rel2id, relemb)
model.set_reladj(reladj)
if not os.path.exists('checkpoint'):
os.mkdir('checkpoint')
ckpt = 'checkpoint/{}.pth.tar'.format(prefix)
if opt.save_ckpt:
ckpt = opt.save_ckpt
if torch.cuda.is_available():
model.cuda()
if not opt.only_test:
if encoder_name == 'bert':
bert_optim = True
else:
bert_optim = False
framework.train(model,
prefix,
batch_size,
trainN,
N,
K,
Q,
pytorch_optim=pytorch_optim,
load_ckpt=opt.load_ckpt,
save_ckpt=ckpt,
na_rate=opt.na_rate,
val_step=opt.val_step,
fp16=opt.fp16,
pair=opt.pair,
train_iter=opt.train_iter,
val_iter=opt.val_iter,
bert_optim=bert_optim)
else:
ckpt = opt.load_ckpt
acc = framework.eval(model,
batch_size,
N,
K,
Q,
opt.test_iter,
na_rate=opt.na_rate,
ckpt=ckpt,
pair=opt.pair)
print("RESULT: %.2f" % (acc * 100))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--train', default='train_wiki', help='train file')
parser.add_argument('--val', default='val_wiki', help='val file')
parser.add_argument('--test', default='test_wiki', help='test file')
parser.add_argument('--adv', default=None, help='adv file')
parser.add_argument('--trainN', default=10, type=int, help='N in train')
parser.add_argument('--N', default=5, type=int, help='N way')
parser.add_argument('--K', default=5, type=int, help='K shot')
parser.add_argument('--Q',
default=5,
type=int,
help='Num of query per class')
parser.add_argument('--batch_size', default=4, type=int, help='batch size')
parser.add_argument('--train_iter',
default=30000,
type=int,
help='num of iters in training')
parser.add_argument('--val_iter',
default=1000,
type=int,
help='num of iters in validation')
parser.add_argument('--test_iter',
default=3000,
type=int,
help='num of iters in testing')
parser.add_argument('--val_step',
default=2000,
type=int,
help='val after training how many iters')
parser.add_argument('--model', default='proto', help='model name')
parser.add_argument('--encoder',
default='cnn',
help='encoder: cnn or bert')
parser.add_argument('--max_length',
default=128,
type=int,
help='max length')
parser.add_argument('--lr', default=1e-1, type=float, help='learning rate')
parser.add_argument('--weight_decay',
default=1e-5,
type=float,
help='weight decay')
parser.add_argument('--dropout',
default=0.0,
type=float,
help='dropout rate')
parser.add_argument('--na_rate',
default=0,
type=int,
help='NA rate (NA = Q * na_rate)')
parser.add_argument('--grad_iter',
default=1,
type=int,
help='accumulate gradient every x iterations')
parser.add_argument('--optim',
default='sgd',
help='sgd / adam / bert_adam')
parser.add_argument('--hidden_size',
default=230,
type=int,
help='hidden size')
parser.add_argument('--load_ckpt', default=None, help='load ckpt')
parser.add_argument('--save_ckpt', default=None, help='save ckpt')
parser.add_argument('--fp16',
action='store_true',
help='use nvidia apex fp16')
parser.add_argument('--only_test', action='store_true', help='only test')
parser.add_argument('--pair', action='store_true', help='use pair model')
opt = parser.parse_args()
trainN = opt.trainN
N = opt.N
K = opt.K
Q = opt.Q
batch_size = opt.batch_size
model_name = opt.model
encoder_name = opt.encoder
max_length = opt.max_length
print("{}-way-{}-shot Few-Shot Relation Classification".format(N, K))
print("model: {}".format(model_name))
print("encoder: {}".format(encoder_name))
print("max_length: {}".format(max_length))
if encoder_name == 'cnn':
try:
glove_mat = np.load('./pretrain/glove/glove_mat.npy')
glove_word2id = json.load(
open('./pretrain/glove/glove_word2id.json'))
except:
raise Exception(
"Cannot find glove files. Run glove/download_glove.sh to download glove files."
)
sentence_encoder = CNNSentenceEncoder(glove_mat, glove_word2id,
max_length)
elif encoder_name == 'bert':
if opt.pair:
sentence_encoder = BERTPAIRSentenceEncoder(
'./pretrain/bert-base-uncased', max_length)
else:
sentence_encoder = BERTSentenceEncoder(
'./pretrain/bert-base-uncased', max_length)
else:
raise NotImplementedError
if opt.pair:
train_data_loader = get_loader_pair(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
val_data_loader = get_loader_pair(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
test_data_loader = get_loader_pair(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
else:
train_data_loader = get_loader(opt.train,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
val_data_loader = get_loader(opt.val,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
test_data_loader = get_loader(opt.test,
sentence_encoder,
N=N,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.adv:
adv_data_loader = get_loader_unsupervised(opt.adv,
sentence_encoder,
N=trainN,
K=K,
Q=Q,
na_rate=opt.na_rate,
batch_size=batch_size)
if opt.optim == 'sgd':
pytorch_optim = optim.SGD
elif opt.optim == 'adam':
pytorch_optim = optim.Adam
elif opt.optim == 'bert_adam':
from pytorch_transformers import AdamW
pytorch_optim = AdamW
else:
raise NotImplementedError
if opt.adv:
d = Discriminator(opt.hidden_size)
framework = FewShotREFramework(train_data_loader,
val_data_loader,
test_data_loader,
adv_data_loader,
adv=opt.adv,
d=d)
else:
framework = FewShotREFramework(train_data_loader, val_data_loader,
test_data_loader)
prefix = '-'.join(
[model_name, encoder_name, opt.train, opt.val,
str(N), str(K)])
if opt.adv is not None:
prefix += '-adv_' + opt.adv
if opt.na_rate != 0:
prefix += '-na{}'.format(opt.na_rate)
if model_name == 'proto':
model = Proto(sentence_encoder, hidden_size=opt.hidden_size)
elif model_name == 'gnn':
model = GNN(sentence_encoder, N)
elif model_name == 'snail':
print("HINT: SNAIL works only in PyTorch 0.3.1")
model = SNAIL(sentence_encoder, N, K)
elif model_name == 'metanet':
model = MetaNet(N, K, sentence_encoder.embedding, max_length)
elif model_name == 'siamese':
model = Siamese(sentence_encoder,
hidden_size=opt.hidden_size,
dropout=opt.dropout)
elif model_name == 'pair':
model = Pair(sentence_encoder, hidden_size=opt.hidden_size)
else:
raise NotImplementedError
if not os.path.exists('checkpoint'):
os.mkdir('checkpoint')
ckpt = 'checkpoint/{}.pth.tar'.format(prefix)
if opt.save_ckpt:
ckpt = opt.save_ckpt
if torch.cuda.is_available():
model.cuda()
if not opt.only_test:
if encoder_name == 'bert':
bert_optim = True
else:
bert_optim = False
framework.train(model,
prefix,
batch_size,
trainN,
N,
K,
Q,
pytorch_optim=pytorch_optim,
load_ckpt=opt.load_ckpt,
save_ckpt=ckpt,
na_rate=opt.na_rate,
val_step=opt.val_step,
fp16=opt.fp16,
pair=opt.pair,
train_iter=opt.train_iter,
val_iter=opt.val_iter,
bert_optim=bert_optim)
else:
ckpt = opt.load_ckpt
acc = 0
his_acc = []
total_test_round = 5
for i in range(total_test_round):
cur_acc = framework.eval(model,
batch_size,
N,
K,
Q,
opt.test_iter,
na_rate=opt.na_rate,
ckpt=ckpt,
pair=opt.pair)
his_acc.append(cur_acc)
acc += cur_acc
acc /= total_test_round
nhis_acc = np.array(his_acc)
error = nhis_acc.std() * 1.96 / (nhis_acc.shape[0]**0.5)
print("RESULT: %.2f\\pm%.2f" % (acc * 100, error * 100))
result_file = open('./result.txt', 'a+')
result_file.write(
"test data: %12s | model: %45s | acc: %.6f\n | error: %.6f\n" %
(opt.test, prefix, acc, error))
result_file = open('./result_detail.txt', 'a+')
result_detail = {
'test': opt.test,
'model': prefix,
'acc': acc,
'his': his_acc
}
result_file.write("%s\n" % (json.dumps(result_detail)))
class Detector:
#RUNS USING GPU if available and pytoch has CUDA support
def __init__(self, chpt_path, max_length=128):
"""
Initializer
"""
self.checkpoint_path = chpt_path
self.bert_pretrained_checkpoint = 'bert-base-uncased'
self.max_length = max_length
self.sentence_encoder = BERTPAIRSentenceEncoder(
self.bert_pretrained_checkpoint, self.max_length)
self.model = Pair(self.sentence_encoder, hidden_size=768)
if torch.cuda.is_available():
self.model = self.model.cuda()
self.model.eval()
# self.nlp_coref = spacy.load("en_core_web_sm")
# neuralcoref.add_to_pipe(self.nlp_coref)
self.nlp_no_coref = spacy.load("en_core_web_sm")
self.load_model()
def __load_model_from_checkpoint__(self, ckpt):
'''
ckpt: Path of the checkpoint
return: Checkpoint dict
'''
if os.path.isfile(ckpt):
checkpoint = torch.load(ckpt)
print("Successfully loaded checkpoint '%s'" % ckpt)
return checkpoint
else:
raise Exception("No checkpoint found at '%s'" % ckpt)
def bert_tokenize(self, tokens, head_indices, tail_indices):
word = self.sentence_encoder.tokenize(tokens, head_indices,
tail_indices)
return word
def load_model(self):
"""
Loads the model from the checkpoint
"""
state_dict = self.__load_model_from_checkpoint__(
self.checkpoint_path)['state_dict']
own_state = self.model.state_dict()
for name, param in state_dict.items():
if name not in own_state:
continue
own_state[name].copy_(param)
# def spacy_tokenize_coref(self,sentence):
# """
# Tokenizes the sentence using spacy
# """
# return list(map(str, self.nlp_coref(sentence)))
def spacy_tokenize_no_coref(self, sentence):
"""
Tokenizes the sentence using spacy
"""
try:
return list(map(str, self.nlp_no_coref(sentence)))
except TypeError as e:
print("problem sentence: '{}'".format(sentence))
raise e
# def get_head_tail_pairs(self,sentence):
# """
# Gets pairs of heads and tails of named entities so that relation identification can be done on these.
# """
# acceptable_entity_types = ['PERSON', 'NORP', 'ORG', 'GPE', 'PRODUCT', 'EVENT', 'LAW', 'LOC', 'FAC']
# doc = self.nlp_coref(sentence)
# entity_info = [(X.text, X.label_) for X in doc.ents]
# entity_info = set(map(lambda x:x[0], filter(lambda x:x[1] in acceptable_entity_types, entity_info)))
# return combinations(entity_info, 2)
def _get_indices_alt(self, tokens, tokenized_head, tokenized_tail):
"""
Alternative implemention for getting the indices of the head and tail if exact matches cannot be done.
"""
head_indices = None
tail_indices = None
for i in range(len(tokens)):
if tokens[i] in tokenized_head[0] or tokenized_head[0] in tokens[i]:
broke = False
for k, j in zip(tokens[i:i + len(tokenized_head)],
tokenized_head):
if k not in j and j not in k:
broke = True
break
if not broke:
head_indices = list(range(i, i + len(tokenized_head)))
break
for i in range(len(tokens)):
if tokens[i] in tokenized_tail[0] or tokenized_tail[0] in tokens[i]:
broke = False
for k, j in zip(tokens[i:i + len(tokenized_tail)],
tokenized_tail):
if k not in j and j not in k:
broke = True
break
if not broke:
tail_indices = list(range(i, i + len(tokenized_tail)))
break
return head_indices, tail_indices
def _calculate_conf(self, logits, order, pred):
exp = list(float(i) for i in logits[0][0])
exp = [math.exp(i) for i in exp]
if pred == 'NA':
return exp[-1] * 100 / sum(exp)
return exp[order.index(pred)] * 100 / sum(exp)
def run_detection_algorithm(self, query, relation_data):
"""
Runs the algorithm/model on the given query using the given support data.
"""
N = len(relation_data)
K = len(relation_data[0]['examples'])
Q = 1
head = query['head']
tail = query['tail']
fusion_set = {'word': [], 'mask': [], 'seg': []}
tokens = self.spacy_tokenize_no_coref(query['sentence'])
print("head: '{}' tail: '{}' sentence: '{}'".format(
head, tail, query['sentence']))
tokenized_head = self.spacy_tokenize_no_coref(head)
tokenized_tail = self.spacy_tokenize_no_coref(tail)
head_indices = None
tail_indices = None
for i in range(len(tokens)):
if tokens[i] == tokenized_head[0] and tokens[
i:i + len(tokenized_head)] == tokenized_head:
head_indices = list(range(i, i + len(tokenized_head)))
break
for i in range(len(tokens)):
if tokens[i] == tokenized_tail[0] and tokens[
i:i + len(tokenized_tail)] == tokenized_tail:
tail_indices = list(range(i, i + len(tokenized_tail)))
break
if head_indices is None or tail_indices is None:
head_indices, tail_indices = self._get_indices_alt(
tokens, tokenized_head, tokenized_tail)
if head_indices is None or tail_indices is None:
print(tokenized_head)
print(tokenized_tail)
print(tokens)
raise ValueError(
"Head/Tail indices error: head: {} \n tail: {} \n sentence: {}"
.format(head, tail, query['sentence']))
bert_query_tokens = self.bert_tokenize(tokens, head_indices,
tail_indices)
for relation in relation_data:
for ex in relation['examples']:
tokens = self.spacy_tokenize_no_coref(ex['sentence'])
tokenized_head = self.spacy_tokenize_no_coref(
ex['head']
) #head and tail spelling and punctuation should match the corefered output exactly
tokenized_tail = self.spacy_tokenize_no_coref(ex['tail'])
head_indices = None
tail_indices = None
for i in range(len(tokens)):
if tokens[i] == tokenized_head[0] and tokens[
i:i + len(tokenized_head)] == tokenized_head:
head_indices = list(range(i, i + len(tokenized_head)))
break
for i in range(len(tokens)):
if tokens[i] == tokenized_tail[0] and tokens[
i:i + len(tokenized_tail)] == tokenized_tail:
tail_indices = list(range(i, i + len(tokenized_tail)))
break
if head_indices is None or tail_indices is None:
raise ValueError(
"Head/Tail indices error: head: {} \n tail: {} \n sentence: {}"
.format(ex['head'], ex['tail'], ex['sentence']))
bert_relation_example_tokens = self.bert_tokenize(
tokens, head_indices, tail_indices)
SEP = self.sentence_encoder.tokenizer.convert_tokens_to_ids(
['[SEP]'])
CLS = self.sentence_encoder.tokenizer.convert_tokens_to_ids(
['[CLS]'])
word_tensor = torch.zeros((self.max_length)).long()
new_word = CLS + bert_relation_example_tokens + SEP + bert_query_tokens + SEP
for i in range(min(self.max_length, len(new_word))):
word_tensor[i] = new_word[i]
mask_tensor = torch.zeros((self.max_length)).long()
mask_tensor[:min(self.max_length, len(new_word))] = 1
seg_tensor = torch.ones((self.max_length)).long()
seg_tensor[:min(self.max_length,
len(bert_relation_example_tokens) + 1)] = 0
fusion_set['word'].append(word_tensor)
fusion_set['mask'].append(mask_tensor)
fusion_set['seg'].append(seg_tensor)
fusion_set['word'] = torch.stack(fusion_set['word'])
fusion_set['seg'] = torch.stack(fusion_set['seg'])
fusion_set['mask'] = torch.stack(fusion_set['mask'])
if torch.cuda.is_available():
fusion_set['word'] = fusion_set['word'].cuda()
fusion_set['seg'] = fusion_set['seg'].cuda()
fusion_set['mask'] = fusion_set['mask'].cuda()
logits, pred = self.model(fusion_set, N, K, Q)
gc.collect()
order = list(r['name'] for r in relation_data)
pred_relation = relation_data[
pred.item()]['name'] if pred.item() < len(relation_data) else 'NA'
return {
'sentence': query['sentence'],
'head': head,
'tail': tail,
'pred_relation': pred_relation,
'conf': int(self._calculate_conf(logits, order, pred_relation))
} #returns (sentence, head, tail, prediction relation name)
def print_result(self, sentence, head, tail, prediction):
"""
Helper function to print the results to the stdout.
"""
print('Sentence: \"{}\", head: \"{}\", tail: \"{}\", prediction: {}'.
format(sentence, head, tail, prediction))