def main(cfgpath, global_step):
# parsing config.json
proj_dir = Path.cwd()
params = json.load((proj_dir / cfgpath).open())
# create dataset
batch_size = params['training'].get('batch_size')
tr_filepath = params['filepath'].get('tr')
val_filepath = params['filepath'].get('val')
tr_ds = create_dataset(tr_filepath, batch_size, True)
val_ds = create_dataset(val_filepath, batch_size, False)
# create pre_processor
vocab = pickle.load((proj_dir / params['filepath'].get('vocab')).open(mode='rb'))
pre_processor = PreProcessor(vocab=vocab, tokenizer=MeCab().morphs, pad_idx=1)
# create model
model = SenCNN(num_classes=2, vocab=vocab)
# create optimizer & loss_fn
epochs = params['training'].get('epochs')
learning_rate = params['training'].get('learning_rate')
opt = tf.optimizers.Adam(learning_rate=learning_rate)
loss_fn = tf.losses.SparseCategoricalCrossentropy(from_logits=True)
writer = tf.summary.create_file_writer(logdir='./runs/exp')
# training
for epoch in tqdm(range(epochs), desc='epochs'):
tr_loss = 0
tf.keras.backend.set_learning_phase(1)
for step, mb in tqdm(enumerate(tr_ds), desc='steps'):
x_mb, y_mb = pre_processor.convert2idx(mb)
with tf.GradientTape() as tape:
mb_loss = loss_fn(y_mb, model(x_mb))
grads = tape.gradient(target=mb_loss, sources=model.trainable_variables)
opt.apply_gradients(grads_and_vars=zip(grads, model.trainable_variables))
tr_loss += mb_loss.numpy()
if tf.equal(opt.iterations % global_step, 0):
with writer.as_default():
val_loss = evaluate(model, val_ds, loss_fn, pre_processor.convert2idx)
tf.summary.scalar('tr_loss', tr_loss / (step + 1), step=opt.iterations)
tf.summary.scalar('val_loss', val_loss, step=opt.iterations)
tf.keras.backend.set_learning_phase(1)
else:
tr_loss /= (step + 1)
val_loss = evaluate(model, val_ds, loss_fn, pre_processor.convert2idx)
tqdm.write('epoch : {}, tr_loss : {:.3f}, val_loss : {:.3f}'.format(epoch + 1, tr_loss, val_loss))
ckpt_path = proj_dir / params['filepath'].get('ckpt')
ckpt = tf.train.Checkpoint(model=model)
ckpt.save(ckpt_path)
def main(cfgpath):
# parsing config.json
proj_dir = Path.cwd()
params = json.load((proj_dir / cfgpath).open())
# create dataset
batch_size = params['training'].get('batch_size')
tr_filepath = params['filepath'].get('tr')
val_filepath = params['filepath'].get('val')
tr_ds = create_dataset(tr_filepath, batch_size, True)
val_ds = create_dataset(val_filepath, batch_size, False)
# create pre_processor
vocab = pickle.load(
(proj_dir / params['filepath'].get('vocab')).open(mode='rb'))
pre_processor = PreProcessor(vocab=vocab, tokenizer=Okt)
# create model
model = SenCNN(num_classes=2, vocab=vocab)
# create optimizer & loss_fn
epochs = params['training'].get('epochs')
learning_rate = params['training'].get('learning_rate')
opt = tf.optimizers.Adam(learning_rate=learning_rate)
loss_fn = tf.losses.SparseCategoricalCrossentropy()
# training
for epoch in tqdm(range(epochs), desc='epochs'):
tr_loss = 0
tf.keras.backend.set_learning_phase(1)
for step, mb in tqdm(enumerate(tr_ds), desc='steps'):
x_mb, y_mb = pre_processor.convert2idx(mb)
with tf.GradientTape() as tape:
mb_loss = loss_fn(y_mb, model(x_mb))
grads = tape.gradient(target=mb_loss,
sources=model.trainable_variables)
opt.apply_gradients(
grads_and_vars=zip(grads, model.trainable_variables))
tr_loss += mb_loss.numpy()
else:
tr_loss /= (step + 1)
tf.keras.backend.set_learning_phase(0)
val_loss = 0
for step, mb in tqdm(enumerate(val_ds), desc='steps'):
x_mb, y_mb = pre_processor.convert2idx(mb)
mb_loss = loss_fn(y_mb, model(x_mb))
val_loss += mb_loss.numpy()
else:
val_loss /= (step + 1)
tqdm.write('epoch : {}, tr_loss : {:.3f}, val_loss : {:.3f}'.format(
epoch + 1, tr_loss, val_loss))
def main(self):
batch_size = self._batch_size
tr_filepath = 'data/train.txt'
val_filepath = 'data/val.txt'
tr_ds = SmCnn.create_dataset(self,
tr_filepath,
batch_size,
shuffle=True)
val_ds = SmCnn.create_dataset(self,
val_filepath,
batch_size,
shuffle=False)
vocab = pd.read_pickle('data/vocab.pkl')
pre_processor = PreProcessor(vocab=vocab, tokenizer=Mecab())
# create model
model = SmCnn()
# create optimizer & loss_fn
epochs = self._epochs
learning_rate = self._learning_rate
opt = tf.keras.optimizers.Adam(learning_rate=learning_rate)
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy()
# training
for epoch in tqdm(range(epochs), desc='steps'):
tr_loss = 0
tf.keras.backend.set_learning_phase(1)
for step, mb in tqdm(enumerate(tr_ds), desc='steps'):
x_mb, y_mb = pre_processor.convert2idx(mb)
with tf.GradientTape() as tape:
mb_loss = loss_fn(y_mb, model(x_mb))
grads = tape.gradient(target=mb_loss,
sources=model.trainable_variables)
opt.apply_gradients(
grads_and_vars=zip(grads, model.trainable_variables))
tr_loss += mb_loss.numpy()
else:
tr_loss /= (step + 1)
tf.keras.backend.set_learning_phase(0)
val_loss = 0
for step, mb in tqdm(enumerate(val_ds), desc='steps'):
x_mb, y_mb = pre_processor.convert2idx(mb)
mb_loss = loss_fn(y_mb, model(x_mb))
val_loss += mb_loss.numpy()
else:
val_loss /= (step + 1)
tqdm.write(
'epoch : {}, tr_loss : {:.3f}, val_loss : {:.3f}'.format(
epoch + 1, tr_loss, val_loss))
def get_preprocessor(ptr_config_info, model_config):
with open(ptr_config_info.vocab, mode='rb') as io:
vocab = pickle.load(io)
if model_config.type == 'etri':
ptr_tokenizer = ETRITokenizer.from_pretrained(ptr_config_info.tokenizer, do_lower_case=False)
pad_sequence = PadSequence(length=model_config.length, pad_val=vocab.to_indices(vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab, split_fn=ptr_tokenizer.tokenize, pad_fn=pad_sequence)
elif model_config.type == 'skt':
ptr_tokenizer = SentencepieceTokenizer(ptr_config_info.tokenizer)
pad_sequence = PadSequence(length=model_config.length, pad_val=vocab.to_indices(vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab, split_fn=ptr_tokenizer, pad_fn=pad_sequence)
return preprocessor
def main(cfgpath):
# parsing config.json
proj_dir = Path.cwd()
params = json.load((proj_dir / cfgpath).open())
# create dataset
batch_size = params['training'].get('batch_size')
tr_filepath = params['filepath'].get('tr')
val_filepath = params['filepath'].get('val')
tst_filepath = params['filepath'].get('tst')
tr_ds = create_dataset(tr_filepath, batch_size, False, False)
val_ds = create_dataset(val_filepath, batch_size, False, False)
tst_ds = create_dataset(tst_filepath, batch_size, False, False)
# create pre_processor
vocab = pickle.load(
(proj_dir / params['filepath'].get('vocab')).open(mode='rb'))
pre_processor = PreProcessor(vocab=vocab,
tokenizer=MeCab().morphs,
pad_idx=1)
# create model
model = SenCNN(num_classes=2, vocab=vocab)
ckpt = tf.train.Checkpoint(model=model)
ckpt.restore(save_path=tf.train.latest_checkpoint(proj_dir / 'checkpoint'))
# evluation
tr_acc = get_accuracy(model, tr_ds, pre_processor.convert2idx)
val_acc = get_accuracy(model, val_ds, pre_processor.convert2idx)
tst_acc = get_accuracy(model, tst_ds, pre_processor.convert2idx)
print('tr_acc: {:.2%}, val_acc : {:.2%}, tst_acc: {:.2%}'.format(
tr_acc, val_acc, tst_acc))
def predict(sentence1, sentence2):
ptr_dir = "C:/Users/aaaaa/workspace/fact-check/BERT_pairwise_text_classification/pretrained"
data_dir = "C:/Users/aaaaa/workspace/fact-check/BERT_pairwise_text_classification/data"
caseType = "skt"
model_dir = "C:/Users/aaaaa/workspace/fact-check/BERT_pairwise_text_classification/experiments/base_model"
checkpoint_model_file = "best_skt.tar"
# ptr_dir = "BERT_pairwise_text_classification/pretrained"
# data_dir = "BERT_pairwise_text_classification/data"
# caseType = "skt"
# model_dir = "BERT_pairwise_text_classification/experiments/base_model"
# checkpoint_model_file = "best_skt.tar"
# ptr_dir = "pretrained"
# data_dir = "data"
# caseType = "skt"
# model_dir = "experiments/base_model"
# checkpoint_model_file = "best_skt.tar"
ptr_dir = Path(ptr_dir)
data_dir = Path(data_dir)
model_dir = Path(model_dir)
checkpoint_model_file = Path(checkpoint_model_file)
ptr_config = Config(ptr_dir / 'config_skt.json')
data_config = Config(data_dir / 'config.json')
model_config = Config(model_dir / 'config.json')
# vocab
with open(os.path.join(ptr_dir, ptr_config.vocab), mode='rb') as io:
vocab = pickle.load(io)
ptr_tokenizer = SentencepieceTokenizer(os.path.join(ptr_dir, ptr_config.tokenizer))
pad_sequence = PadSequence(length=model_config.length, pad_val=vocab.to_indices(vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab, split_fn=ptr_tokenizer, pad_fn=pad_sequence)
# model (restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint(checkpoint_model_file)
config = BertConfig(os.path.join(ptr_dir, ptr_config.config))
model = PairwiseClassifier(config, num_classes=model_config.num_classes, vocab=preprocessor.vocab)
model.load_state_dict(checkpoint['model_state_dict'])
device = torch.device('cpu')
model.to(device)
transform = preprocessor.preprocess
if model.training:
model.eval()
indices, token_types = [torch.tensor([elm]) for elm in transform(sentence1, sentence2)]
with torch.no_grad():
label = model(indices, token_types)
label = label.max(dim=1)[1]
label = label.numpy()[0]
return label
def get_preprocessor(dataset_config, coarse_split_fn, fine_split_fn):
with open(dataset_config.fine_vocab, mode="rb") as io:
fine_vocab = pickle.load(io)
with open(dataset_config.coarse_vocab, mode="rb") as io:
coarse_vocab = pickle.load(io)
preprocessor = PreProcessor(coarse_vocab=coarse_vocab, fine_vocab=fine_vocab,
coarse_split_fn=coarse_split_fn,
fine_split_fn=fine_split_fn)
return preprocessor
data_config = Config(data_dir / 'config.json')
model_config = Config(model_dir / 'config.json')
# vocab
with open(ptr_config.vocab, mode='rb') as io:
vocab = pickle.load(io)
# tokenizer
if args.type == 'etri':
ptr_tokenizer = ETRITokenizer.from_pretrained(ptr_config.tokenizer,
do_lower_case=False)
pad_sequence = PadSequence(length=model_config.length,
pad_val=vocab.to_indices(
vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab,
split_fn=ptr_tokenizer.tokenize,
pad_fn=pad_sequence)
elif args.type == 'skt':
ptr_tokenizer = SentencepieceTokenizer(ptr_config.tokenizer)
pad_sequence = PadSequence(length=model_config.length,
pad_val=vocab.to_indices(
vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab,
split_fn=ptr_tokenizer,
pad_fn=pad_sequence)
# model (restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint('best_{}.tar'.format(
args.type))
config = BertConfig(ptr_config.config)
def main():
tr_filepath = Path.cwd() / 'data' / 'train.txt'
val_filepath = Path.cwd() / 'data' / 'val.txt'
with open(Path.cwd() / 'data/vocab.pkl', mode='rb') as f:
vocab = pickle.load(f)
tr_ds = create_dataset(str(tr_filepath), 128, shuffle=True)
val_ds = create_dataset(str(val_filepath), 128,
shuffle=False) # 평가 데이터는 셔플 ㄴㄴ
tokenized = Okt()
pre_processor = PreProcessor(vocab=vocab, tokenizer=tokenized)
# create model
model = SenCNN(num_classes=2, vocab=vocab)
# create optimizer & loss_fn
epochs = 10
learning_rate = 1e-3
opt = tf.optimizers.Adam(learning_rate=learning_rate)
loss_fn = tf.losses.SparseCategoricalCrossentropy(from_logits=True)
# metrics
tr_loss_metric = tf.keras.metrics.Mean(name='train_loss')
tr_accuracy_metric = tf.keras.metrics.SparseCategoricalAccuracy(
name='train_accuracy')
val_loss_metric = tf.keras.metrics.Mean(name='validation_loss')
val_accuracy_metric = tf.keras.metrics.SparseCategoricalAccuracy(
name='validation_accuracy')
# training
for epoch in tqdm(range(epochs), desc='epochs'):
# trining data
tf.keras.backend.set_learning_phase(1) # train mode
for _, mb in tqdm(enumerate(tr_ds), desc='steps'):
x_mb, y_mb = pre_processor.convert2idx(mb)
x_mb = pre_processor.pad_sequences(x_mb, 70)
x_mb, y_mb = pre_processor.convert_to_tensor(x_mb, y_mb)
with tf.GradientTape() as tape:
mb_loss = loss_fn(y_mb, model(x_mb))
grads = tape.gradient(target=mb_loss,
sources=model.trainable_variables)
opt.apply_gradients(
grads_and_vars=zip(grads, model.trainable_variables))
tr_loss_metric.update_state(mb_loss)
tr_accuracy_metric(y_mb, model(x_mb))
tr_mean_loss = tr_loss_metric.result()
tr_mean_accuracy = tr_accuracy_metric.result()
# test data
tf.keras.backend.set_learning_phase(0) # test mode
for _, mb in tqdm(enumerate(val_ds), desc='steps'):
x_mb, y_mb = pre_processor.convert2idx(mb)
x_mb = pre_processor.pad_sequences(x_mb, 70)
x_mb, y_mb = pre_processor.convert_to_tensor(x_mb, y_mb)
mb_loss = loss_fn(y_mb, model(x_mb))
val_loss_metric.update_state(mb_loss)
val_accuracy_metric.update_state(y_mb, model(x_mb))
val_mean_loss = val_loss_metric.result()
val_mean_accuracy = val_accuracy_metric.result()
tqdm.write(
'epoch : {}, tr_accuracy : {:.3f}, tr_loss : {:.3f}, val_accuracy : {:.3f}, val_loss : {:.3f}'
.format(epoch + 1, tr_mean_accuracy, tr_mean_loss,
val_mean_accuracy, val_mean_loss))
ckpt_path = Path.cwd() / 'checkpoint/ckpt'
ckpt = tf.train.Checkpoint(model=model)
ckpt.save(ckpt_path)
# vocab
vocab = pickle.load(open(ptr_config.vocab, mode='rb'))
# tokenizer
if args.tokenizer == 'ranked':
print('[RANKED TOKENIZER]')
ptr_tokenizer = KBertRankedTokenizer(ptr_config.tokenizer,
do_lower_case=False)
else:
ptr_tokenizer = BertTokenizer.from_pretrained(ptr_config.tokenizer,
do_lower_case=False)
print('[BERT TOKENIZER]')
pad_sequence = PadSequence(length=model_config.length,
pad_val=vocab.to_indices(vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab,
split_fn=ptr_tokenizer.tokenize,
pad_fn=pad_sequence,
subchar=args.subchar)
# model
config = BertConfig(ptr_config.config)
model = SentenceClassifier(config,
num_classes=model_config.num_classes,
vocab=preprocessor.vocab)
bert_pretrained = torch.load(ptr_config.bert)
model.load_state_dict(bert_pretrained, strict=False)
# training
tr_ds = Corpus(data_config.train, preprocessor.preprocess)
tr_dl = DataLoader(tr_ds,
batch_size=model_config.batch_size,
shuffle=True,
model_dir = Path(args.model_dir)
data_config = Config(data_dir / 'config.json')
model_config = Config(model_dir / 'config.json')
# Vocab and Tokenizer
ptr_dir = Path("pretrained")
vocab_filepath = ptr_dir / "{}-vocab.pkl".format(args.type)
with open(vocab_filepath, mode='rb') as io:
vocab = pickle.load(io)
ptr_tokenizer = BertTokenizer.from_pretrained(args.type,
do_lower_case="uncased"
in args.type)
ptr_tokenizer = Tokenizer(vocab, ptr_tokenizer.tokenize)
preprocessor = PreProcessor(ptr_tokenizer, model_config.max_len)
# Load Model
config_filepath = ptr_dir / "{}-config.json".format(args.type)
config = BertConfig.from_pretrained(config_filepath,
output_hidden_states=False)
model = BIIN(config,
vocab,
model_config.hidden_size,
enc_num_layers=len(model_config.hidden_size))
# Data Loader
tr_ds = Corpus(data_config.tr_path,
preprocessor.preprocess,
sep='\t',
doc_col='question1',
if __name__ == "__main__":
args = parser.parse_args()
data_dir = Path(args.data_dir)
model_dir = Path(args.model_dir)
data_config = Config(data_dir / "config.json")
model_config = Config(model_dir / "config.json")
# tokenizer
with open(data_config.fine_vocab, mode="rb") as io:
fine_vocab = pickle.load(io)
with open(data_config.coarse_vocab, mode="rb") as io:
coarse_vocab = pickle.load(io)
preprocessor = PreProcessor(
coarse_vocab=coarse_vocab,
fine_vocab=fine_vocab,
coarse_split_fn=split_morphs,
fine_split_fn=split_jamos,
)
# model
model = SAN(model_config.num_classes, coarse_vocab, fine_vocab,
model_config.fine_embedding_dim, model_config.hidden_dim, model_config.multi_step,
model_config.prediction_drop_ratio)
# training
tr_ds = Corpus(data_config.train, preprocessor.preprocess)
tr_dl = DataLoader(
tr_ds,
batch_size=model_config.batch_size,
shuffle=True,
num_workers=4,
