def main():
# Set hyperparameters
hp = {
'LAYERS': 1,
'THRESHOLD': 2,
'DECAY': 3,
'EPOCH': 25,
'WINDOW_SIZE': 3,
'DIM_UNI': 32,
'DIM_BI': 16,
'DIM_WORD': 16,
'DIM_CTYPE': 8,
'DIM_TAGEMB': 16,
'DIM_HIDDEN': 100,
'LEARNING_RATE': 0.075,
'DROPOUT_RATE': 0.2,
'TRAINSET': args.train,
'TESTSET': args.test,
'DEVSET': args.dev,
'DICTIONARY': args.dict,
'HYPERPARAMS': 'data/' + args.output + '.hp',
'MODEL': 'data/' + args.output + '.model',
'VOCAB': 'data/' + args.output + '.dict',
'EPOCH_MODEL': 'data/epoch.model',
'TMP_PRED': 'data/pred',
'TMP_GOLD': 'data/gold'
}
# Setup vocabuary files
vocabs = prepro.create_vocabs_from_trainset(trainset=hp['TRAINSET'],
fn_dictionary=hp['DICTIONARY'],
fn_vocabs=hp['VOCAB'])
# Update hyper-parameters
hp['VOCAB_SIZE_UNI'] = len(vocabs[0])
hp['VOCAB_SIZE_BI'] = len(vocabs[1])
hp['VOCAB_SIZE_WORD'] = len(vocabs[2])
hp['VOCAB_SIZE_POSTAG'] = len(vocabs[3])
# Preprocess
ws = hp['WINDOW_SIZE']
TrainData = prepro.from_file(filename=hp['TRAINSET'],
window_size=ws,
vocabs=vocabs)
TestData = prepro.from_file(filename=hp['TESTSET'],
window_size=ws,
vocabs=vocabs)
DevData = prepro.from_file(filename=hp['DEVSET'],
window_size=ws,
vocabs=vocabs)
# Construct networks
_model = model.Model(hp=hp)
# Start training
fit(hp,
model=_model,
train_data=TrainData,
test_data=TestData,
dev_data=DevData)
def fit(train_file,
dev_file,
test_file,
model_name,
dict_file=None,
emb_file=None,
delimiter='\t',
newline='EOS',
layers=1,
min_count=3,
decay=1,
epoch=10,
window_size=3,
dim_uni=32,
dim_bi=16,
dim_word=16,
dim_ctype=8,
dim_tagemb=16,
dim_hidden=100,
learning_rate=0.1,
dropout_rate=0.3,
seed=1234):
"""Train a joint word segmentation and sequence labeling (e.g, POS-tagging, NER) model.
args:
- train_file (str): Path to a train file.
- dev_file (str): Path to a development file for early stopping.
- test_file (str): Path to a test file for evaluation.
- model_name (str): Output model filename.
- dict_file (str, optional): Path to a dictionary file.
- emb_file (str, optional): Path to a pre-trained embedding file (word2vec format).
- delimiter (str, optional): Separate word and tag in each line by 'delimiter'.
- newline (str, optional): Separate lines in the file by 'newline'.
- layers (int, optional): RNN Layer size.
- min_count (int, optional): Ignores all words with total frequency lower than this.
- decay (int, optional): Learning rate decay.
- epoch (int, optional): Epoch size.
- window_size (int, optional): Window size of the context characters for word segmentation.
- dim_uni (int, optional): Dimensionality of the char-unigram vectors.
- dim_bi (int, optional): Dimensionality of the char-bigram vectors.
- dim_word (int, optional): Dimensionality of the word vectors.
- dim_ctype (int, optional): Dimensionality of the character-type vectors.
- dim_tagemb (int, optional): Dimensionality of the tag vectors.
- dim_hidden (int, optional): Dimensionality of the BiLSTM's hidden layer.
- learning_rate (float, optional): Learning rate of SGD.
- dropout_rate (float, optional): Dropout rate of the input vector for BiLSTMs.
- seed (int, optional): Random seed.
return:
- Nothing. After finish training, however,
save the three model files (*.vocabs, *.params, *.hp) in the current directory.
"""
random.seed(seed)
hp = OrderedDict({
'LAYERS': layers,
'THRESHOLD': min_count,
'DECAY': decay,
'EPOCH': epoch,
'WINDOW_SIZE': window_size,
'DIM_UNI': dim_uni,
'DIM_BI': dim_bi,
'DIM_WORD': dim_word,
'DIM_CTYPE': dim_ctype,
'DIM_TAGEMB': dim_tagemb,
'DIM_HIDDEN': dim_hidden,
'LEARNING_RATE': learning_rate,
'DROPOUT_RATE': dropout_rate,
'SEED': seed,
'TRAINSET': train_file,
'TESTSET': test_file,
'DEVSET': dev_file,
'DICTIONARY': dict_file,
'EMBEDDING': emb_file,
'HYPERPARAMS': model_name + '.hp',
'MODEL': model_name + '.params',
'VOCAB': model_name + '.vocabs',
'EPOCH_MODEL': model_name + '_epoch.params'
})
# Preprocess
vocabs = prepro.create_vocabs_from_trainset(trainset=hp['TRAINSET'],
fn_dictionary=hp['DICTIONARY'],
fn_vocabs=hp['VOCAB'],
delimiter=delimiter,
newline=newline)
if emb_file is not None:
embs, dim_word = prepro.embedding_loader(fn_embedding=hp['EMBEDDING'],
word2id=vocabs[2])
hp['DIM_WORD'] = dim_word
else:
embs = None
TrainData = prepro.from_file(filename=hp['TRAINSET'],
window_size=hp['WINDOW_SIZE'],
vocabs=vocabs,
delimiter=delimiter,
newline=newline)
TestData = prepro.from_file(filename=hp['TESTSET'],
window_size=hp['WINDOW_SIZE'],
vocabs=vocabs,
delimiter=delimiter,
newline=newline)
DevData = prepro.from_file(filename=hp['DEVSET'],
window_size=hp['WINDOW_SIZE'],
vocabs=vocabs,
delimiter=delimiter,
newline=newline)
# Update hyper-parameters
hp['NUM_TRAIN'] = len(TrainData.ws_data)
hp['NUM_TEST'] = len(TestData.ws_data)
hp['NUM_DEV'] = len(DevData.ws_data)
hp['VOCAB_SIZE_UNI'] = len(vocabs[0])
hp['VOCAB_SIZE_BI'] = len(vocabs[1])
hp['VOCAB_SIZE_WORD'] = len(vocabs[2])
hp['VOCAB_SIZE_POSTAG'] = len(vocabs[3])
# Construct networks
_model = model.Model(hp=hp, embs=embs)
# Start training
_start(hp,
model=_model,
train_data=TrainData,
test_data=TestData,
dev_data=DevData)