def __pretrain_bertnrdj(dataset, n_labels, seq_length, n_steps, batch_size,
dropout, n_layers, load_model_file, dst_model_file):
init_logging('log/{}-pre-bertnrdj-{}-{}.log'.format(
cur_script_name, utils.get_machine_name(), str_today),
mode='a',
to_stdout=True)
dataset_files = config.DATA_FILES[dataset]
print('init robert ...')
bert_config = bertmodel.BertConfig.from_json_file(config.BERT_CONFIG_FILE)
robert_model = robert.Robert(
bert_config,
n_labels=n_labels,
seq_length=config.BERT_SEQ_LEN,
is_train=False,
init_checkpoint=dataset_files['bert_init_checkpoint'])
print('done')
yelp_tv_idxs_file = os.path.join(
config.RES_DIR,
'yelp/eng-part/yelp-rest-sents-r9-tok-eng-p0_04-tvidxs.txt')
amazon_tv_idxs_file = os.path.join(
config.RES_DIR, 'amazon/laptops-reivews-sent-tok-text-tvidxs.txt')
tv_idxs_file = amazon_tv_idxs_file if dataset == 'se14l' else yelp_tv_idxs_file
print('loading data ...')
idxs_train, idxs_valid = datautils.load_train_valid_idxs(tv_idxs_file)
logging.info('{} valid samples'.format(len(idxs_valid)))
# idxs_valid = set(idxs_valid)
valid_aspect_terms_list = __load_terms_list(
idxs_valid, dataset_files['pretrain_aspect_terms_file'])
valid_opinion_terms_list = __load_terms_list(
idxs_valid, dataset_files['pretrain_opinion_terms_file'])
print('done')
bertnrdj_model = BertNRDJ(n_labels,
config.BERT_EMBED_DIM,
hidden_size_lstm=hidden_size_lstm,
batch_size=batch_size,
model_file=load_model_file,
n_lstm_layers=n_layers)
bertnrdj_model.pretrain(
robert_model=robert_model,
train_aspect_tfrec_file=dataset_files[
'pretrain_train_aspect_tfrec_file'],
valid_aspect_tfrec_file=dataset_files[
'pretrain_valid_aspect_tfrec_file'],
train_opinion_tfrec_file=dataset_files[
'pretrain_train_opinion_tfrec_file'],
valid_opinion_tfrec_file=dataset_files[
'pretrain_valid_opinion_tfrec_file'],
valid_tokens_file=dataset_files['pretrain_valid_token_file'],
seq_length=seq_length,
valid_aspect_terms_list=valid_aspect_terms_list,
valid_opinion_terms_list=valid_opinion_terms_list,
n_steps=n_steps,
batch_size=batch_size,
dropout=dropout,
save_file=dst_model_file)
def __train(word_vecs_file, train_tok_texts_file, train_sents_file,
train_valid_split_file, test_tok_texts_file, test_sents_file,
load_model_file, task):
init_logging('log/{}-train-{}-{}.log'.format(
os.path.splitext(os.path.basename(__file__))[0],
utils.get_machine_name(), str_today),
mode='a',
to_stdout=True)
dst_aspects_file, dst_opinions_file = None, None
# n_train = 1000
n_train = -1
n_tags = 5
batch_size = 64
lr = 0.001
share_lstm = False
logging.info(word_vecs_file)
logging.info('load model {}'.format(load_model_file))
logging.info(test_sents_file)
print('loading data ...')
vocab, word_vecs_matrix = datautils.load_word_vecs(word_vecs_file)
logging.info('word vec dim: {}, n_words={}'.format(
word_vecs_matrix.shape[1], word_vecs_matrix.shape[0]))
train_data, valid_data, test_data = modelutils.get_data_semeval(
train_sents_file, train_tok_texts_file, train_valid_split_file,
test_sents_file, test_tok_texts_file, vocab, n_train, task)
print('done')
test_f1s = list()
for i in range(5):
logging.info('turn {}'.format(i))
model = RINANTE(n_tags,
word_vecs_matrix,
share_lstm,
hidden_size_lstm=hidden_size_lstm,
model_file=load_model_file,
batch_size=batch_size,
lamb=lamb)
test_af1, _ = model.train(train_data,
valid_data,
test_data,
vocab,
n_epochs=n_epochs,
lr=lr,
dst_aspects_file=dst_aspects_file,
dst_opinions_file=dst_opinions_file)
test_f1s.append(test_af1)
logging.info('r={} test_f1={:.4f}'.format(i, test_af1))
tf.reset_default_graph()
logging.info('avg_test_f1={:.4f}'.format(sum(test_f1s) / len(test_f1s)))
def __train_nrdj(word_vecs_file, train_tok_texts_file, train_sents_file,
train_valid_split_file, test_tok_texts_file, test_sents_file,
load_model_file, task):
init_logging('log/{}-train-{}-{}.log'.format(
os.path.splitext(os.path.basename(__file__))[0],
utils.get_machine_name(), str_today),
mode='a',
to_stdout=True)
dst_aspects_file = '/home/hldai/data/aspect/semeval14/nrdj-aspects.txt'
dst_opinions_file = '/home/hldai/data/aspect/semeval14/nrdj-opinions.txt'
# dst_aspects_file, dst_opinions_file = None, None
# n_train = 1000
n_train = -1
label_opinions = True
# label_opinions = False
n_tags = 5 if label_opinions else 3
# n_tags = 5 if task == 'train' else 3
batch_size = 20
lr = 0.001
share_lstm = False
logging.info(word_vecs_file)
logging.info('load model {}'.format(load_model_file))
logging.info(test_sents_file)
print('loading data ...')
vocab, word_vecs_matrix = __load_word_vecs(word_vecs_file)
logging.info('word vec dim: {}, n_words={}'.format(
word_vecs_matrix.shape[1], word_vecs_matrix.shape[0]))
train_data, valid_data, test_data = datautils.get_data_semeval(
train_sents_file, train_tok_texts_file, train_valid_split_file,
test_sents_file, test_tok_texts_file, vocab, n_train, label_opinions)
print('done')
nrdj = NeuRuleDoubleJoint(n_tags,
word_vecs_matrix,
share_lstm,
hidden_size_lstm=hidden_size_lstm,
model_file=load_model_file,
batch_size=batch_size)
nrdj.train(train_data,
valid_data,
test_data,
vocab,
n_epochs=n_epochs,
lr=lr,
dst_aspects_file=dst_aspects_file,
dst_opinions_file=dst_opinions_file)
def __train_bertnrdj(dataset, n_labels, batch_size, model_file, dropout,
n_epochs, learning_rate, start_eval_epoch, n_layers):
init_logging('log/{}-bertnrdj-{}-{}.log'.format(cur_script_name,
utils.get_machine_name(),
str_today),
mode='a',
to_stdout=True)
dataset_files = config.DATA_FILES[dataset]
n_train, data_valid = bldatautils.load_train_data_bert_ol(
dataset_files['train_sents_file'],
dataset_files['train_valid_split_file'],
dataset_files['bert_valid_tokens_file'])
data_test = bldatautils.load_test_data_bert_ol(
dataset_files['test_sents_file'],
dataset_files['bert_test_tokens_file'])
bert_config = bertmodel.BertConfig.from_json_file(config.BERT_CONFIG_FILE)
bm = robert.Robert(bert_config,
n_labels=n_labels,
seq_length=config.BERT_SEQ_LEN,
is_train=False,
init_checkpoint=dataset_files['bert_init_checkpoint'])
logging.info('batch_size={}, learning_rate={}, dropout={}'.format(
batch_size, learning_rate, dropout))
# model_file = dataset_files['pretrained_bertnrdj_file']
# model_file = None
bertnrdj_model = BertNRDJ(n_labels,
config.BERT_EMBED_DIM,
hidden_size_lstm=hidden_size_lstm,
batch_size=batch_size,
model_file=model_file,
n_lstm_layers=n_layers)
bertnrdj_model.train(
robert_model=bm,
train_tfrec_file=dataset_files['train_tfrecord_file'],
valid_tfrec_file=dataset_files['valid_tfrecord_file'],
test_tfrec_file=dataset_files['test_tfrecord_file'],
seq_length=config.BERT_SEQ_LEN,
n_train=n_train,
data_valid=data_valid,
data_test=data_test,
dropout=dropout,
start_eval_spoch=start_eval_epoch,
n_epochs=n_epochs,
lr=learning_rate,
)
def __train(word_vecs_file, train_tok_texts_file, train_sents_file,
train_valid_split_file, test_tok_texts_file, test_sents_file,
load_model_file, task):
init_logging('log/{}-train-{}-{}.log'.format(
os.path.splitext(os.path.basename(__file__))[0],
utils.get_machine_name(), str_today),
mode='a',
to_stdout=True)
dst_aspects_file, dst_opinions_file = None, None
# n_train = 1000
n_train = -1
n_tags = 5
batch_size = 32
lr = 0.001
share_lstm = False
logging.info(word_vecs_file)
logging.info('load model {}'.format(load_model_file))
logging.info(test_sents_file)
print('loading data ...')
vocab, word_vecs_matrix = datautils.load_word_vecs(word_vecs_file)
logging.info('word vec dim: {}, n_words={}'.format(
word_vecs_matrix.shape[1], word_vecs_matrix.shape[0]))
train_data, valid_data, test_data = modelutils.get_data_semeval(
train_sents_file, train_tok_texts_file, train_valid_split_file,
test_sents_file, test_tok_texts_file, vocab, n_train, task)
print('done')
model = RINANTE(n_tags,
word_vecs_matrix,
share_lstm,
hidden_size_lstm=hidden_size_lstm,
model_file=load_model_file,
batch_size=batch_size,
lamb=lamb)
model.train(train_data,
valid_data,
test_data,
vocab,
n_epochs=n_epochs,
lr=lr,
dst_aspects_file=dst_aspects_file,
dst_opinions_file=dst_opinions_file)
def __pre_train_nrdj(word_vecs_file,
tok_texts_file,
aspect_terms_file,
opinion_terms_file,
dst_model_file,
task,
lamb,
lstm_l2,
train_word_embeddings=False,
load_model_file=None):
init_logging('log/{}-pre-{}-{}.log'.format(
os.path.splitext(os.path.basename(__file__))[0],
utils.get_machine_name(), str_today),
mode='a',
to_stdout=True)
# n_train = 1000
n_train = -1
label_opinions = True
# label_opinions = False
n_tags = 5 if label_opinions else 3
# n_tags = 5 if task == 'train' else 3
batch_size = 32
lr = 0.001
share_lstm = False
logging.info(word_vecs_file)
logging.info(aspect_terms_file)
logging.info(opinion_terms_file)
logging.info('dst: {}'.format(dst_model_file))
print('loading data ...')
vocab, word_vecs_matrix = __load_word_vecs(word_vecs_file)
# train_data_src1, valid_data_src1 = __get_data_amazon(vocab, config.AMAZON_TERMS_TRUE1_FILE)
# train_data_src1, valid_data_src1 = __get_data_amazon(vocab, config.AMAZON_TERMS_TRUE3_FILE)
train_data_src1, valid_data_src1 = datautils.get_data_amazon(
vocab, aspect_terms_file, tok_texts_file, 'aspect')
train_data_src2, valid_data_src2 = datautils.get_data_amazon(
vocab, opinion_terms_file, tok_texts_file, 'opinion')
print('done')
logging.info('train_word_embeddings={} lstm_l2={}'.format(
train_word_embeddings, lstm_l2))
nrdj = NeuRuleDoubleJoint(n_tags,
word_vecs_matrix,
share_lstm,
hidden_size_lstm=hidden_size_lstm,
train_word_embeddings=train_word_embeddings,
lamb=lamb,
lstm_l2_src=lstm_l2,
model_file=load_model_file,
batch_size=batch_size)
nrdj.pre_train(train_data_src1,
valid_data_src1,
train_data_src2,
valid_data_src2,
vocab,
n_epochs=50,
lr=lr,
save_file=dst_model_file)