mean_square_error = tf.reduce_mean(tf.square(output_score - y_))
train_step = tf.train.AdamOptimizer(learning_rate).minimize(
mean_square_error)
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
# DONOTCHANGE: Reserved for nsml use
bind_model(sess=sess, config=config)
# DONOTCHANGE: They are reserved for nsml
if config.pause:
nsml.paused(scope=locals())
if config.mode == 'train':
# 데이터를 로드합니다.
dataset = MovieReviewDataset(DATASET_PATH, config.strmaxlen)
dataset_len = len(dataset)
if config.validate > 0:
validate_len = int(dataset_len * config.validate)
ds_review = np.array([np.array(review) for review, _ in dataset])
ds_label = np.array([int(label) for _, label in dataset])
global_perm = np.random.RandomState(
seed=777).permutation(dataset_len)
ds_review, ds_label = ds_review[global_perm], ds_label[global_perm]
dataset = MovieReviewDataset(remake=True,
new_review=ds_review[validate_len:],
model = get_model(config)
model.summary()
# DONOTCHANGE: Reserved for nsml use
print("nsml binding...")
bind_model(model, config)
# DONOTCHANGE: They are reserved for nsml
if config.pause:
nsml.paused(scope=locals())
# 학습 모드일 때 사용합니다. (기본값)
if config.mode == 'train':
# 데이터를 로드합니다.
print("data loading...")
dataset = MovieReviewDataset(DATASET_PATH, config.strmaxlen)
x_pre = np.array(dataset.reviews_pre)
x_post = np.array(dataset.reviews_post)
y = np.array(dataset.labels)
# epoch마다 학습을 수행합니다.
nsml_callback = Nsml_Callback()
# checkpoint = ModelCheckpoint('./best.hdf5', monitor='val_loss', save_best_only=True, mode='min', period=1)
# dataset_val = MovieReviewDataset_val(DATASET_PATH, config.strmaxlen)
# x_val = np.array(dataset_val.reviews)
# y_val = np.array(dataset_val.labels)
print("model training...")
hist = model.fit(
[x_pre, x_post],
[y, y, y, y, y, y, y, y, y, y, y],
args.add_argument('--strmaxlen', type=int, default=200)
args.add_argument('--embedding', type=int, default=8)
config = args.parse_args()
learning_rate = 1e-3
grad_clip = True
dropout = 0.2
n_char = 4500
n_embed = 256
n_hidden = 256
if not HAS_DATASET and not IS_ON_NSML: # It is not running on nsml
DATASET_PATH = '../sample_data/movie_review/'
if config.mode == 'train':
dataset = MovieReviewDataset(DATASET_PATH, flip=True)
else:
dataset = MovieReviewDataset('', build=False, flip=False)
model = Movie(n_char, n_embed, n_hidden, dropout)
model_run = Movie(n_char, n_embed, n_hidden, dropout)
if GPU_NUM:
model = model.cuda()
model_run.cuda()
accumulate(model_run, model, 0)
# DONOTCHANGE: Reserved for nsml use
bind_model(model_run, dataset, config)
criterion = nn.MSELoss()
# DONOTCHANGE: Reserved for nsml use
bind_model(model, config)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=0.01, weight_decay=0.0001)
# DONOTCHANGE: They are reserved for nsml
if config.pause:
nsml.paused(scope=locals())
# 학습 모드일 때 사용합니다. (기본값)
if config.mode == 'train':
# 데이터를 로드합니다.
t0 = time.time()
dataset = MovieReviewDataset(DATASET_PATH, config.strmaxlen, max_size=config.max_dataset)
print("dataset loaded %.2f s" % (time.time() - t0))
pin_memory = USE_GPU > 0
if config.no_eval:
train_loader = DataLoader(dataset=dataset, batch_size=config.batch,
shuffle=True, collate_fn=collate_fn,
num_workers=2, pin_memory=pin_memory)
eval_loader = []
else:
train_sampler, eval_sampler = dataset.get_sampler()
train_loader = DataLoader(dataset=dataset, batch_size=config.batch,
sampler=train_sampler, collate_fn=collate_fn,
num_workers=2, pin_memory=pin_memory)
eval_loader = DataLoader(dataset=dataset, batch_size=config.batch,
sampler=eval_sampler, collate_fn=collate_fn,
num_workers=2, pin_memory=pin_memory)
lgb_model = object
models = (model, lgb_model, vect_word, vect_char)
# DONOTCHANGE: Reserved for nsml use
print("nsml binding...")
bind_model(models, config)
# DONOTCHANGE: They are reserved for nsml
if config.pause:
nsml.paused(scope=locals())
# 학습 모드일 때 사용합니다. (기본값)
if config.mode == 'train':
# 데이터를 로드합니다.
print("data loading...")
dataset = MovieReviewDataset(DATASET_PATH)
# X_trn, X_val, Y_trn, Y_val= trn_val_seperation(dataset, 144570)
X_trn, X_val, Y_trn, Y_val= trn_val_seperation(dataset, 3)
# Vectorizer를 학습합니다
vect_word, vect_char = vect_fit(X_trn, vect_word, vect_char)
# Text를 Vector화 합니다
X_trn = vect_transform(X_trn, vect_word, vect_char)
X_val = vect_transform(X_val, vect_word, vect_char)
#Dataset 구성
train_data = lgb.Dataset(X_trn, Y_trn)
valid_data = lgb.Dataset(X_val, Y_val, reference=train_data)
gc.collect()
word_embed: we,
sent_len: sl,
chars: cs,
word_len: wl,
sylls: ss,
y_: label,
is_training: train
}
# DONOTCHANGE: Reserved for nsml
if config.pause:
nsml.paused(scope=locals())
if config.mode == 'train':
# 데이터를 로드합니다.
dataset = MovieReviewDataset(DATASET_PATH, max_sent_len, max_word_len,
max_syll_num)
dataset_len = len(dataset)
one_batch_size = dataset_len // config.batch
if dataset_len % config.batch != 0:
one_batch_size += 1
if config.debug:
debugset = MovieReviewDataset(DEBUG3_PATH, max_sent_len,
max_word_len, max_syll_num)
debugset_len = len(debugset)
one_debug_size = debugset_len // config.batch
if debugset_len % config.batch != 0:
one_debug_size += 1
train_step = 0
best_ema = 99999.0
# DONOTCHANGE: They are reserved for nsml
if config.pause:
nsml.paused(scope=locals())
### Training mode
# 학습 모드일 때 사용합니다. (기본값)
if config.mode == 'train':
# 데이터를 로드합니다.
if not HAS_DATASET and not IS_ON_NSML: # It is not running on nsml
DATASET_PATH = '../sample_data/movie_review/'
corpus = DP.Corpus(DATASET_PATH, total_train)
print('[*]', 'Load corpus')
# Load training data
train_dataset = MovieReviewDataset(DATASET_PATH, config.strmaxlen,
True, corpus)
print('[*]', 'Load train dataset')
train_loader = DataLoader(dataset=train_dataset,
batch_size=config.batch,
shuffle=True,
collate_fn=collate_fn,
num_workers=1)
total_train = len(train_loader)
# Load validation data
test_dataset = MovieReviewDataset(DATASET_PATH,
config.strmaxlen,
True,
corpus,
test=True)
print('[*]', 'Load test dataset')
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
# DONOTCHANGE: Reserved for nsml
bind_model(sess=sess, config=config)
def get_feed_dict(w, c, s, y, train=False):
return {wx: w, cx_: c, sx_: s, y_: y, is_training: train}
# DONOTCHANGE: Reserved for nsml
if config.pause:
nsml.paused(scope=locals())
if config.mode == 'train':
# 데이터를 로드합니다.
dataset = MovieReviewDataset(DATASET_PATH, max_word_num, max_char_num,
max_syll_num)
dataset_len = len(dataset)
one_batch_size = dataset_len // config.batch
if dataset_len % config.batch != 0:
one_batch_size += 1
if config.debug:
debugset = MovieReviewDataset(DEBUG3_PATH, max_word_num,
max_char_num, max_syll_num)
debugset_len = len(debugset)
one_debug_size = debugset_len // config.batch
if debugset_len % config.batch != 0:
one_debug_size += 1
train_step = 0
best_ema = 99999.0
# Select model
args.add_argument('--model', type=str, default='SentpieceModel', choices=['SentpieceModel', 'regression', 'classification', 'bilstmwithattn', 'cnntext'])
config = args.parse_args()
print('HAS_DATASET :', HAS_DATASET)
print('IS_ON_NSML :', IS_ON_NSML)
print('DATASET_PATH :', DATASET_PATH)
print(config)
sp = []
wp_vocab = []
preprcess_infer = {}
if config.mode == 'train':
sp, wp_vocab = build_vocab(config.mode, DATASET_PATH, vocab_size=config.vocab_size)
dataset = MovieReviewDataset(DATASET_PATH, config.strmaxlen,
max_word_len=config.max_words_len, max_wp_len=config.max_wp_len, n_class=11)
vocab_size = len(dataset.i2wp)
else:
vocab_size = 19488 #19475
model_type = {
'SentpieceModel': SentpieceModel(vocab_size,
char_emb_size=config.char_embedding,
word_emb_size=config.word_embedding,
hidden_size=config.hidden_dim,
max_wp_len=config.max_wp_len,
max_words_len=config.max_words_len)
}
models = [
SentpieceModel(vocab_size,
for m in range(num_models):
models.append(Model(sess, "model" + str(m), config))
tf.global_variables_initializer().run()
# DONOTCHANGE: Reserved for nsml
bind_model(sess=sess, config=config, model=models)
# DONOTCHANGE: Reserved for nsml
if config.pause:
nsml.paused(scope=locals())
if config.mode == 'train':
# 데이터를 로드합니다.
dataset = MovieReviewDataset(DATASET_PATH,
config.strmaxlen,
is_train=True)
dataset_len = len(dataset)
one_batch_size = dataset_len // config.batch
if dataset_len % config.batch != 0:
one_batch_size += 1
# epoch마다 학습을 수행합니다.
for epoch in range(config.epochs):
#avg_loss = 0.0
avg_cost_list = np.zeros(len(models))
for i, (data,
labels) in enumerate(_batch_loader(dataset, config.batch)):
labels = np.reshape(labels, (-1, 1))
onehot_label = sess.run(
tf.reshape(tf.one_hot(labels, depth=11, dtype=tf.float32),
model = CNNReg(config.vocasize, config.embedding, config.maxlen, GPU_NUM)
if GPU_NUM: model = model.cuda()
# DONOTCHANGE: Reserved for nsml use
bind_model(model, config)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=config.lr)
# DONOTCHANGE: They are reserved for nsml
if config.pause: nsml.paused(scope=locals())
# 학습 모드일 때 사용합니다. (기본값)
if config.mode == 'train':
dataset = MovieReviewDataset(DATASET_PATH, config.vocasize,
config.minlen, config.maxlen)
train_loader = DataLoader(dataset=dataset,
batch_size=config.batch,
shuffle=True,
collate_fn=lambda data: zip(*data),
num_workers=2)
total_batch = len(train_loader)
# epoch마다 학습을 수행합니다.
for epoch in range(config.epochs):
avg_loss = 0.0
for i, (data, labels) in enumerate(train_loader):
predictions = model(data)
labels = Variable(torch.from_numpy(np.array(labels)))
if GPU_NUM: labels = labels.cuda()