def main(gpu, args):
rank = args.nr * args.gpus + gpu
dist.init_process_group(backend='nccl',
init_method='env://',
world_size=args.world_size,
rank=rank)
with open(args.config) as config_file:
params = json.load(config_file)
params['nodes'] = args.nodes
params['gpus'] = args.gpus
params['nr'] = args.nr
params['world_size'] = args.world_size
params['rank'] = rank
random.seed(params['seed'])
np.random.seed(params['seed'])
torch.manual_seed(params['seed'])
preproc = Preprocessor(sp_model=params['sp_model'])
params['vocab_size'] = preproc.vocab_size
logger = Logger(params['logfile'])
logger.write('Loading data...')
train_dataloader, eval_dataloader = prepare_dataset(
dataset_dir=params['dataset_dir'],
train_sets=params['train_sets'],
eval_sets=params['eval_sets'],
batch_size=params['batch_size'],
preproc=preproc,
world_size=args.world_size,
rank=rank)
logger.write('Model initialization...')
model, num_params = make_model(params)
logger.write(f'Total number of parameters: {num_params}')
torch.cuda.set_device(gpu)
model.cuda()
optimizer = torch.optim.Adadelta(model.parameters(),
lr=params['lr'],
eps=1e-8,
rho=0.95,
weight_decay=params['weight_decay'])
model, optimizer = amp.initialize(min_loss_scale=1.0,
models=model,
optimizers=optimizer,
opt_level=params['opt_level'])
model = DDP(model)
train(data=[train_dataloader, eval_dataloader, preproc],
model=model,
optimizer=optimizer,
logger=logger,
params=params)
def fit(self, data):
"""Fit the model to the given data.
Args:
data(pandas.DataFrame): dataset to fit the model.
Returns:
None
"""
self.preprocessor = Preprocessor(
continuous_columns=self.continuous_columns)
data = self.preprocessor.fit_transform(data)
self.metadata = self.preprocessor.metadata
dataflow = TGANDataFlow(data, self.metadata)
batch_data = BatchData(dataflow, self.batch_size)
input_queue = QueueInput(batch_data)
self.model = self.get_model(training=True)
if self.trainer == 'GANTrainer':
trainer = GANTrainer(model=self.model, input_queue=input_queue)
elif self.trainer == 'SeparateGANTrainer':
trainer = SeparateGANTrainer(model=self.model,
input_queue=input_queue)
else:
raise ValueError(
'Incorrect trainer name. Use GANTrainer or SeparateGANTrainer')
# trainer = SeparateGANTrainer(model=self.model, input_queue=input_queue)
self.restore_path = os.path.join(self.model_dir, 'checkpoint')
if os.path.isfile(self.restore_path) and self.restore_session:
session_init = SaverRestore(self.restore_path)
with open(os.path.join(self.log_dir, 'stats.json')) as f:
starting_epoch = json.load(f)[-1]['epoch_num'] + 1
else:
session_init = None
starting_epoch = 1
action = 'k' if self.restore_session else None
logger.set_logger_dir(self.log_dir, action=action)
callbacks = []
if self.save_checkpoints:
callbacks.append(ModelSaver(checkpoint_dir=self.model_dir))
trainer.train_with_defaults(callbacks=callbacks,
steps_per_epoch=self.steps_per_epoch,
max_epoch=self.max_epoch,
session_init=session_init,
starting_epoch=starting_epoch)
self.prepare_sampling()
def train(args):
if args.seed is not None:
utils.random.seed_everything(args.seed)
device = torch.device("cuda" if args.cuda else "cpu")
preprocessor = Preprocessor()
preprocessor.build_vocab(args.train_file, cache_dir=args.cache_dir)
if args.embed_file:
preprocessor.load_embeddings(args.embed_file, cache_dir=args.cache_dir)
loader_config = dict(
preprocessor=preprocessor,
batch_size=args.batch_size,
device=device,
cache_dir=args.cache_dir,
)
train_dataloader = create_dataloader(args.train_file, **loader_config, shuffle=True)
eval_dataloader = None
if args.eval_file:
eval_dataloader = create_dataloader(args.eval_file, **loader_config, shuffle=False)
model = build_model(
word_vocab_size=len(preprocessor.vocabs["word"]),
pretrained_word_vocab_size=len(preprocessor.vocabs["pretrained_word"]),
postag_vocab_size=len(preprocessor.vocabs["postag"]),
pretrained_word_embeddings=preprocessor.pretrained_word_embeddings,
n_deprels=len(preprocessor.vocabs["deprel"]),
)
model.to(device)
trainer = create_trainer(
model, lr=args.learning_rate, max_steps=args.max_steps, eval_interval=args.eval_interval
)
trainer.add_callback(utils.training.PrintCallback(printer=logger.info))
if eval_dataloader:
deprel_map = {v: k for k, v in preprocessor.vocabs["deprel"].mapping.items()}
trainer.add_callback(EvaluateCallback(args.eval_file, deprel_map), priority=0)
if args.save_dir:
torch.save(preprocessor, os.path.join(args.save_dir, "preprocessor.pt"))
trainer.add_callback(
utils.training.SaveCallback(args.save_dir, monitor="eval/UAS", mode="max")
)
with logging_redirect_tqdm(loggers=[logger]):
trainer.fit(train_dataloader, eval_dataloader)
def single_experiment(args, i, logger):
print(f'\033[91m======================{args.dataset} exp: {i}====================\033[00m')
np.random.seed(int(args.seed + i)) # for the reproducibility
random.seed(int(args.seed + i))
torch.manual_seed(int(args.seed + i))
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
args.seed += 1
server = Server(args, logger) # 서버와 로컬이 만들어짐
data = Preprocessor(args) # Data 불러오는 곳
data.distribute_data(server) # 서버에 데이터를 전달함
server.train(exp_id=i) # 서버 Training 진행함
print("\033[91m=\033[00m" * 50 + "\n")
ret_model = server.get_global_model()
del data
del server
gc.collect()
torch.cuda.empty_cache()
logger.save_data()
return ret_model
def preprocess_SNLI_data(inputdir,
embeddings_file,
targetdir,
lowercase=False,
ignore_punctuation=False,
num_words=None,
stopwords=[],
labeldict={},
bos=None,
eos=None):
"""
Preprocess the data from the SNLI corpus so it can be used by the
ESIM model.
Compute a worddict from the train set, and transform the words in
the sentences of the corpus to their indices, as well as the labels.
Build an embedding matrix from pretrained word vectors.
The preprocessed data is saved in pickled form in some target directory.
Args:
inputdir: The path to the directory containing the NLI corpus.
embeddings_file: The path to the file containing the pretrained
word vectors that must be used to build the embedding matrix.
targetdir: The path to the directory where the preprocessed data
must be saved.
lowercase: Boolean value indicating whether to lowercase the premises
and hypotheseses in the input data. Defautls to False.
ignore_punctuation: Boolean value indicating whether to remove
punctuation from the input data. Defaults to False.
num_words: Integer value indicating the size of the vocabulary to use
for the word embeddings. If set to None, all words are kept.
Defaults to None.
stopwords: A list of words that must be ignored when preprocessing
the data. Defaults to an empty list.
bos: A string indicating the symbol to use for beginning of sentence
tokens. If set to None, bos tokens aren't used. Defaults to None.
eos: A string indicating the symbol to use for end of sentence tokens.
If set to None, eos tokens aren't used. Defaults to None.
"""
if not os.path.exists(targetdir):
os.makedirs(targetdir)
# Retrieve the train, dev and test data files from the dataset directory.
train_file = ""
dev_file = ""
test_file = ""
for file in os.listdir(inputdir):
if fnmatch.fnmatch(file, "*_train.txt"):
train_file = file
elif fnmatch.fnmatch(file, "*_dev.txt"):
dev_file = file
elif fnmatch.fnmatch(file, "*_test.txt"):
test_file = file
# -------------------- Train data preprocessing -------------------- #
preprocessor = Preprocessor(lowercase=lowercase,
ignore_punctuation=ignore_punctuation,
num_words=num_words,
stopwords=stopwords,
labeldict=labeldict,
bos=bos,
eos=eos)
print(20*"=", " Preprocessing train set ", 20*"=")
print("\t* Reading data...")
data = preprocessor.read_data(os.path.join(inputdir, train_file))
print("\t* Computing worddict and saving it...")
preprocessor.build_worddict(data)
with open(os.path.join(targetdir, "worddict.pkl"), "wb") as pkl_file:
pickle.dump(preprocessor.worddict, pkl_file)
print("\t* Transforming words in premises and hypotheses to indices...")
transformed_data = preprocessor.transform_to_indices(data)
print("\t* Saving result...")
with open(os.path.join(targetdir, "train_data.pkl"), "wb") as pkl_file:
pickle.dump(transformed_data, pkl_file)
# -------------------- Validation data preprocessing -------------------- #
print(20*"=", " Preprocessing dev set ", 20*"=")
print("\t* Reading data...")
data = preprocessor.read_data(os.path.join(inputdir, dev_file))
print("\t* Transforming words in premises and hypotheses to indices...")
transformed_data = preprocessor.transform_to_indices(data)
print("\t* Saving result...")
with open(os.path.join(targetdir, "dev_data.pkl"), "wb") as pkl_file:
pickle.dump(transformed_data, pkl_file)
# -------------------- Test data preprocessing -------------------- #
print(20*"=", " Preprocessing test set ", 20*"=")
print("\t* Reading data...")
data = preprocessor.read_data(os.path.join(inputdir, test_file))
print("\t* Transforming words in premises and hypotheses to indices...")
transformed_data = preprocessor.transform_to_indices(data)
print("\t* Saving result...")
with open(os.path.join(targetdir, "test_data.pkl"), "wb") as pkl_file:
pickle.dump(transformed_data, pkl_file)
# -------------------- Embeddings preprocessing -------------------- #
print(20*"=", " Preprocessing embeddings ", 20*"=")
print("\t* Building embedding matrix and saving it...")
embed_matrix = preprocessor.build_embedding_matrix(embeddings_file)
with open(os.path.join(targetdir, "embeddings.pkl"), "wb") as pkl_file:
pickle.dump(embed_matrix, pkl_file)
from data import Data, Preprocessor
import sys
input_file = './data/chat-200w.txt'
output_file = './data/combine.txt'
stop_words_file = './data/stop_words.txt'
num = None
if __name__ == '__main__':
data_loader = Data(input_file, num)
data_loader.seperate_conversation()
preprocessor = Preprocessor(stop_words_file)
data_loader.preprocess(preprocessor, format=3)
data = data_loader.conversation_QAQAQ_preprocessed
f = open(output_file, 'w')
sys.stdout = f
for i, conv in enumerate(data):
print('Conversation %d:' % i)
for sen in conv:
if not sen[1]:
sen[1] = ['EMOJI']
print(str(sen[0]) + ' ' + ''.join(sen[1]))
print()
f.close()