def set_eval_inference_latency_mode(self):
"""
Evaluate Inference Latency Mode
- Pipeline
1. read raw_data (DataReader)
2. load vocabs from checkpoint (DataReader, Token)
3. define raw_to_tensor_fn (DataReader, Token)
4. define and load model
5. run!
"""
data_reader, token_makers = self._create_data_and_token_makers()
# Token & Vocab
vocabs = utils.load_vocabs(self.model_checkpoint)
for token_name, token_maker in token_makers.items():
token_maker.set_vocab(vocabs[token_name])
text_handler = TextHandler(token_makers, lazy_indexing=False)
_, helpers = data_reader.read()
raw_examples = helpers["valid"]["examples"]
cuda_device = self.config.cuda_devices[
0] if self.config.use_gpu else None
raw_to_tensor_fn = text_handler.raw_to_tensor_fn(
data_reader, cuda_device=cuda_device)
# Model
model = self._create_model(token_makers,
checkpoint=self.model_checkpoint)
self.set_trainer(model)
return raw_examples, raw_to_tensor_fn
def test(config):
NSML_SESSEION = 'team_6/19_tcls_qa/80' # NOTE: need to hard code
NSML_CHECKPOINT = '13800' # NOTE: nghhhhed to hard code
assert NSML_CHECKPOINT is not None, "You must insert NSML Session's checkpoint for submit"
assert NSML_SESSEION is not None, "You must insert NSML Session's name for submit"
set_global_seed(config.seed_num)
token_makers = create_by_factory(TokenMakersFactory, config.token)
tokenizers = token_makers["tokenizers"]
del token_makers["tokenizers"]
config.data_reader.tokenizers = tokenizers
data_reader = create_by_factory(DataReaderFactory, config.data_reader)
def bind_load_vocabs(config, token_makers):
CHECKPOINT_FNAME = "checkpoint.bin"
def load(dir_path):
checkpoint_path = os.path.join(dir_path, CHECKPOINT_FNAME)
checkpoint = torch.load(checkpoint_path)
vocabs = {}
token_config = config.token
for token_name in token_config.names:
token = getattr(token_config, token_name, {})
vocab_config = getattr(token, "vocab", {})
texts = checkpoint["vocab_texts"][token_name]
if type(vocab_config) != dict:
vocab_config = vars(vocab_config)
vocabs[token_name] = Vocab(token_name,
**vocab_config).from_texts(texts)
for token_name, token_maker in token_makers.items():
token_maker.set_vocab(vocabs[token_name])
return token_makers
nsml.bind(load=load)
bind_load_vocabs(config, token_makers)
nsml.load(checkpoint=NSML_CHECKPOINT, session=NSML_SESSEION)
# Raw to Tensor Function
text_handler = TextHandler(token_makers, lazy_indexing=False)
raw_to_tensor_fn = text_handler.raw_to_tensor_fn(
data_reader,
cuda_device=device,
)
# Model & Optimizer
model = create_model(token_makers, ModelFactory, config.model, device)
trainer = Trainer(model, metric_key="f1")
if nsml.IS_ON_NSML:
bind_nsml(model, trainer=trainer, raw_to_tensor_fn=raw_to_tensor_fn)
if config.nsml.pause:
nsml.paused(scope=locals())
def set_predict_mode(self, preload=False):
"""
Predict Mode
- Pipeline
1. read raw_data (Argument)
2. load vocabs from checkpoint (DataReader, Token)
3. define raw_to_tensor_fn (DataReader, Token)
4. define and load model
5. run!
"""
data_reader, token_makers = self._create_data_and_token_makers()
# Token & Vocab
vocabs = utils.load_vocabs(self.model_checkpoint)
for token_name, token_maker in token_makers.items():
token_maker.set_vocab(vocabs[token_name])
text_handler = TextHandler(token_makers, lazy_indexing=False)
# Set predict config
if self.argument.interactive:
raw_features = {
feature_name: ""
for feature_name in data_reader.text_columns
}
else:
raw_features = {}
for feature_name in data_reader.text_columns:
feature = getattr(self.argument, feature_name, None)
# if feature is None:
# raise ValueError(f"--{feature_name} argument is required!")
raw_features[feature_name] = feature
cuda_device = self.config.cuda_devices[
0] if self.config.use_gpu else None
raw_to_tensor_fn = text_handler.raw_to_tensor_fn(
data_reader,
cuda_device=cuda_device,
helper=self.model_checkpoint.get("predict_helper", {}))
# Model
model = self._create_model(token_makers,
checkpoint=self.model_checkpoint)
self.set_trainer(model)
arguments = vars(self.argument)
if preload:
self.predict_settings = {
"raw_to_tensor_fn": raw_to_tensor_fn,
"arguments": arguments
}
else:
return raw_features, raw_to_tensor_fn, arguments
def set_eval_mode(self):
"""
Evaluate Mode
- Pipeline
1. read raw_data (DataReader)
2. load vocabs from checkpoint (DataReader, Token)
3. indexing tokens (DataReader, Token)
4. convert to DataSet (DataReader)
5. create DataLoader (DataLoader)
6. define and load model
7. run!
"""
data_reader, token_makers = self._create_data_and_token_makers()
# DataReader
datas, helpers = data_reader.read()
# Token & Vocab
vocabs = utils.load_vocabs(self.model_checkpoint)
for token_name, token_maker in token_makers.items():
token_maker.set_vocab(vocabs[token_name])
text_handler = TextHandler(token_makers, lazy_indexing=False)
text_handler.index(datas, data_reader.text_columns)
# iterator
vocab = vocabs[next(iter(vocabs))]
datasets = data_reader.convert_to_dataset(datas,
vocab,
helpers=helpers) # with name
self.config.iterator.cuda_devices = self.config.cuda_devices
_, valid_loader, _ = self._create_by_factory(
DataLoaderFactory,
self.config.iterator,
param={"datasets": datasets})
# Model
model = self._create_model(token_makers,
checkpoint=self.model_checkpoint)
self.set_trainer(model)
return valid_loader
def train_and_evaluate(config):
token_makers = create_by_factory(TokenMakersFactory, config.token)
tokenizers = token_makers["tokenizers"]
del token_makers["tokenizers"]
config.data_reader.tokenizers = tokenizers
if nsml.IS_ON_NSML:
config.data_reader.train_file_path = os.path.join(
DATASET_PATH, "train", "train_data",
config.data_reader.train_file_path)
config.data_reader.valid_file_path = os.path.join(
DATASET_PATH, "train", "train_data",
config.data_reader.valid_file_path)
data_reader = create_by_factory(DataReaderFactory, config.data_reader)
datas, helpers = data_reader.read()
# Vocab & Indexing
text_handler = TextHandler(token_makers, lazy_indexing=True)
texts = data_reader.filter_texts(datas)
token_counters = text_handler.make_token_counters(texts)
text_handler.build_vocabs(token_counters)
text_handler.index(datas, data_reader.text_columns)
# Iterator
datasets = data_reader.convert_to_dataset(datas, helpers=helpers)
train_loader = create_data_loader(datasets["train"],
batch_size=config.iterator.batch_size,
shuffle=True,
cuda_device_id=device)
valid_loader = create_data_loader(datasets["valid"],
batch_size=config.iterator.batch_size,
shuffle=False,
cuda_device_id=device)
# Model & Optimizer
model = create_model(token_makers,
ModelFactory,
config.model,
device,
helpers=helpers)
model_parameters = [
param for param in model.parameters() if param.requires_grad
]
optimizer = get_optimizer_by_name("adam")(model_parameters)
if IS_ON_NSML:
bind_nsml(model, optimizer=optimizer)
# Trainer
trainer_config = vars(config.trainer)
trainer_config["model"] = model
trainer = Trainer(**trainer_config)
trainer.train_and_evaluate(train_loader, valid_loader, optimizer)
def set_train_mode(self):
"""
Training Mode
- Pipeline
1. read raw_data (DataReader)
2. build vocabs (DataReader, Token)
3. indexing tokens (DataReader, Token)
4. convert to DataSet (DataReader)
5. create DataLoader (DataLoader)
6. define model and optimizer
7. run!
"""
logger.info("Config. \n" + pretty_json_dumps(self.config_dict) + "\n")
data_reader, token_makers = self._create_data_and_token_makers()
datas, helpers = data_reader.read()
# Token & Vocab
text_handler = TextHandler(token_makers, lazy_indexing=True)
texts = data_reader.filter_texts(datas)
token_counters = text_handler.make_token_counters(texts, config=self.config)
text_handler.build_vocabs(token_counters)
text_handler.index(datas, data_reader.text_columns)
# iterator
datasets = data_reader.convert_to_dataset(datas, helpers=helpers) # with name
self.config.iterator.cuda_devices = self.config.cuda_devices
train_loader, valid_loader, test_loader = self._create_by_factory(
DataLoaderFactory, self.config.iterator, param={"datasets": datasets}
)
checkpoint_dir = Path(self.config.trainer.log_dir) / "checkpoint"
checkpoints = None
if checkpoint_dir.exists():
checkpoints = self._load_exist_checkpoints(checkpoint_dir) # contain model and optimizer
if checkpoints is None:
model = self._create_model(token_makers, helpers=helpers)
op_dict = self._create_by_factory(
OptimizerFactory, self.config.optimizer, param={"model": model}
)
else:
model = self._create_model(token_makers, checkpoint=checkpoints)
op_dict = self._create_by_factory(
OptimizerFactory, self.config.optimizer, param={"model": model}
)
utils.load_optimizer_checkpoint(op_dict["optimizer"], checkpoints)
self.set_trainer(model, op_dict=op_dict)
return train_loader, valid_loader, op_dict["optimizer"]
def re_train_and_evaluate(config):
NSML_SESSEION = 'team_6/19_tcls_qa/258' # NOTE: need to hard code
NSML_CHECKPOINT = '1' # NOTE: nghhhhed to hard code
assert NSML_CHECKPOINT is not None, "You must insert NSML Session's checkpoint for submit"
assert NSML_SESSEION is not None, "You must insert NSML Session's name for submit"
token_makers = create_by_factory(TokenMakersFactory, config.token)
tokenizers = token_makers["tokenizers"]
del token_makers["tokenizers"]
config.data_reader.tokenizers = tokenizers
if nsml.IS_ON_NSML:
config.data_reader.train_file_path = os.path.join(
DATASET_PATH, "train", "train_data",
config.data_reader.train_file_path)
config.data_reader.valid_file_path = os.path.join(
DATASET_PATH, "train", "train_data",
config.data_reader.valid_file_path)
data_reader = create_by_factory(DataReaderFactory, config.data_reader)
datas, helpers = data_reader.read()
# Vocab & Indexing
text_handler = TextHandler(token_makers, lazy_indexing=True)
texts = data_reader.filter_texts(datas)
token_counters = text_handler.make_token_counters(texts)
text_handler.build_vocabs(token_counters)
text_handler.index(datas, data_reader.text_columns)
def bind_load_vocabs(config, token_makers):
CHECKPOINT_FNAME = "checkpoint.bin"
def load(dir_path):
checkpoint_path = os.path.join(dir_path, CHECKPOINT_FNAME)
checkpoint = torch.load(checkpoint_path)
vocabs = {}
token_config = config.token
for token_name in token_config.names:
token = getattr(token_config, token_name, {})
vocab_config = getattr(token, "vocab", {})
texts = checkpoint["vocab_texts"][token_name]
if type(vocab_config) != dict:
vocab_config = vars(vocab_config)
vocabs[token_name] = Vocab(token_name,
**vocab_config).from_texts(texts)
for token_name, token_maker in token_makers.items():
token_maker.set_vocab(vocabs[token_name])
return token_makers
nsml.bind(load=load)
bind_load_vocabs(config, token_makers)
nsml.load(checkpoint=NSML_CHECKPOINT, session=NSML_SESSEION)
# Raw to Tensor Function
text_handler = TextHandler(token_makers, lazy_indexing=False)
raw_to_tensor_fn = text_handler.raw_to_tensor_fn(
data_reader,
cuda_device=device,
)
# Iterator
datasets = data_reader.convert_to_dataset(datas, helpers=helpers)
train_loader = create_data_loader(datasets["train"],
batch_size=config.iterator.batch_size,
shuffle=True,
cuda_device_id=device)
valid_loader = create_data_loader(datasets["valid"],
batch_size=config.iterator.batch_size,
shuffle=False,
cuda_device_id=device)
# Model & Optimizer
model = create_model(token_makers,
ModelFactory,
config.model,
device,
helpers=helpers)
model_parameters = [
param for param in model.parameters() if param.requires_grad
]
optimizer = get_optimizer_by_name("adam")(model_parameters)
def bind_load_model(config, model, **kwargs):
CHECKPOINT_FNAME = "checkpoint.bin"
def load(dir_path):
checkpoint_path = os.path.join(dir_path, CHECKPOINT_FNAME)
checkpoint = torch.load(checkpoint_path)
model.load_state_dict(checkpoint["weights"])
model.config = checkpoint["config"]
model.metrics = checkpoint["metrics"]
model.init_params = checkpoint["init_params"],
model.predict_helper = checkpoint["predict_helper"],
model.train_counter = TrainCounter(display_unit=1000)
# model.vocabs = load_vocabs(checkpoint)
if "optimizer" in kwargs:
kwargs["optimizer"].load_state_dict(checkpoint["optimizer"][0])
print(f"Model reload checkpoints...! {checkpoint_path}")
nsml.bind(load=load)
bind_load_model(config, model, optimizer=optimizer)
nsml.load(checkpoint=NSML_CHECKPOINT, session=NSML_SESSEION)
if IS_ON_NSML:
bind_nsml(model, optimizer=optimizer)
# Trainer
trainer_config = vars(config.trainer)
trainer_config["model"] = model
trainer = Trainer(**trainer_config)
trainer.train_and_evaluate(train_loader, valid_loader, optimizer)
