def test_no_torch_utils_tensorboard_package(dirname):
from tensorboardX import SummaryWriter
with patch.dict("sys.modules", {"torch.utils.tensorboard": None}):
tb_logger = TensorboardLogger(log_dir=dirname)
assert isinstance(tb_logger.writer, SummaryWriter), type(tb_logger.writer)
tb_logger.close()
def add_tensorboard(engine_train, optimizer, model, log_dir):
"""Creates an ignite logger object and adds training elements such as weight and gradient histograms
Args:
engine_train (:obj:`ignite.engine`): the train engine to attach to the logger
optimizer (:obj:`torch.optim`): the model's optimizer
model (:obj:`torch.nn.Module`): the model being trained
log_dir (string): path to where tensorboard data should be saved
"""
# Create a logger
tb_logger = TensorboardLogger(log_dir=log_dir)
# Attach the logger to the trainer to log training loss at each iteration
tb_logger.attach(engine_train,
log_handler=OutputHandler(
tag="training",
output_transform=lambda loss: {"loss": loss}),
event_name=Events.ITERATION_COMPLETED)
# Attach the logger to the trainer to log optimizer's parameters, e.g. learning rate at each iteration
tb_logger.attach(engine_train,
log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.EPOCH_COMPLETED)
# Attach the logger to the trainer to log model's weights as a histogram after each epoch
tb_logger.attach(engine_train,
log_handler=WeightsHistHandler(model),
event_name=Events.EPOCH_COMPLETED)
# Attach the logger to the trainer to log model's gradients as a histogram after each epoch
tb_logger.attach(engine_train,
log_handler=GradsHistHandler(model),
event_name=Events.EPOCH_COMPLETED)
tb_logger.close()
def setup_training_loggers(name, level=logging.INFO):
log_setup = logging.getLogger(name)
os.makedirs(os.path.join(cfg.tb_log_folder, name), exist_ok=True)
tb_logger = TensorboardLogger(os.path.join(cfg.tb_log_folder, name))
os.makedirs(os.path.join(cfg.checkpoint_log_folder, name), exist_ok=True)
checkpoint_handler = ModelCheckpoint(os.path.join(
cfg.checkpoint_log_folder, name),
'checkpoint',
save_interval=1,
n_saved=3,
require_empty=False)
if len(log_setup.handlers) == 2: # Logger already set up for current run
return log_setup, tb_logger, checkpoint_handler
fileHandler = logging.FileHandler(os.path.join(cfg.validation_log_folder,
name + '.log'),
mode='a')
formatter = logging.Formatter('%(levelname)s: %(asctime)s %(message)s',
datefmt='%m/%d/%Y %I:%M:%S %p')
fileHandler.setFormatter(formatter)
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(formatter)
log_setup.setLevel(level)
log_setup.addHandler(fileHandler)
log_setup.addHandler(consoleHandler)
return logging.getLogger(name), tb_logger, checkpoint_handler
def get_tensorboard_logger(trainer: Engine, evaluators: ThreeEvaluators,
metric_names: List[str]) -> TensorboardLogger:
"""
creates a ``tensorboard`` logger which read metrics from given evaluators and attaches it to a given trainer
:param trainer: an ``ignite`` trainer to attach to
:param evaluators: a triple of train, validation, and test evaluators to get metrics from
:param metric_names: a list of metrics to log during validation and testing
"""
tb_logger = TensorboardLogger(log_dir=f"runs/{datetime.now()}",
flush_secs=1)
training_loss = OutputHandler(
"training",
["running_loss"],
global_step_transform=global_step_from_engine(trainer),
)
tb_logger.attach(trainer, training_loss, Events.EPOCH_COMPLETED)
validation_loss = OutputHandler(
"validation",
metric_names,
global_step_transform=global_step_from_engine(trainer),
)
tb_logger.attach(evaluators.validation, validation_loss, Events.COMPLETED)
test_loss = OutputHandler(
"test",
metric_names,
global_step_transform=global_step_from_engine(trainer),
)
tb_logger.attach(evaluators.test, test_loss, Events.COMPLETED)
return tb_logger
def add_logging_and_checkpoint_saving(trainer, evaluator, metrics, model, optimizer, args, prefix=""):
""" Add to training engine tensorboard logging, progress bar with average loss, checkpoint saving and save training config. """
# Add progress bar with average loss
RunningAverage(output_transform=lambda x: x).attach(trainer, prefix + "loss")
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=[prefix + "loss"])
evaluator.add_event_handler(Events.COMPLETED,
lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics)))
# Add tensorboard logging with training and evaluation metrics
tb_logger = TensorboardLogger(log_dir=None)
tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=[prefix + "loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.ITERATION_STARTED)
@evaluator.on(Events.COMPLETED)
def tb_log_metrics(engine):
for name in metrics.keys():
tb_logger.writer.add_scalar(name, engine.state.metrics[name], trainer.state.iteration)
# Add checkpoint saving after each epoch - take care of distributed encapsulation ('getattr()')
checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir, 'checkpoint', save_interval=1, n_saved=3)
trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {'mymodel': getattr(model, 'module', model)})
# Save training configuration
torch.save(args, os.path.join(tb_logger.writer.log_dir, CONFIG_NAME))
return checkpoint_handler, tb_logger
def test_integration_as_context_manager(dirname):
n_epochs = 5
data = list(range(50))
losses = torch.rand(n_epochs * len(data))
losses_iter = iter(losses)
def update_fn(engine, batch):
return next(losses_iter)
with TensorboardLogger(log_dir=dirname) as tb_logger:
trainer = Engine(update_fn)
def dummy_handler(engine, logger, event_name):
global_step = engine.state.get_event_attrib_value(event_name)
logger.writer.add_scalar("test_value", global_step, global_step)
tb_logger.attach(trainer,
log_handler=dummy_handler,
event_name=Events.EPOCH_COMPLETED)
trainer.run(data, max_epochs=n_epochs)
# Check if event files are present
written_files = os.listdir(dirname)
written_files = [f for f in written_files if "tfevents" in f]
assert len(written_files) > 0
def custom_setup(self):
if self.tensorboard_logs:
tb_logger = TensorboardLogger(log_dir=self.tensorboard_logs)
tb_logger.attach(self.trainer,
log_handler=OutputHandler(
tag="training",
output_transform=lambda loss: {'loss': loss}),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(self.evaluator,
log_handler=OutputHandler(
tag="validation",
metric_names=["LossMetric"],
another_engine=self.trainer),
event_name=Events.EPOCH_COMPLETED)
if self.optional_tensorboard_features:
tb_logger.attach(self.trainer,
log_handler=OptimizerParamsHandler(
self.optimizer),
event_name=Events.ITERATION_STARTED)
tb_logger.attach(self.trainer,
log_handler=WeightsScalarHandler(self.model),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(self.trainer,
log_handler=WeightsHistHandler(self.model),
event_name=Events.EPOCH_COMPLETED)
tb_logger.attach(self.trainer,
log_handler=GradsScalarHandler(self.model),
event_name=Events.ITERATION_COMPLETED)
# This is important to close the tensorboard file logger
@self.trainer.on(Events.COMPLETED)
def end_tensorboard(trainer):
logger.info("Training completed")
tb_logger.close()
if self.embeddings_name:
@self.trainer.on(Events.COMPLETED)
def log_embeddings(trainer):
if hasattr(self.model, self.embeddings_name) and hasattr(
self.dataset_splits, "vectorizer") and TENSORBOARD:
logger.info(
f"Logging embeddings ({self.embeddings_name}) to Tensorboard!"
)
embeddings = getattr(self.model,
self.embeddings_name).weight.data
metadata = [
str(self.dataset_splits.vectorizer.data_vocab.
_id2token[token_index]).encode('utf-8')
for token_index in range(embeddings.shape[0])
]
self.writer.add_embedding(
mat=embeddings,
metadata=metadata,
global_step=self.trainer.state.epoch)
def test_no_tensorboardX_nor_torch_utils_tensorboard():
with patch.dict("sys.modules", {
"tensorboardX": None,
"torch.utils.tensorboard": None
}):
with pytest.raises(
RuntimeError,
match=
r"This contrib module requires either tensorboardX or torch"):
TensorboardLogger(log_dir=None)
def setup_ignite(engine: Engine,
params: SimpleNamespace,
exp_source,
run_name: str,
extra_metrics: Iterable[str] = ()):
simplefilter("ignore", category=UserWarning)
handler = EndOfEpisodeHandler(exp_source,
bound_avg_reward=params.stop_reward)
handler.attach(engine)
EpisodeFPSHandler().attach(engine)
@engine.on(EpisodeEvents.EPISODE_COMPLETED)
def episode_completed(trainer: Engine):
passed = trainer.state.metrics.get("time_passed", 0)
print(
"Episode {}: reward={:.0f}, steps={}, speed={:.1f} f/s, elapsed={}"
.format(trainer.state.episode, trainer.state.episode_reward,
trainer.state.episode_steps,
trainer.state.metrics.get("avg_fps", 0),
timedelta(seconds=int(passed))))
@engine.on(EpisodeEvents.BOUND_REWARD_REACHED)
def game_solved(trainer: Engine):
passed = trainer.state.metrics["time_passed"]
print(
f"Game solved in {timedelta(seconds=int(passed))} after {trainer.state.episode}"
f" episodes and {trainer.state.iteration} iterations!")
trainer.should_terminate = True
now = datetime.now().isoformat(timespec="minutes").replace(":", "-")
logdir = f"runs/{now}-{params.run_name}-{run_name}"
tb = TensorboardLogger(log_dir=logdir)
run_avg = RunningAverage(output_transform=lambda v: v["loss"])
run_avg.attach(engine, "avg_loss")
metrics = ["reward", "steps", "avg_reward"]
handler = OutputHandler(tag="episodes", metric_names=metrics)
event = EpisodeEvents.EPISODE_COMPLETED
tb.attach(engine, log_handler=handler, event_name=event)
# write to tensorboard every 100 iterations
PeriodicEvents().attach(engine)
metrics = ["avg_loss", "avg_fps"]
metrics.extend(extra_metrics)
handler = OutputHandler(tag="train",
metric_names=metrics,
output_transform=lambda a: a)
event = PeriodEvents.ITERS_100_COMPLETED
tb.attach(engine, log_handler=handler, event_name=event)
def train():
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
model = Bert_SQG()
optimizer = AdamW(model.parameters(), lr=3e-5)
ds = dataloader.BertSQG_DataClass()
dl = DataLoader(ds, num_workers=4, batch_size=4)
scheduler = PiecewiseLinear(optimizer, "lr",
[(0, 3e-5),
(EPOCHS * len(ds) // BATCH_SIZE, 0.0)])
metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1))}
def update(engine, batch):
model.train()
for i in range(0, len(batch) - 1):
x = batch[i].to(device)
y = batch[i + 1].to(device)
y_prime = model(x)
loss = criterion(y_prime[-1], y[-1]) / ITERATION_STEP
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
if engine.state.iteration % ITERATION_STEP == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item()
trainer = Engine(update)
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=["loss"])
tb_logger = TensorboardLogger(log_dir='./logs')
tb_logger.attach(trainer,
log_handler=OutputHandler(tag="training",
metric_names=["loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer,
log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.ITERATION_STARTED)
checkpoint_handler = ModelCheckpoint('./checkpoint',
'_checkpoint',
save_interval=1,
n_saved=3)
trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler,
{'bert_sqg': getattr(model, 'module', model)})
trainer.run(dl, max_epochs=EPOCHS)
tb_loger.close()
def add_tensorboard_logging(self, logging_dir=None):
# Add TensorBoard logging
if logging_dir is None:
os.path.join(self.config.DIRS.WORKING_DIR, 'tb_logs')
else:
os.path.join(logging_dir, 'tb_logs')
print('Tensorboard logging saving to:: {} ...'.format(logging_dir),
end='')
self.tb_logger = TensorboardLogger(log_dir=logging_dir)
# Logging iteration loss
self.tb_logger.attach_output_handler(
engine=self.train_engine,
event_name=Events.ITERATION_COMPLETED,
tag='training',
output_transform=lambda loss: {"batch loss": loss})
# Logging epoch training metrics
self.tb_logger.attach_output_handler(
engine=self.train_evaluator,
event_name=Events.EPOCH_COMPLETED,
tag="training",
metric_names=[
"loss", "accuracy", "precision", "recall", "f1", "topKCatAcc"
],
global_step_transform=global_step_from_engine(self.train_engine),
)
# Logging epoch validation metrics
self.tb_logger.attach_output_handler(
engine=self.evaluator,
event_name=Events.EPOCH_COMPLETED,
tag="validation",
metric_names=[
"loss", "accuracy", "precision", "recall", "f1", "topKCatAcc"
],
global_step_transform=global_step_from_engine(self.train_engine),
)
# Attach the logger to the trainer to log model's weights as a histogram after each epoch
self.tb_logger.attach(self.train_engine,
event_name=Events.EPOCH_COMPLETED,
log_handler=WeightsHistHandler(self.model))
# Attach the logger to the trainer to log model's gradients as a histogram after each epoch
self.tb_logger.attach(self.train_engine,
event_name=Events.EPOCH_COMPLETED,
log_handler=GradsHistHandler(self.model))
print('Tensorboard Logging...', end='')
print('done')
def __init__(self, name, model, log_dir, lr, lr_decay_step, adam=False):
"""
Initialize to train the given model.
:param name: The name of the model to be trained.
:param model: The model to be trained.
:param log_dir: String. The log directory of the tensorboard.
:param lr: Float. The learning rate.
:param lr_decay_step: Integer. The amount of steps the learning rate decays.
:param adam: Bool. Whether to use adam optimizer or not.
"""
super(Trainer, self).__init__(self.update_model)
self.model = model
# tqdm
ProgressBar(persist=True).attach(self)
# Optimizer
params = [p for p in model.parameters() if p.requires_grad]
if adam:
self.optimizer = torch.optim.Adam(params, lr=lr)
else:
self.optimizer = torch.optim.SGD(params, lr=lr, momentum=0.9)
# Scheduler
if lr_decay_step > 0:
self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, step_size=lr_decay_step, gamma=0.1)
self.add_event_handler(Events.EPOCH_COMPLETED, lambda e: e.scheduler.step())
else:
self.scheduler = None
# Terminate if nan values found
self.add_event_handler(Events.ITERATION_COMPLETED, TerminateOnNan())
# Tensorboard logging
self.tb_logger = TensorboardLogger(log_dir=os.path.join(log_dir, name))
self.add_event_handler(Events.COMPLETED, lambda x: self.tb_logger.close())
self.tb_logger.attach(self,
log_handler=OptimizerParamsHandler(self.optimizer),
event_name=Events.EPOCH_COMPLETED)
self.tb_logger.attach(self,
log_handler=OutputHandler(tag='training', output_transform=lambda x: {
'rpn_box_loss': round(self.state.output['loss_rpn_box_reg'].item(), 4),
'rpn_cls_loss': round(self.state.output['loss_objectness'].item(), 4),
'roi_box_loss': round(self.state.output['loss_box_reg'].item(), 4),
'roi_cls_loss': round(self.state.output['loss_classifier'].item(), 4)
}),
event_name=Events.EPOCH_COMPLETED)
# Run on GPU (cuda) if available
if torch.cuda.is_available():
torch.cuda.set_device(int(get_free_gpu()))
model.cuda(torch.cuda.current_device())
def evaluate(config):
device = torch.device('cuda' if config['use_cuda'] else 'cpu')
model = architecture.Model().to(device)
train_state = dict(model=model)
print('Loading model checkpoint')
workflow.ignite.handlers.ModelCheckpoint.load(
train_state, 'model/checkpoints', device
)
@workflow.ignite.decorators.evaluate(model)
def evaluate_batch(engine, examples):
predictions = model.predictions(
architecture.FeatureBatch.from_examples(examples)
)
loss = predictions.loss(examples)
return dict(
examples=examples,
predictions=predictions.cpu().detach(),
loss=loss,
)
evaluate_data_loaders = {
f'evaluate_{name}': datastream.data_loader(
batch_size=config['eval_batch_size'],
num_workers=config['n_workers'],
collate_fn=tuple,
)
for name, datastream in datastream.evaluate_datastreams().items()
}
tensorboard_logger = TensorboardLogger(log_dir='tb')
for desciption, data_loader in evaluate_data_loaders.items():
engine = evaluator(
evaluate_batch, desciption,
metrics.evaluate_metrics(),
tensorboard_logger,
)
engine.run(data=data_loader)
def logging_board(model_name="densenet121"):
from ignite.contrib.handlers.tensorboard_logger import (
TensorboardLogger,
OutputHandler,
OptimizerParamsHandler,
GradsHistHandler,
)
tb_logger = TensorboardLogger("board/" + model_name)
tb_logger.attach(
trainer,
log_handler=OutputHandler(
tag="training", output_transform=lambda loss: {"loss": loss}),
event_name=Events.ITERATION_COMPLETED,
)
tb_logger.attach(
val_evaluator,
log_handler=OutputHandler(
tag="validation",
metric_names=["accuracy", "loss"],
another_engine=trainer,
),
event_name=Events.EPOCH_COMPLETED,
)
tb_logger.attach(
trainer,
log_handler=OptimizerParamsHandler(IGTrainer.optimizer),
event_name=Events.ITERATION_STARTED,
)
tb_logger.attach(
trainer,
log_handler=GradsHistHandler(IGTrainer.model),
event_name=Events.EPOCH_COMPLETED,
)
tb_logger.close()
def setup(self, training_metrics):
def metric_name(n) -> str:
if n.endswith('Accuracy'):
n = 'acc'
else:
n = n[:-6] if n.endswith('Metric') else n
return n
def print_metrics(metrics) -> str:
rv = ''
metric_keys = sorted(k for k in metrics)
for k in metric_keys:
if k == 'Accuracy':
rv += f'{metric_name(k)}: {metrics[k]:.3}'
else:
rv += f'{metric_name(k)}: {metrics[k]:.6}'
return rv
if self.seed:
set_seed_everywhere(self.seed, self.cuda)
pbar = ProgressBar()
names = []
for k, v in training_metrics.items():
name = f'r{k}'
names.append(name)
RunningAverage(v).attach(self.trainer, name)
RunningAverage(None,
output_transform=lambda x: x[-1] * self.
loss_accumulation_steps).attach(self.trainer, 'rloss')
names.append('rloss')
pbar.attach(self.trainer, names)
pbar = ProgressBar()
pbar.attach(self.evaluator)
# A few events handler. To add / modify the events handler, you need to extend the __init__ method of RunnerABC
# Ignite provides the necessary abstractions and a furnished repository of useful tools
@self.trainer.on(Events.EPOCH_COMPLETED)
def log_validation_results(trainer):
self.evaluator.run(self.dataset_splits.val_data_loader())
metrics = self.evaluator.state.metrics
logger.info(
f"Validation Results - Epoch: {trainer.state.epoch} {print_metrics(metrics)}"
)
if self.scheduler:
self.scheduler.step(
metrics[self.loss_metric.__class__.__name__])
@self.trainer.on(Events.COMPLETED)
def log_test_results(trainer):
self.evaluator.run(self.dataset_splits.test_data_loader())
metrics = self.evaluator.state.metrics
logger.info(
f"Test Results - Epoch: {trainer.state.epoch} {print_metrics(metrics)}"
)
if self.tensorboard_logs:
tb_logger = TensorboardLogger(log_dir=self.tensorboard_logs)
tb_logger.attach(self.trainer,
log_handler=OutputHandler(
tag="training",
output_transform=lambda loss: {'loss': loss}),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(self.evaluator,
log_handler=OutputHandler(
tag="validation",
metric_names=["LossMetric"],
another_engine=self.trainer),
event_name=Events.EPOCH_COMPLETED)
tb_logger.attach(self.trainer,
log_handler=OptimizerParamsHandler(
self.optimizer),
event_name=Events.ITERATION_STARTED)
tb_logger.attach(self.trainer,
log_handler=WeightsScalarHandler(self.model),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(self.trainer,
log_handler=WeightsHistHandler(self.model),
event_name=Events.EPOCH_COMPLETED)
tb_logger.attach(self.trainer,
log_handler=GradsScalarHandler(self.model),
event_name=Events.ITERATION_COMPLETED)
# This is important to close the tensorboard file logger
@self.trainer.on(Events.COMPLETED)
def end_tensorboard(trainer):
logger.info("Training completed")
tb_logger.close()
if self.embeddings_name:
@self.trainer.on(Events.COMPLETED)
def log_embeddings(trainer):
if hasattr(self.model, self.embeddings_name) and hasattr(
self.dataset_splits, "vectorizer"):
logger.info(
f"Logging embeddings ({self.embeddings_name}) to Tensorboard!"
)
embeddings = getattr(self.model,
self.embeddings_name).weight.data
metadata = [
str(self.dataset_splits.vectorizer.data_vocab.
_id2token[token_index]).encode('utf-8')
for token_index in range(embeddings.shape[0])
]
self.writer.add_embedding(
mat=embeddings,
metadata=metadata,
global_step=self.trainer.state.epoch)
def _create_tb_logger(self) -> TensorboardLogger:
now = datetime.now().isoformat(timespec='minutes')
now = now.replace(":", "")
logdir = f"{self.logfolder}/{now}-{self.run_name}"
return TensorboardLogger(log_dir=logdir)
# Prepare metrics - note how we compute distributed metrics
RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(), output_transform=lambda x: (x[0][0], x[1][0]))}
metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], args)})
metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"])
for name, metric in metrics.items():
metric.attach(evaluator, name)
# On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train
if args.local_rank in [-1, 0]:
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=["loss"])
evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics)))
log_dir = make_logdir(args,args.model_checkpoint)
tb_logger = TensorboardLogger(log_dir)
tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]), event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED)
tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED)
checkpoint_handler = ModelCheckpoint(log_dir, 'checkpoint', save_interval=1, n_saved=3)
trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {'mymodel': getattr(model, 'module', model)}) # "getattr" takes care of distributed encapsulation
torch.save(args, log_dir + '/model_training_args.bin')
getattr(model, 'module', model).config.to_json_file(os.path.join(log_dir, CONFIG_NAME))
tokenizer.save_pretrained(log_dir)
# Run the training
trainer.run(train_loader, max_epochs=args.n_epochs)
def train():
parser = ArgumentParser()
parser.add_argument("--train_path", type=str, default='data/spolin-train-acl.json', help="Set data path")
parser.add_argument("--valid_path", type=str, default='data/spolin-valid.json', help="Set data path")
parser.add_argument("--correct_bias", type=bool, default=False, help="Set to true to correct bias for Adam optimizer")
parser.add_argument("--lr", type=float, default=2e-5, help="Set learning rate")
parser.add_argument("--n_epochs", type=int, default=4, help="Set number of epochs")
parser.add_argument("--num_warmup_steps", type=float, default=1000, help="Set number of warm-up steps")
parser.add_argument("--num_total_steps", type=float, default=10000, help="Set number of total steps")
parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu", help="Device (cuda or cpu)")
parser.add_argument("--max_grad_norm", type=float, default=1.0, help="Set maximum gradient normalization.")
parser.add_argument("--pretrained_path", type=str, default='bert-base-uncased', help="Choose which pretrained model to use (bert-base-uncased, roberta-base, roberta-large, roberta-large-mnli)")
parser.add_argument("--batch_size", type=int, default=32, help="Provide the batch size")
parser.add_argument("--random_seed", type=int, default=42, help="Set the random seed")
parser.add_argument("--test", action='store_true', help="If true, run with small dataset for testing code")
parser.add_argument("--base", action='store_true', help="If true, run with base experiment configuration (training with spont only) for comparison")
args = parser.parse_args()
logging.basicConfig(level=logging.INFO)
logger.info("Arguments: {}".format(pformat(args)))
if 'roberta' in args.pretrained_path:
# initialize tokenizer and model
logger.info("Initialize model and tokenizer.")
tokenizer = RobertaTokenizer.from_pretrained(args.pretrained_path, cache_dir = '../pretrained_models')
model = RobertaForSequenceClassification.from_pretrained(args.pretrained_path, cache_dir='../pretrained_models')
### START MODEL MODIFICATION
# Pretrained model was not trained with token type ids.
# fix token type embeddings for finetuning. Without this, the model can only take 0s as valid input for token_type_ids
model.config.type_vocab_size = 2
model.roberta.embeddings.token_type_embeddings = torch.nn.Embedding(2, model.config.hidden_size)
model.roberta.embeddings.token_type_embeddings.weight.data.normal_(mean=0.0, std=model.config.initializer_range)
### END MOD
elif 'bert' in args.pretrained_path:
model = BertForSequenceClassification.from_pretrained(args.pretrained_path, cache_dir='../pretrained_models')
tokenizer = BertTokenizer.from_pretrained(args.pretrained_path, cache_dir='../pretrained_models')
model.to(args.device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0}
]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.lr,
correct_bias = args.correct_bias)
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.num_warmup_steps, t_total=args.num_total_steps)
logger.info("Prepare datasets")
logger.info("Loading train set...")
train_data = get_data(args.train_path)
valid_data = get_data(args.valid_path)
cornell_valid_data = {k: {'cornell': valid_data[k]['cornell']} for k in valid_data.keys()}
spont_valid_data = {k: {'spont': valid_data[k]['spont']} for k in valid_data.keys()}
train_loader, train_sampler = get_data_loaders(args, train_data, args.train_path, tokenizer)
logger.info("Loading validation set...")
valid_p = Path(args.valid_path)
cornell_valid_loader, cornell_valid_sampler = get_data_loaders(args, cornell_valid_data, f"{str(valid_p.parent)}/cornell_{valid_p.name}", tokenizer)
spont_valid_loader, spont_valid_sampler = get_data_loaders(args, spont_valid_data, f"{str(valid_p.parent)}/spont_{valid_p.name}", tokenizer)
# Training function and trainer
def update(engine, batch):
model.train()
batch = tuple(input_tensor.to(args.device) for input_tensor in batch)
b_input_ids, b_input_mask, b_input_segment, b_labels = batch
optimizer.zero_grad()
#roberta has issues with token_type_ids
loss, logits = model(b_input_ids, token_type_ids=b_input_segment, attention_mask=b_input_mask, labels=b_labels)
# loss, logits = model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask, labels=b_labels)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step()
return loss.item(), logits, b_labels
trainer = Engine(update)
# Evaluation function and evaluator
def inference(engine, batch):
model.eval()
batch = tuple(input_tensor.to(args.device) for input_tensor in batch)
b_input_ids, b_input_mask, b_input_segment, b_labels = batch
with torch.no_grad():
#roberta has issues with token_type_ids
# loss, logits = model(b_input_ids, token_type_ids = None, attention_mask=b_input_mask, labels=b_labels)
loss, logits = model(b_input_ids, token_type_ids = b_input_segment, attention_mask=b_input_mask, labels=b_labels)
label_ids = b_labels
return logits, label_ids, loss.item()
cornell_evaluator = Engine(inference)
spont_evaluator = Engine(inference)
trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: cornell_evaluator.run(cornell_valid_loader))
trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: spont_evaluator.run(spont_valid_loader))
RunningAverage(output_transform=lambda x: x[0]).attach(trainer, "loss")
RunningAverage(Accuracy(output_transform=lambda x: (x[1], x[2]))).attach(trainer, "accuracy")
if torch.cuda.is_available():
GpuInfo().attach(trainer, name='gpu')
recall = Recall(output_transform=lambda x: (x[0], x[1]))
precision = Precision(output_transform=lambda x: (x[0], x[1]))
F1 = (precision * recall * 2 / (precision + recall)).mean()
accuracy = Accuracy(output_transform=lambda x: (x[0], x[1]))
metrics = {"recall": recall, "precision": precision, "f1": F1, "accuracy": accuracy, "loss": Average(output_transform=lambda x: x[2])}
for name, metric in metrics.items():
metric.attach(cornell_evaluator, name)
metric.attach(spont_evaluator, name)
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=['loss', 'accuracy'])
pbar.attach(trainer, metric_names=['gpu:0 mem(%)', 'gpu:0 util(%)'])
cornell_evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Cornell validation metrics:\n %s" % pformat(cornell_evaluator.state.metrics)))
spont_evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Spont validation metrics:\n %s" % pformat(spont_evaluator.state.metrics)))
tb_logger = TensorboardLogger(log_dir=None)
tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]), event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED)
tb_logger.attach(cornell_evaluator, log_handler=OutputHandler(tag="valid", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED)
tb_logger.attach(spont_evaluator, log_handler=OutputHandler(tag="valid", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED)
# tb_logger.writer.log_dir -> tb_logger.writer.logdir (this is the correct attribute name as seen in: https://tensorboardx.readthedocs.io/en/latest/_modules/tensorboardX/writer.html#SummaryWriter)
checkpoint_handler = ModelCheckpoint(tb_logger.writer.logdir, 'checkpoint', save_interval=1, n_saved=5)
trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {'mymodel': getattr(model, 'module', model)}) # "getattr" take care of distributed encapsulation
torch.save(args, tb_logger.writer.logdir + '/model_training_args.bin')
getattr(model, 'module', model).config.to_json_file(os.path.join(tb_logger.writer.logdir, CONFIG_NAME))
tokenizer.save_vocabulary(tb_logger.writer.logdir)
trainer.run(train_loader, max_epochs = args.n_epochs)
if args.n_epochs > 0:
os.rename(checkpoint_handler._saved[-1][1][-1], os.path.join(tb_logger.writer.logdir, WEIGHTS_NAME)) # TODO: PR in ignite to have better access to saved file paths (cleaner)
tb_logger.close()
def main():
args = get_args()
if 'e-SNLI-VE' in args.data_path:
args.no_image = False
else:
args.no_image = True
if not args.no_image:
args.no_premise = True
args.with_expl = True
'''Setup'''
t = datetime.today()
output_dir = os.path.join(args.output_folder,
f"{t.month}_{t.day}_{t.hour}_{t.minute}_{t.second}")
if not os.path.exists(output_dir):
os.makedirs(output_dir, exist_ok=True)
# logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes
logging.basicConfig(filename=os.path.join(output_dir, 'app.log'),
filemode='a',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
# This is a logger.warning: it will be printed by all distributed processes
logger.warning(f"Running process {args.local_rank}")
logger.info(f"Arguments: {pformat(args)}")
logger.info(f'Image not used:{args.no_image}')
logger.info(f'Premise not used:{args.no_premise}')
logger.info(f'Explanations used:{args.with_expl}')
'''Initialize distributed training if needed'''
args.distributed = (args.local_rank != -1)
if args.distributed:
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
logger.info(
"Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning")
tokenizer = GPT2Tokenizer.from_pretrained(args.model_checkpoint)
tokenizer.add_special_tokens(SPECIAL_TOKENS_DICT)
if args.no_image:
model = GPT2LMHeadModel.from_pretrained(args.model_checkpoint)
else:
import image_gpt2_291
model = image_gpt2_291.GPT2LMHeadModel.from_pretrained(
args.model_checkpoint)
model.resize_token_embeddings(len(tokenizer))
model.to(args.device)
optimizer = AdamW(model.parameters(), lr=args.lr)
'''
Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last)
'''
if args.fp16:
from apex import amp # Apex is only required if we use fp16 training
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16)
if args.distributed:
model = DistributedDataParallel(model,
device_ids=[args.local_rank],
output_device=args.local_rank)
model = model.module
logger.info("Prepare datasets")
train_loader, val_loader = get_data_loaders(args, tokenizer)
'''Training function and trainer'''
def train(engine, batch):
model.train()
batch = tuple(input_tensor.to(args.device) for input_tensor in batch)
if args.no_image:
input_ids, lm_label, label, input_mask = batch
else:
image, input_ids, lm_label, label, input_mask = batch
if args.no_image:
output = model(input_ids=input_ids,
# attention_mask=input_mask,
labels=lm_label)
else:
output = model(input_ids=input_ids,
images=image,
# attention_mask=input_mask,
labels=lm_label)
loss, logits, _ = output
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_norm)
if engine.state.iteration % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
if not args.with_expl:
lbl_accuracy = torch.eq(label, logits.argmax(
dim=1)).float().sum() / len(label)
return {
'loss': loss.item(),
'lbl_accuracy': lbl_accuracy.item()
}
else:
if engine.state.iteration % (args.gradient_accumulation_steps * 500) == 0:
input_output = list(zip(input_ids, logits))
random_item = random.choice(input_output)
in_sent = tokenizer.decode(list(filter(
lambda x: x != tokenizer.eos_token_id,
random_item[0])))
out_expl = tokenizer.decode(random_item[1].argmax(dim=1),
skip_special_tokens=True)
logger.info(f'MODEL INPUT: {in_sent}')
logger.info(f'GEN. EXPL {out_expl}')
logger.info('--------------------------------')
return {
'loss': loss.item(),
}
'''Validation function and validator (validator output is the input of the metrics)'''
def validation(engine, batch):
model.eval()
with torch.no_grad():
batch = tuple(input_tensor.to(args.device)
for input_tensor in batch)
if args.no_image:
input_ids, lm_label, label, input_mask = batch
else:
image, input_ids, lm_label, label, input_mask = batch
if args.no_image:
output = model(input_ids=input_ids,
# attention_mask=input_mask
)
else:
output = model(input_ids=input_ids,
images=image,
# attention_mask=input_mask
)
logits, _ = output
logits_shifted = logits[..., :-1, :].contiguous().view(-1,
logits.size(-1))
labels_shifted = lm_label[..., 1:].contiguous().view(-1)
return logits_shifted, labels_shifted
'''Engines'''
trainer = Engine(train)
validator = Engine(validation)
# t_total = len(
# train_loader) // args.gradient_accumulation_steps * args.n_epochs
# scheduler = get_linear_schedule_with_warmup(
# optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
'''Linearly decrease the learning rate from lr to zero'''
scheduler = PiecewiseLinear(optimizer, "lr",
[(0, args.lr), (args.n_epochs * len(train_loader), 0.0)])
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
'''
Attach validation to trainer: we evaluate when we start the training and at the end of each epoch
'''
trainer.add_event_handler(Events.EPOCH_COMPLETED,
lambda _: validator.run(val_loader))
if args.eval_before_start:
trainer.add_event_handler(Events.STARTED,
lambda _: validator.run(val_loader))
'''Prepare metrics - note how we compute distributed metrics'''
RunningAverage(output_transform=lambda x: x['loss']).attach(
trainer, "loss")
RunningAverage(output_transform=lambda x: math.exp(
average_distributed_scalar(x['loss'], args))).attach(trainer, "ppl")
if not args.with_expl:
RunningAverage(output_transform=lambda x: 100 * x['lbl_accuracy']).attach(
trainer, "lbl_accuracy")
metrics = {}
metrics["lbl_loss"] = Loss(torch.nn.CrossEntropyLoss(),
output_transform=lambda x: (x[0], x[1]))
metrics["loss"] = MetricsLambda(
lambda l, a: average_distributed_scalar(
l / a.gradient_accumulation_steps, a), metrics["lbl_loss"], args)
metrics["ppl"] = MetricsLambda(math.exp, metrics["loss"])
if not args.with_expl:
metrics["lbl_accuracy"] = 100 * \
Accuracy(output_transform=lambda x: (x[0], x[1]))
for name, metric in metrics.items():
metric.attach(validator, name)
'''
On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train
'''
if args.local_rank in [-1, 0]:
pbar = ProgressBar(persist=True)
pbar.attach(trainer,
metric_names=["loss", 'ppl'] if args.with_expl else ["loss", 'lbl_accuracy', 'ppl'])
validator.add_event_handler(Events.COMPLETED,
lambda _: pbar.log_message(
"Validation: %s" % pformat(validator.state.metrics)))
tb_logger = TensorboardLogger(log_dir=output_dir)
tb_logger.attach(trainer,
log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.ITERATION_STARTED)
tb_logger.attach(trainer,
log_handler=OutputHandler(
tag="training",
metric_names=["loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer,
log_handler=OutputHandler(
tag="training",
metric_names=["ppl"] if args.with_expl else ["lbl_accuracy", "ppl"]),
event_name=Events.EPOCH_COMPLETED)
tb_logger.attach(validator,
log_handler=OutputHandler(
tag="validation",
metric_names=[
'ppl', 'loss'] if args.with_expl else['ppl', 'loss', 'lbl_accuracy'],
global_step_transform=lambda *args, **kwargs: trainer.state.iteration),
event_name=Events.EPOCH_COMPLETED)
checkpoint_handler = ModelCheckpoint(output_dir,
'checkpoint',
n_saved=8,
require_empty=False)
trainer.add_event_handler(Events.EPOCH_COMPLETED(every=1),
checkpoint_handler,
{'mymodel': getattr(model, 'module', model)})
# "getattr" take care of distributed encapsulation
torch.save(args, os.path.join(output_dir, 'model_training_args.bin'))
getattr(model, 'module', model).config.to_json_file(
os.path.join(output_dir, CONFIG_NAME))
tokenizer.save_vocabulary(output_dir)
'''Run the training'''
trainer.run(train_loader, max_epochs=args.n_epochs)
def train(args):
logger.info("Prepare tokenizer, pretrained model and optimizer.")
tokenizer, _, vocab = get_kogpt2_tokenizer()
model = get_kogpt2_model()
model.to(args.device)
optimizer = AdamW(model.parameters(), lr=args.lr, correct_bias=True)
logger.info("Prepare datasets")
train_loader, val_loader = get_data_loaders(args, tokenizer, vocab)
def update(engine, batch):
model.train()
batch = tuple(input_tensor.to(args.device) for input_tensor in batch)
input_ids, labels, token_type_ids = batch
loss, *_ = model(input_ids,
token_type_ids=token_type_ids,
labels=labels)
loss = loss / args.gradient_accumulation_steps
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_norm)
if engine.state.iteration % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item()
trainer = Engine(update)
def inference(engine, batch):
model.eval()
with torch.no_grad():
batch = tuple(
input_tensor.to(args.device) for input_tensor in batch)
input_ids, labels, token_type_ids = batch
# logger.info(tokenizer.decode(input_ids[0, -1, :].tolist()))
# if we dont send labels to model, it doesnt return losses
logits, *_ = model(input_ids, token_type_ids=token_type_ids)
logits_flat_shifted = logits[..., :-1, :].contiguous().view(
-1, logits.size(-1))
labels_flat_shifted = labels[..., 1:].contiguous().view(-1)
return (logits_flat_shifted), (labels_flat_shifted)
evaluator = Engine(inference)
# Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch
trainer.add_event_handler(Events.EPOCH_COMPLETED,
lambda _: evaluator.run(val_loader))
if args.n_epochs < 1:
trainer.add_event_handler(Events.COMPLETED,
lambda _: evaluator.run(val_loader))
if args.eval_before_start:
trainer.add_event_handler(Events.STARTED,
lambda _: evaluator.run(val_loader))
# Linearly decrease the learning rate from lr to zero
scheduler = PiecewiseLinear(optimizer, "lr",
[(0, args.lr),
(args.n_epochs * len(train_loader), 0.0)])
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
# Prepare metrics - note how we compute distributed metrics
RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
metrics = {
"nll":
Loss(torch.nn.CrossEntropyLoss(ignore_index=-100),
output_transform=lambda x: (x[0], x[1])),
"accuracy":
Accuracy(output_transform=lambda x: (x[0], x[1]))
}
for name, metric in metrics.items():
metric.attach(evaluator, name)
# On the main process: add progress bar, tensorboard, checkpoints and save model,
# configuration and tokenizer before we start to train
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=["loss"])
evaluator.add_event_handler(
Events.COMPLETED, lambda _: pbar.log_message(
"Validation: %s" % pformat(evaluator.state.metrics)))
log_dir = make_logdir("kogpt2_personachat")
tb_logger = TensorboardLogger(log_dir)
tb_logger.attach(trainer,
log_handler=OutputHandler(tag="training",
metric_names=["loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer,
log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.ITERATION_STARTED)
tb_logger.attach(
evaluator,
log_handler=OutputHandler(
tag="validation",
metric_names=list(metrics.keys()),
global_step_transform=global_step_from_engine(trainer)),
event_name=Events.EPOCH_COMPLETED)
checkpoint_handler = ModelCheckpoint(log_dir,
'checkpoint',
save_interval=1,
n_saved=3)
trainer.add_event_handler(
Events.EPOCH_COMPLETED, checkpoint_handler,
{'mymodel': getattr(model, 'module', model)
}) # "getattr" takes care of distributed encapsulation
torch.save(args, log_dir + '/model_training_args.bin')
getattr(model, 'module',
model).config.to_json_file(os.path.join(log_dir, CONFIG_NAME))
# tokenizer.save_pretrained(log_dir)
# Run the training
trainer.run(train_loader, max_epochs=args.n_epochs)
# On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method)
# TODO: PR in ignite to have better access to saved file paths (cleaner)
os.rename(os.path.join(log_dir, checkpoint_handler._saved[-1][1]),
os.path.join(log_dir, WEIGHTS_NAME))
tb_logger.close()
def train():
os.environ['CUDA_VISIBLE_DEVICES'] = '7'
parser = ArgumentParser()
parser.add_argument('--gpt2', action='store_true', help="use gpt2")
parser.add_argument("--model_checkpoint", type=str, default="uer/gpt2-chinese-cluecorpussmall", help="Path or URL of the model")
parser.add_argument("--from_step", type=int, default=-1, help="Init learning rate from this step")
parser.add_argument('--pretrained', action='store_true', help="If False train from scratch")
parser.add_argument("--data_path", type=str, default="data/autocloze.json",
help="Path or url of the dataset. ")
parser.add_argument("--train_path", type=str, default="data/toy_train.txt",
help="Path of the train dataset for dist dataset. ")
parser.add_argument("--valid_path", type=str, default="data/toy_valid.txt",
help="Path of the valid dataset for dist dataset. ")
#--------------------------------------------------------------
parser.add_argument("--dataset_cache", type=str, default="dataset_zh",
help="Path or url of the dataset cache")
parser.add_argument('--log_file', '-log_file', type=str, default="", help="Output logs to a file under this path")
parser.add_argument("--num_workers", type=int, default=8, help="Number of subprocesses for data loading")
parser.add_argument("--n_epochs", type=int, default=40, help="Number of training epochs")
parser.add_argument("--train_batch_size", type=int, default=1, help="Batch size for training")
parser.add_argument("--valid_batch_size", type=int, default=1, help="Batch size for validation")
parser.add_argument("--max_history", type=int, default=15, help="Number of previous exchanges to keep in history")
parser.add_argument("--scheduler", type=str, default="noam", choices=['noam', 'linear'], help="method of optim")
parser.add_argument("--n_emd", type=int, default=768, help="Number of n_emd in config file (for noam)")
parser.add_argument("--lr", type=float, default=5e-5, help="Learning rate")
parser.add_argument("--eval_before_start", action='store_true',
help="If true start with a first evaluation before training")
parser.add_argument("--warmup_steps", type=int, default=5000, help="Warm up steps")
parser.add_argument("--valid_steps", type=int, default=5000, help="Perfom validation every X steps")
parser.add_argument("--gradient_accumulation_steps", type=int, default=64,
help="Accumulate gradients on several steps")
parser.add_argument("--max_norm", type=float, default=1.0, help="Clipping gradient norm")
parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu",
help="Device (cuda or cpu)")
parser.add_argument("--fp16", type=str, default="",
help="Set to O0, O1, O2 or O3 for fp16 training (see apex documentation)")
parser.add_argument("--local_rank", type=int, default=-1,
help="Local rank for distributed training (-1: not distributed)")
args = parser.parse_args()
print('cuda ',torch.cuda.is_available())
# logging is set to INFO (resp. WARN) for main (resp. auxiliary) process.
# logger.info => log main process only, logger.warning => log all processes
logging.basicConfig(level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning("Running process %d", args.local_rank)
logger.info("Arguments: %s", pformat(args))
# Initialize distributed training if needed
args.distributed = (args.local_rank != -1)
'''if args.distributed:
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
'''
args.device = torch.device("cuda")
print('device ',args.device)
logger.info("Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning")
#model_class = OpenAIGPTLMHeadModel if not args.gpt2 else GPT2LMHeadModel
#config_class = OpenAIGPTConfig if not args.gpt2 else GPT2Config
model_class = GPT2LMHeadModel
config_class = GPT2Config
tokenizer_class = BertTokenizer
print('pretrained:',args.pretrained)
if args.pretrained:
print("----------------pretrained")
tokenizer = BertTokenizer.from_pretrained(args.model_checkpoint, do_lower_case=True)
model = GPT2LMHeadModel.from_pretrained(args.model_checkpoint)
else:
tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-cluecorpussmall")
model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-cluecorpussmall",from_tf=True)
#print('generate')
#print(text_generator("这是很久之前的事情了", max_length=100, do_sample=True))
#args.device=torch.device("cuda", 2)
model.to(args.device)
optimizer = AdamW([{'params': model.parameters(), 'initial_lr': args.lr}], lr=args.lr, correct_bias=True)
logger.info("Prepare datasets")
loader_class = build_dist_loaders if not args.data_path else build_dataloaders
train_loader, val_loader, train_sampler, valid_sampler = loader_class(args, tokenizer, logger)
logger.info("Prepare datasets ends")
# Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last)
if args.fp16:
from apex import amp # Apex is only required if we use fp16 training
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16)
if args.distributed:
model = DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank)
model=model.module
#if isinstance(model,torch.nn.DataParallel):
#print('params:',params_count(model))
#tokens_embed = model.transformer.get_input_embeddings()
# Training function and trainer
def update(engine, batch):
input_ids, token_type_ids, lm_labels = tuple(input_tensor.to(args.device) for input_tensor in batch)
#for i in range(input_ids.size()[0]):
# for j in range(input_ids.size()[1]):
# if input_ids[i,j]==-1:
# input_ids[i,j]=-100
# if lm_labels[i,j]==-1:
# lm_labels[i,j]=-100
#one=torch.tensor(-100)
#input_ids=torch.where(input_ids==-1,one,input_ids)
#lm_labels=torch.where(lm_labels==-1,one,lm_labels)
#print('traindata',input_ids,lm_labels)
#lm_labels=input_ids
r'''input_shape = input_ids.siz`e`()
input_ids = input_ids.view(-1, input_shape[-1])
inputs_embeds = tokens_embed(input_ids) * math.sqrt(tokens_embed.embedding_dim)'''
model.train()
#(lm_loss), *_ = model(inputs_embeds=inputs_embeds, labels=lm_labels,return_dict=0)
(lm_loss), *_ = model(input_ids=input_ids, labels=lm_labels,return_dict=False)
#print('lm_loss',lm_loss)
loss = lm_loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_norm)
if engine.state.iteration % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item(), optimizer.param_groups[0]['lr']
trainer = Engine(update)
# Evaluation function and evaluator (evaluator output is the input of the metrics)
cntepoch=0
def inference(engine, batch):
model.eval()
with torch.no_grad():
input_ids, token_type_ids, lm_labels = tuple(input_tensor.to(args.device) for input_tensor in batch)
# logger.info(tokenizer.decode(input_ids[0, -1, :].tolist()))
#one = torch.tensor(-100)
#input_ids=torch.where(input_ids==-1,one,input_ids)
#print('validdata',input_ids,lm_labels)
#lm_labels=input_ids
r'''input_shape = input_ids.size()
input_ids = input_ids.view(-1, input_shape[-1])
inputs_embeds = tokens_embed(input_ids) * math.sqrt(tokens_embed.embedding_dim)'''
#lm_logits, *_ = model(inputs_embeds=inputs_embeds,return_dict=0)
lm_logits, *_ = model(input_ids=input_ids,return_dict=False)
lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(-1, lm_logits.size(-1))
lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1)
return lm_logits_flat_shifted, lm_labels_flat_shifted
cntepoch+=1
torch.save(args, tb_logger.writer.logdir + '_%s/model_training_args.bin'%(str(cntepoch)))
evaluator = Engine(inference)
# Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch
trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: evaluator.run(val_loader))
if args.n_epochs < 1:
trainer.add_event_handler(Events.COMPLETED, lambda _: evaluator.run(val_loader))
if args.eval_before_start:
trainer.add_event_handler(Events.STARTED, lambda _: evaluator.run(val_loader))
# Evaluation during training
@trainer.on(Events.ITERATION_STARTED)
def log_iterations(engine):
# if engine.state.iteration % max(int(0.1 * len(train_loader)), 1) == 0:
if engine.state.iteration % args.valid_steps == 0:
evaluator.run(val_loader)
# Make sure distributed data samplers split the dataset nicely between the distributed processes
if args.distributed:
trainer.add_event_handler(Events.EPOCH_STARTED, lambda engine: train_sampler.set_epoch(engine.state.epoch))
evaluator.add_event_handler(Events.EPOCH_STARTED, lambda engine: valid_sampler.set_epoch(engine.state.epoch))
# noam decrease the learning rate
# model_size = model.config.n_embd
model_size = args.n_emd
noam_lambda = lambda step: (
model_size ** (-0.5) * min((step + 1) ** (-0.5), (step + 1) * args.warmup_steps ** (-1.5)))
noam_scheduler = LambdaLR(optimizer, lr_lambda=noam_lambda, last_epoch=args.from_step)
scheduler = LRScheduler(noam_scheduler)
if args.scheduler == "linear":
scheduler = PiecewiseLinear(optimizer, "lr", [(0, args.lr), (args.n_epochs * len(train_loader), 0.0)])
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
# Prepare metrics - note how we compute distributed metrics
RunningAverage(output_transform=lambda x: x[0]).attach(trainer, "loss")
RunningAverage(output_transform=lambda x: x[1]).attach(trainer, "lr")
metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-100), output_transform=lambda x: (x[0], x[1]))}
metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], args)})
metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"])
for name, metric in metrics.items():
metric.attach(evaluator, name)
# On the main process: add progress bar, tensorboard, checkpoints
# And save model, configuration and tokenizer before we start to train
if args.local_rank in [-1, 0]:
pbar = ProgressBar(persist=True, mininterval=2)
pbar.attach(trainer, metric_names=["loss", "lr"])
evaluator.add_event_handler(Events.COMPLETED,
lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics)))
tb_logger = TensorboardLogger(log_dir=None)
tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED)
tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()),
another_engine=trainer),
event_name=Events.EPOCH_COMPLETED)
checkpoint_handler = ModelCheckpoint(tb_logger.writer.logdir, 'checkpoint', save_interval=1, n_saved=6)
# save model after evaluation
evaluator.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {
'mymodel': getattr(model, 'module', model)})
trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {
'mymodel': getattr(model, 'module', model)}) # "getattr" take care of distributed encapsulation
torch.save(args, tb_logger.writer.logdir + '/model_training_args.bin')
getattr(model, 'module', model).config.to_json_file(os.path.join(tb_logger.writer.logdir, CONFIG_NAME))
tokenizer.save_vocabulary(tb_logger.writer.logdir)
# Run the training
trainer.run(train_loader, max_epochs=args.n_epochs)
# On the main process: close tensorboard logger and rename the last checkpoint
# (for easy re-loading with OpenAIGPTModel.from_pretrained method)
if args.local_rank in [-1, 0] and args.n_epochs > 0:
os.rename(checkpoint_handler._saved[-1][1][-1],
os.path.join(tb_logger.writer.logdir,
WEIGHTS_NAME)) # TODO: PR in ignite to have better access to saved file paths (cleaner)
tb_logger.close()
def attach_handlers(run, model, optimizer, learning_rule, trainer, evaluator, train_loader, val_loader, params):
# Metrics
UnitConvergence(model[0], learning_rule.norm).attach(trainer.engine, 'unit_conv')
# Tqdm logger
pbar = ProgressBar(persist=True, bar_format=config.IGNITE_BAR_FORMAT)
pbar.attach(trainer.engine, metric_names='all')
tqdm_logger = TqdmLogger(pbar=pbar)
# noinspection PyTypeChecker
tqdm_logger.attach_output_handler(
evaluator.engine,
event_name=Events.COMPLETED,
tag="validation",
global_step_transform=global_step_from_engine(trainer.engine),
)
# Evaluator
evaluator.attach(trainer.engine, Events.EPOCH_COMPLETED(every=100), train_loader, val_loader)
# Learning rate scheduling
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer=optimizer,
lr_lambda=lambda epoch: 1 - epoch / params['epochs'])
lr_scheduler = LRScheduler(lr_scheduler)
trainer.engine.add_event_handler(Events.EPOCH_COMPLETED, lr_scheduler)
# Early stopping
mc_handler = ModelCheckpoint(config.MODELS_DIR, run.replace('/', '-'), n_saved=1, create_dir=True,
require_empty=False,
global_step_transform=global_step_from_engine(trainer.engine))
trainer.engine.add_event_handler(Events.EPOCH_COMPLETED, mc_handler, {'m': model})
# Create a TensorBoard logger
tb_logger = TensorboardLogger(log_dir=os.path.join(config.TENSORBOARD_DIR, run))
images, labels = next(iter(train_loader))
tb_logger.writer.add_graph(copy.deepcopy(model).cpu(), images)
tb_logger.writer.add_hparams(params, {})
# noinspection PyTypeChecker
tb_logger.attach_output_handler(
evaluator.engine,
event_name=Events.COMPLETED,
tag="validation",
metric_names="all",
global_step_transform=global_step_from_engine(trainer.engine),
)
# noinspection PyTypeChecker
tb_logger.attach_output_handler(
trainer.engine,
event_name=Events.EPOCH_COMPLETED,
tag="train",
metric_names=["unit_conv"]
)
input_shape = tuple(next(iter(train_loader))[0].shape[1:])
tb_logger.attach(trainer.engine,
log_handler=WeightsImageHandler(model, input_shape),
event_name=Events.EPOCH_COMPLETED)
tb_logger.attach(trainer.engine, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.EPOCH_STARTED)
# tb_logger.attach(trainer.engine,
# log_handler=WeightsScalarHandler(model, layer_names=['linear1', 'linear2']),
# event_name=Events.EPOCH_COMPLETED)
# tb_logger.attach(trainer.engine,
# log_handler=WeightsHistHandler(model, layer_names=['linear1', 'linear2']),
# event_name=Events.EPOCH_COMPLETED)
# tb_logger.attach(trainer.engine,
# log_handler=ActivationsHistHandler(model, layer_names=['batch_norm', 'repu']),
# event_name=Events.ITERATION_COMPLETED)
# tb_logger.attach(trainer.engine,
# log_handler=NumActivationsScalarHandler(model, layer_names=['repu']),
# event_name=Events.ITERATION_COMPLETED)
# tb_logger.attach(trainer.engine,
# log_handler=ActivationsScalarHandler(model, reduction=torch.mean,
# layer_names=['batch_norm', 'repu']),
# event_name=Events.ITERATION_COMPLETED)
# tb_logger.attach(trainer.engine,
# log_handler=ActivationsScalarHandler(model, reduction=torch.std,
# layer_names=['batch_norm', 'repu']),
# event_name=Events.ITERATION_COMPLETED)
return tb_logger
def train_network(model: nn.Module, training_loader: DataLoader,
validation_loader: DataLoader):
"""Trains the given neural network model.
Parameters
----------
model (nn.Module): The PyTorch model to be trained
training_loader (DataLoader): Training data loader
validation_loader (DataLoader): Validation data loader
"""
device = "cuda:0" if cast(Any, torch).cuda.is_available() else "cpu"
if device == "cuda:0":
model.cuda()
optimizer = cast(Any, torch).optim.Adam(model.parameters(), lr=0.001)
criterion = nn.MSELoss()
trainer = create_supervised_trainer(model,
optimizer,
criterion,
device=device)
save_handler = Checkpoint(
{
"model": model,
"optimizer": optimizer,
"trainer": trainer
},
DiskSaver("dist/models", create_dir=True),
n_saved=2,
)
trainer.add_event_handler(Events.EPOCH_COMPLETED(every=100), save_handler)
# Create a logger
tb_logger = TensorboardLogger(log_dir="logs/training" +
datetime.now().strftime("-%Y%m%d-%H%M%S"),
flush_secs=1)
tb_logger.attach_output_handler(
trainer,
event_name=Events.ITERATION_COMPLETED,
tag="training",
output_transform=lambda loss: {"loss": loss},
)
# Training evaluator
training_evaluator = create_supervised_evaluator(model,
metrics={
"r2": R2Score(),
"MSELoss":
Loss(criterion)
},
device=device)
tb_logger.attach_output_handler(
training_evaluator,
event_name=Events.EPOCH_COMPLETED,
tag="training",
metric_names=["MSELoss", "r2"],
global_step_transform=global_step_from_engine(trainer),
)
# Validation evaluator
evaluator = create_supervised_evaluator(model,
metrics={
"r2": R2Score(),
"MSELoss": Loss(criterion)
},
device=device)
tb_logger.attach_output_handler(
evaluator,
event_name=Events.EPOCH_COMPLETED,
tag="validation",
metric_names=["MSELoss", "r2"],
global_step_transform=global_step_from_engine(trainer),
)
@trainer.on(Events.EPOCH_COMPLETED(every=10))
def log_training_results(trainer):
training_evaluator.run(training_loader)
metrics = training_evaluator.state.metrics
print(
f"Training Results - Epoch: {trainer.state.epoch}",
f" Avg r2: {metrics['r2']:.2f} Avg loss: {metrics['MSELoss']:.2f}",
)
@trainer.on(Events.EPOCH_COMPLETED(every=10))
def log_validation_results(trainer):
evaluator.run(validation_loader)
metrics = evaluator.state.metrics
print(
f"Validation Results - Epoch: {trainer.state.epoch}",
f" Avg r2: {metrics['r2']:.2f} Avg loss: {metrics['MSELoss']:.2f}\n",
)
trainer.run(training_loader, max_epochs=int(1e6))
def trainer(
train_batch,
evaluate_batch,
evaluate_data_loaders,
metrics,
optimizers,
):
'''
Create standard trainer with evaluators.
Parameters
----------
train_batch : function
function that trains on given batch
evaluate_batch : function
function that evaluates a given batch
evaluate_data_loaders: list
data loaders that yield batches to evaluate on
metrics : dict
dict with one dict each for 'train' and evaluate data loader. Wrap a
metric with trainer.Progress to show in progress bar.
optimizers : dict
dict with optimizers for logging
Returns
-------
tuple
trainer engine
list of evaluator engines
tensorboard logger
'''
trainer = ignite.engine.Engine(train_batch)
for name, metric in metrics.get(PROGRESS_DESC, dict()).items():
metric.attach(trainer, name)
for name, metric in metrics.get(TRAIN_DESC, dict()).items():
metric.attach(trainer, name)
evaluators = {
evaluator_name: ignite.engine.Engine(evaluate_batch)
for evaluator_name in evaluate_data_loaders.keys()
}
for evaluator_name, evaluator in evaluators.items():
for metric_name, metric in metrics[evaluator_name].items():
metric.attach(evaluator, metric_name)
tensorboard_logger = TensorboardLogger(log_dir='tb')
EpochLogger().attach(trainer)
# Order of attaching progress bars is important for vscode / atom
ProgressBar(desc=TRAIN_DESC).attach(trainer,
metric_names=list(
metrics.get(PROGRESS_DESC,
dict()).keys()))
tensorboard_logger.attach(
trainer,
OutputHandler(
tag=PROGRESS_DESC,
metric_names=list(metrics.get(PROGRESS_DESC, dict()).keys()),
),
Events.ITERATION_COMPLETED,
)
MetricsLogger(TRAIN_DESC).attach(trainer,
metrics.get(TRAIN_DESC, dict()).keys())
tensorboard_logger.attach(
trainer,
OutputHandler(
tag=TRAIN_DESC,
metric_names=list(metrics.get(TRAIN_DESC, dict()).keys()),
),
Events.ITERATION_COMPLETED,
)
def run_evaluator(evaluator_desc):
return lambda engine: evaluators[evaluator_desc].run(
evaluate_data_loaders[evaluator_desc])
for evaluator_desc, evaluator in evaluators.items():
evaluator_metric_names = list(metrics[evaluator_desc].keys())
trainer.add_event_handler(
Events.EPOCH_COMPLETED,
run_evaluator(evaluator_desc),
)
ProgressBar(desc=evaluator_desc).attach(evaluator)
MetricsLogger(evaluator_desc).attach(evaluator, evaluator_metric_names)
tensorboard_logger.attach(
evaluator,
OutputHandler(
tag=evaluator_desc,
metric_names=evaluator_metric_names,
global_step_transform=global_step_from_engine(trainer),
),
Events.EPOCH_COMPLETED,
)
if type(optimizers) is not dict:
optimizers = dict(optimizer=optimizers)
for name, optimizer in optimizers.items():
tensorboard_logger.attach(
trainer,
log_handler=OptimizerParamsHandler(
tag=f'{TRAIN_DESC}/{name}',
param_name='lr',
optimizer=optimizer,
),
event_name=Events.ITERATION_COMPLETED,
)
return trainer, evaluators, tensorboard_logger
def run(train_batch_size, val_batch_size, epochs, lr, momentum, log_dir):
train_loader, val_loader = get_data_loaders(train_batch_size,
val_batch_size)
model = Net()
device = "cpu"
if torch.cuda.is_available():
device = "cuda"
model.to(device) # Move model before creating optimizer
optimizer = SGD(model.parameters(), lr=lr, momentum=momentum)
criterion = nn.CrossEntropyLoss()
trainer = create_supervised_trainer(model,
optimizer,
criterion,
device=device)
trainer.logger = setup_logger("Trainer")
if sys.version_info > (3, ):
from ignite.contrib.metrics.gpu_info import GpuInfo
try:
GpuInfo().attach(trainer)
except RuntimeError:
print(
"INFO: By default, in this example it is possible to log GPU information (used memory, utilization). "
"As there is no pynvml python package installed, GPU information won't be logged. Otherwise, please "
"install it : `pip install pynvml`")
metrics = {"accuracy": Accuracy(), "loss": Loss(criterion)}
train_evaluator = create_supervised_evaluator(model,
metrics=metrics,
device=device)
train_evaluator.logger = setup_logger("Train Evaluator")
validation_evaluator = create_supervised_evaluator(model,
metrics=metrics,
device=device)
validation_evaluator.logger = setup_logger("Val Evaluator")
@trainer.on(Events.EPOCH_COMPLETED)
def compute_metrics(engine):
train_evaluator.run(train_loader)
validation_evaluator.run(val_loader)
tb_logger = TensorboardLogger(log_dir=log_dir)
tb_logger.attach_output_handler(
trainer,
event_name=Events.ITERATION_COMPLETED(every=100),
tag="training",
output_transform=lambda loss: {"batchloss": loss},
metric_names="all",
)
for tag, evaluator in [("training", train_evaluator),
("validation", validation_evaluator)]:
tb_logger.attach_output_handler(
evaluator,
event_name=Events.EPOCH_COMPLETED,
tag=tag,
metric_names=["loss", "accuracy"],
global_step_transform=global_step_from_engine(trainer),
)
tb_logger.attach_opt_params_handler(
trainer,
event_name=Events.ITERATION_COMPLETED(every=100),
optimizer=optimizer)
tb_logger.attach(trainer,
log_handler=WeightsScalarHandler(model),
event_name=Events.ITERATION_COMPLETED(every=100))
tb_logger.attach(trainer,
log_handler=WeightsHistHandler(model),
event_name=Events.EPOCH_COMPLETED(every=100))
tb_logger.attach(trainer,
log_handler=GradsScalarHandler(model),
event_name=Events.ITERATION_COMPLETED(every=100))
tb_logger.attach(trainer,
log_handler=GradsHistHandler(model),
event_name=Events.EPOCH_COMPLETED(every=100))
def score_function(engine):
return engine.state.metrics["accuracy"]
model_checkpoint = ModelCheckpoint(
log_dir,
n_saved=2,
filename_prefix="best",
score_function=score_function,
score_name="validation_accuracy",
global_step_transform=global_step_from_engine(trainer),
)
validation_evaluator.add_event_handler(Events.COMPLETED, model_checkpoint,
{"model": model})
# kick everything off
trainer.run(train_loader, max_epochs=epochs)
tb_logger.close()
end_value=LR / 4,
cycle_size=TOTAL_UPDATE_STEPS // 1)
evaluator.add_event_handler(Events.EPOCH_COMPLETED, checkpointer,
{'_': mude})
trainer.add_event_handler(Events.ITERATION_COMPLETED, nan_handler)
trainer.add_event_handler(Events.ITERATION_COMPLETED, coslr)
GpuInfo().attach(trainer, name='gpu')
pbar.attach(trainer,
output_transform=lambda output: {'loss': output['loss']},
metric_names=[f"gpu:{args.gpu} mem(%)"])
# FIRE
tb_logger = TensorboardLogger(log_dir=TENSORBOARD_RUN_LOG_DIR_PATH)
tb_logger.attach(
trainer,
log_handler=OutputHandler(
tag='training',
output_transform=lambda output: {'loss': output['loss']}),
event_name=Events.ITERATION_COMPLETED(
every=LOG_TRAINING_PROGRESS_EVERY_N))
tb_logger.attach(
evaluator,
log_handler=OutputHandler(
tag='validation',
metric_names='all',
global_step_transform=global_step_from_engine(trainer)),
event_name=Events.EPOCH_COMPLETED)
tb_logger.attach(trainer,
def train():
config_file = "configs/train_daily_dialog_emotion_action_config.json"
config = Config.from_json_file(config_file)
# logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes
logging.basicConfig(
level=logging.INFO if config.local_rank in [-1, 0] else logging.WARN)
logger.warning(
"Running process %d", config.local_rank
) # This is a logger.warning: it will be printed by all distributed processes
logger.info("Arguments: %s", pformat(config))
# Initialize distributed training if needed
config.distributed = (config.local_rank != -1)
if config.distributed:
torch.cuda.set_device(config.local_rank)
config.device = torch.device("cuda", config.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
logger.info(
"Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning"
)
tokenizer_class = GPT2Tokenizer if "gpt2" in config.model_checkpoint else OpenAIGPTTokenizer
tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint)
model_class = GPT2DoubleHeadsModel if "gpt2" in config.model_checkpoint else OpenAIGPTDoubleHeadsModel
model = model_class.from_pretrained(config.model_checkpoint)
tokenizer.set_special_tokens(SPECIAL_TOKENS)
model.set_num_special_tokens(len(SPECIAL_TOKENS))
model.to(config.device)
optimizer = OpenAIAdam(model.parameters(), lr=config.lr)
# Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last)
if config.fp16:
from apex import amp # Apex is only required if we use fp16 training
model, optimizer = amp.initialize(model,
optimizer,
opt_level=config.fp16)
if config.distributed:
model = DistributedDataParallel(model,
device_ids=[config.local_rank],
output_device=config.local_rank)
logger.info("Prepare datasets")
train_loader, val_loader, train_sampler, valid_sampler = get_data_loaders(
config, tokenizer)
# Training function and trainer
def update(engine, batch):
model.train()
input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids, token_action_ids = tuple(
input_tensor.to(config.device) for input_tensor in batch)
lm_loss, mc_loss = model(input_ids, mc_token_ids, lm_labels, mc_labels,
token_type_ids, token_emotion_ids,
token_action_ids)
loss = (lm_loss * config.lm_coef +
mc_loss * config.mc_coef) / config.gradient_accumulation_steps
if config.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
config.max_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), config.max_norm)
if engine.state.iteration % config.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item()
trainer = Engine(update)
# Evaluation function and evaluator (evaluator output is the input of the metrics)
def inference(engine, batch):
model.eval()
with torch.no_grad():
batch = tuple(
input_tensor.to(config.device) for input_tensor in batch)
input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids, token_action_ids = batch
#logger.info(tokenizer.decode(input_ids[0, -1, :].tolist()))
model_outputs = model(input_ids,
mc_token_ids,
token_type_ids=token_type_ids,
token_emotion_ids=token_emotion_ids,
token_action_ids=token_action_ids)
lm_logits, mc_logits = model_outputs[0], model_outputs[
1] # So we can also use GPT2 outputs
lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(
-1, lm_logits.size(-1))
lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1)
return (lm_logits_flat_shifted,
mc_logits), (lm_labels_flat_shifted, mc_labels)
evaluator = Engine(inference)
# Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch
trainer.add_event_handler(Events.EPOCH_COMPLETED,
lambda _: evaluator.run(val_loader))
if config.n_epochs < 1:
trainer.add_event_handler(Events.COMPLETED,
lambda _: evaluator.run(val_loader))
if config.eval_before_start:
trainer.add_event_handler(Events.STARTED,
lambda _: evaluator.run(val_loader))
# Make sure distributed data samplers split the dataset nicely between the distributed processes
if config.distributed:
trainer.add_event_handler(
Events.EPOCH_STARTED,
lambda engine: train_sampler.set_epoch(engine.state.epoch))
evaluator.add_event_handler(
Events.EPOCH_STARTED,
lambda engine: valid_sampler.set_epoch(engine.state.epoch))
# Linearly decrease the learning rate from lr to zero
scheduler = PiecewiseLinear(optimizer, "lr",
[(0, config.lr),
(config.n_epochs * len(train_loader), 0.0)])
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
# Prepare metrics - note how we compute distributed metrics
RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
metrics = {
"nll":
Loss(torch.nn.CrossEntropyLoss(ignore_index=-1),
output_transform=lambda x: (x[0][0], x[1][0])),
"accuracy":
Accuracy(output_transform=lambda x: (x[0][1], x[1][1]))
}
metrics.update({
"average_nll":
MetricsLambda(average_distributed_scalar, metrics["nll"], config),
"average_accuracy":
MetricsLambda(average_distributed_scalar, metrics["accuracy"], config)
})
metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"])
for name, metric in metrics.items():
metric.attach(evaluator, name)
# On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train
if config.local_rank in [-1, 0]:
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=["loss"])
evaluator.add_event_handler(
Events.COMPLETED, lambda _: pbar.log_message(
"Validation: %s" % pformat(evaluator.state.metrics)))
tb_logger = TensorboardLogger(log_dir=config.log_dir)
tb_logger.attach(trainer,
log_handler=OutputHandler(tag="training",
metric_names=["loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer,
log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.ITERATION_STARTED)
tb_logger.attach(evaluator,
log_handler=OutputHandler(tag="validation",
metric_names=list(
metrics.keys()),
another_engine=trainer),
event_name=Events.EPOCH_COMPLETED)
checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir,
'checkpoint',
save_interval=1,
n_saved=3)
trainer.add_event_handler(
Events.EPOCH_COMPLETED, checkpoint_handler,
{'mymodel': getattr(model, 'module', model)
}) # "getattr" take care of distributed encapsulation
torch.save(config,
tb_logger.writer.log_dir + '/model_training_args.bin')
getattr(model, 'module', model).config.to_json_file(
os.path.join(tb_logger.writer.log_dir, CONFIG_NAME))
tokenizer.save_vocabulary(tb_logger.writer.log_dir)
# Run the training
trainer.run(train_loader, max_epochs=config.n_epochs)
# On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method)
if config.local_rank in [-1, 0] and config.n_epochs > 0:
os.rename(
checkpoint_handler._saved[-1][1][-1],
os.path.join(tb_logger.writer.log_dir, WEIGHTS_NAME)
) # TODO: PR in ignite to have better access to saved file paths (cleaner)
tb_logger.close()
def train():
parser = ArgumentParser()
parser.add_argument("--local_rank", type=int, default=-1)
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.device_count() > 1 else "cpu")
model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
DISTRIBUTED = args.local_rank != -1
if DISTRIBUTED and torch.distributed.is_available():
print("Distributed")
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
#BATCH_SIZE *= 2
def average_distributed_scalar(scalar):
if (not DISTRIBUTED):
return scalar
scalar_t = torch.tensor(
scalar, dtype=torch.float,
device=device) / torch.distributed.get_world_size()
torch.distributed.all_reduce(scalar_t,
op=torch.distributed.ReduceOp.SUM)
return scalar_t.item()
optimizer = AdamW(model.parameters(), lr=6.25e-5)
ds = dataloader.Conv_GPT2_DataClass(tokenizer)
v_ds = dataloader.Conv_GPT2_DataClass(tokenizer, dev=True)
orig_added_tokens = len(tokenizer.encoder)
num_added_tokens = tokenizer.add_special_tokens(
dataloader.ATTR_SPECIAL_TOKENS)
if (num_added_tokens > 0):
model.resize_token_embeddings(new_num_tokens=orig_added_tokens +
num_added_tokens)
model = model.to(device)
train_sampler = torch.utils.data.distributed.DistributedSampler(
ds) if DISTRIBUTED else None
valid_sampler = torch.utils.data.distributed.DistributedSampler(
v_ds) if DISTRIBUTED else None
dl = DataLoader(ds,
sampler=train_sampler,
batch_size=BATCH_SIZE,
shuffle=not DISTRIBUTED)
v_dl = DataLoader(v_ds, sampler=valid_sampler, shuffle=False)
metrics = {
"nll":
Loss(torch.nn.CrossEntropyLoss(ignore_index=-1),
output_transform=lambda x: (x[0][0], x[1][0])),
}
metrics.update({
"average_nll":
MetricsLambda(average_distributed_scalar, metrics["nll"]),
})
metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"])
def update(engine, batch):
model.train()
batch = tuple(t.to(device) for t in batch)
lm_loss, *_ = model(batch[0],
token_type_ids=batch[1],
lm_labels=batch[2])
loss = lm_loss / ITERATION_STEP
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
if engine.state.iteration % ITERATION_STEP == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item()
def inference(engine, batch):
model.eval()
with torch.no_grad():
batch = tuple(t.to(device) for t in batch)
input_ids, token_type_ids, lm_labels = batch
model_outputs = model(input_ids, token_type_ids=token_type_ids)
lm_logits = model_outputs[0]
lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(
-1, lm_logits.size(-1))
lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1)
return lm_logits_flat_shifted, lm_labels_flat_shifted
trainer = Engine(update)
evaluator = Engine(inference)
scheduler = PiecewiseLinear(optimizer, "lr",
[(0, 6.25e-5),
(EPOCHS * len(ds) // BATCH_SIZE, 0.0)])
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
# trainer.add_event_handler(Events.STARTED, lambda _: evaluator.run(v_dl))
if DISTRIBUTED:
trainer.add_event_handler(
Events.EPOCH_STARTED,
lambda engine: train_sampler.set_epoch(engine.state.epoch))
#evaluator.add_event_handler(Events.EPOCH_STARTED, lambda engine: valid_sampler.set_epoch(engine.state.epoch))
RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
for name, metric in metrics.items():
metric.attach(evaluator, name)
if (args.local_rank in [0, -1]):
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=["loss"])
#evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics)))
tb_logger = TensorboardLogger(log_dir='./logs')
tb_logger.attach(trainer,
log_handler=OutputHandler(tag="training",
metric_names=["loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer,
log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.ITERATION_STARTED)
#tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED)
checkpoint_handler = ModelCheckpoint('./checkpoint',
'_checkpoint',
n_saved=3)
trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler,
{'gpt2_qg': getattr(model, 'module', model)})
getattr(model, 'module', model).config.to_json_file(
os.path.join('./checkpoint', 'config'))
tokenizer.save_pretrained('./checkpoint')
trainer.run(dl, max_epochs=EPOCHS)
if (args.local_rank in [0, -1]):
tb_logger.close()
def train():
parser = ArgumentParser()
parser.add_argument("--dataset_path", type=str, default="", help="Path or url of the dataset. If empty download from S3.")
parser.add_argument("--dataset_cache", type=str, default='./dataset_cache', help="Path or url of the dataset cache")
parser.add_argument("--model_checkpoint", type=str, default="openai-gpt", help="Path, url or short name of the model")
parser.add_argument("--num_candidates", type=int, default=2, help="Number of candidates for training")
parser.add_argument("--max_history", type=int, default=2, help="Number of previous exchanges to keep in history")
parser.add_argument("--train_batch_size", type=int, default=4, help="Batch size for training")
parser.add_argument("--valid_batch_size", type=int, default=4, help="Batch size for validation")
parser.add_argument("--gradient_accumulation_steps", type=int, default=8, help="Accumulate gradients on several steps")
parser.add_argument("--lr", type=float, default=6.25e-5, help="Learning rate")
parser.add_argument("--lm_coef", type=float, default=1.0, help="LM loss coefficient")
parser.add_argument("--mc_coef", type=float, default=1.0, help="Multiple-choice loss coefficient")
parser.add_argument("--max_norm", type=float, default=1.0, help="Clipping gradient norm")
parser.add_argument("--n_epochs", type=int, default=3, help="Number of training epochs")
parser.add_argument("--personality_permutations", type=int, default=1, help="Number of permutations of personality sentences")
parser.add_argument("--eval_before_start", action='store_true', help="If true start with a first evaluation before training")
parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu", help="Device (cuda or cpu)")
parser.add_argument("--fp16", type=str, default="", help="Set to O0, O1, O2 or O3 for fp16 training (see apex documentation)")
parser.add_argument("--local_rank", type=int, default=-1, help="Local rank for distributed training (-1: not distributed)")
args = parser.parse_args()
# logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes
logging.basicConfig(level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning("Running process %d", args.local_rank) # This is a logger.warning: it will be printed by all distributed processes
logger.info("Arguments: %s", pformat(args))
# Initialize distributed training if needed
args.distributed = (args.local_rank != -1)
if args.distributed:
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
logger.info("Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning")
tokenizer_class = GPT2Tokenizer if "gpt2" in args.model_checkpoint else OpenAIGPTTokenizer
tokenizer = tokenizer_class.from_pretrained(args.model_checkpoint)
model_class = GPT2LMHeadModel if "gpt2" in args.model_checkpoint else OpenAIGPTLMHeadModel
model = model_class.from_pretrained(args.model_checkpoint)
tokenizer.set_special_tokens(SPECIAL_TOKENS)
model.set_num_special_tokens(len(SPECIAL_TOKENS))
model.to(args.device)
optimizer = OpenAIAdam(model.parameters(), lr=args.lr)
# Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last)
if args.fp16:
from apex import amp # Apex is only required if we use fp16 training
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16)
if args.distributed:
model = DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank)
logger.info("Prepare datasets")
train_loader, val_loader, train_sampler, valid_sampler = get_data_loaders(args, tokenizer)
# Training function and trainer
def update(engine, batch):
model.train()
batch = tuple(input_tensor.to(args.device) for input_tensor in batch)
lm_loss, mc_loss = model(*batch)
loss = (lm_loss * args.lm_coef + mc_loss * args.mc_coef) / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_norm)
if engine.state.iteration % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item()
trainer = Engine(update)
# Evaluation function and evaluator (evaluator output is the input of the metrics)
def inference(engine, batch):
model.eval()
with torch.no_grad():
batch = tuple(input_tensor.to(args.device) for input_tensor in batch)
input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids = batch
logger.info(tokenizer.decode(input_ids[0, -1, :].tolist()))
model_outputs = model(input_ids, mc_token_ids, token_type_ids=token_type_ids)
lm_logits, mc_logits = model_outputs[0], model_outputs[1] # So we can also use GPT2 outputs
lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(-1, lm_logits.size(-1))
lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1)
return (lm_logits_flat_shifted, mc_logits), (lm_labels_flat_shifted, mc_labels)
evaluator = Engine(inference)
# Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch
trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: evaluator.run(val_loader))
if args.n_epochs < 1:
trainer.add_event_handler(Events.COMPLETED, lambda _: evaluator.run(val_loader))
if args.eval_before_start:
trainer.add_event_handler(Events.STARTED, lambda _: evaluator.run(val_loader))
# Make sure distributed data samplers split the dataset nicely between the distributed processes
if args.distributed:
trainer.add_event_handler(Events.EPOCH_STARTED, lambda engine: train_sampler.set_epoch(engine.state.epoch))
evaluator.add_event_handler(Events.EPOCH_STARTED, lambda engine: valid_sampler.set_epoch(engine.state.epoch))
# Linearly decrease the learning rate from lr to zero
scheduler = PiecewiseLinear(optimizer, "lr", [(0, args.lr), (args.n_epochs * len(train_loader), 0.0)])
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
# Prepare metrics - note how we compute distributed metrics
RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1), output_transform=lambda x: (x[0][0], x[1][0])),
"accuracy": Accuracy(output_transform=lambda x: (x[0][1], x[1][1]))}
metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], args),
"average_accuracy": MetricsLambda(average_distributed_scalar, metrics["accuracy"], args)})
metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"])
for name, metric in metrics.items():
metric.attach(evaluator, name)
# On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train
if args.local_rank in [-1, 0]:
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=["loss"])
evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics)))
tb_logger = TensorboardLogger(log_dir=None)
tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]), event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED)
tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED)
checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir, 'checkpoint', save_interval=1, n_saved=3)
trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {'mymodel': getattr(model, 'module', model)}) # "getattr" take care of distributed encapsulation
torch.save(args, tb_logger.writer.log_dir + '/model_training_args.bin')
getattr(model, 'module', model).config.to_json_file(os.path.join(tb_logger.writer.log_dir, CONFIG_NAME))
tokenizer.save_vocabulary(tb_logger.writer.log_dir)
# Run the training
trainer.run(train_loader, max_epochs=args.n_epochs)
# On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method)
if args.local_rank in [-1, 0] and args.n_epochs > 0:
os.rename(checkpoint_handler._saved[-1][1][-1], os.path.join(tb_logger.writer.log_dir, WEIGHTS_NAME)) # TODO: PR in ignite to have better access to saved file paths (cleaner)
tb_logger.close()
def train():
parser = ArgumentParser()
parser.add_argument(
"--dataset_path",
type=str,
default='wikitext-2',
help="One of ('wikitext-103', 'wikitext-2') or a dict of splits paths."
)
parser.add_argument("--dataset_cache",
type=str,
default='./dataset_cache',
help="Path or url of the dataset cache")
parser.add_argument("--embed_dim",
type=int,
default=410,
help="Embeddings dim")
parser.add_argument("--hidden_dim",
type=int,
default=2100,
help="Hidden dimension")
parser.add_argument("--num_max_positions",
type=int,
default=256,
help="Max input length")
parser.add_argument("--num_heads",
type=int,
default=10,
help="Number of heads")
parser.add_argument("--num_layers",
type=int,
default=16,
help="NUmber of layers")
parser.add_argument("--dropout", type=float, default=0.1, help="Dropout")
parser.add_argument("--initializer_range",
type=float,
default=0.02,
help="Dropout")
parser.add_argument("--train_batch_size",
type=int,
default=8,
help="Batch size for training")
parser.add_argument("--valid_batch_size",
type=int,
default=8,
help="Batch size for validation")
parser.add_argument("--lr",
type=float,
default=2.5e-4,
help="Learning rate")
parser.add_argument("--max_norm",
type=float,
default=0.25,
help="Clipping gradient norm")
parser.add_argument("--weight_decay",
type=float,
default=0.0,
help="Weight decay")
parser.add_argument("--n_epochs",
type=int,
default=200,
help="Number of training epochs")
parser.add_argument("--n_warmup",
type=float,
default=1000,
help="Number of warmup iterations")
parser.add_argument("--eval_every",
type=int,
default=-1,
help="Evaluate every X steps (-1 => end of epoch)")
parser.add_argument("--gradient_accumulation_steps",
type=int,
default=1,
help="Accumulate gradient")
parser.add_argument("--device",
type=str,
default="cuda" if torch.cuda.is_available() else "cpu",
help="Device (cuda or cpu)")
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="Local rank for distributed training (-1: not distributed)")
args = parser.parse_args()
# logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log on main process only, logger.warning => log on all processes
logging.basicConfig(
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning(
"Running process %d", args.local_rank
) # This is a logger.warning: it will be printed by all distributed processes
logger.info("Arguments: %s", pformat(
args)) # This is a logger.info: only printed on the first process
# Initialize distributed training if needed
args.distributed = (args.local_rank != -1)
if args.distributed:
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
logger.info("Prepare tokenizer, model and optimizer")
tokenizer = BertTokenizer.from_pretrained(
'bert-base-cased',
do_lower_case=False) # Let's use a pre-defined tokenizer
args.num_embeddings = len(
tokenizer.vocab
) # We need this to create the model at next line (number of embeddings to use)
model = TransformerWithLMHead(args)
model.to(args.device)
optimizer = Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
logger.info("Model has %s parameters",
sum(p.numel() for p in model.parameters() if p.requires_grad))
# Prepare model for distributed training if needed
if args.distributed:
model = DistributedDataParallel(model,
device_ids=[args.local_rank],
output_device=args.local_rank)
logger.info("Prepare datasets")
train_loader, val_loader, train_sampler, valid_sampler, train_num_words, valid_num_words = get_data_loaders(
args, tokenizer)
# Training function and trainer
def update(engine, batch):
model.train()
batch = batch.transpose(0, 1).contiguous().to(
args.device) # to shape [seq length, batch]
logits, loss = model(batch, labels=batch)
loss = loss / args.gradient_accumulation_steps
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_norm)
if engine.state.iteration % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item()
trainer = Engine(update)
# Evaluation function and evaluator (evaluator output is the input of the metrics)
def inference(engine, batch):
model.eval()
with torch.no_grad():
batch = batch.transpose(0, 1).contiguous().to(
args.device) # to shape [seq length, batch]
logits = model(batch)
shift_logits = logits[:-1].view(-1, logits.size(-1))
shift_labels = batch[1:].view(-1)
return shift_logits, shift_labels
evaluator = Engine(inference)
# Attach evaluation to trainer: we evaluate at the end of each epoch and every 'eval_every' iterations if needed
trainer.add_event_handler(Events.EPOCH_COMPLETED,
lambda _: evaluator.run(val_loader))
if args.eval_every > 0:
trainer.add_event_handler(
Events.ITERATION_COMPLETED,
lambda engine: evaluator.run(val_loader)
if engine.state.iteration % args.eval_every == 0 else None)
if args.n_epochs < 1:
trainer.add_event_handler(Events.COMPLETED,
lambda _: evaluator.run(val_loader))
# Make sure distributed data samplers split the dataset nicely between the distributed processes
if args.distributed:
trainer.add_event_handler(
Events.EPOCH_STARTED,
lambda engine: train_sampler.set_epoch(engine.state.epoch))
evaluator.add_event_handler(
Events.EPOCH_STARTED,
lambda engine: valid_sampler.set_epoch(engine.state.epoch))
# Learning rate schedule: linearly warm-up to lr and then decrease the learning rate to zero with cosine schedule
cos_scheduler = CosineAnnealingScheduler(optimizer, 'lr', args.lr, 0.0,
len(train_loader) * args.n_epochs)
scheduler = create_lr_scheduler_with_warmup(cos_scheduler, 0.0, args.lr,
args.n_warmup)
trainer.add_event_handler(Events.ITERATION_STARTED, scheduler)
# Prepare metrics - note how we average distributed metrics using average_distributed_scalar
RunningAverage(output_transform=lambda x: x).attach(trainer, "loss")
metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1))}
metrics.update({
"average_nll":
MetricsLambda(average_distributed_scalar, metrics["nll"], args)
})
metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"])
metrics["average_word_ppl"] = MetricsLambda(
lambda x: math.exp(x * val_loader.dataset.numel() / valid_num_words),
metrics["average_nll"])
for name, metric in metrics.items():
metric.attach(evaluator, name)
# On the main process: add progress bar, tensorboard, checkpoints and save model and configuration before we start to train
if args.local_rank in [-1, 0]:
pbar = ProgressBar(persist=True)
pbar.attach(trainer, metric_names=["loss"])
evaluator.add_event_handler(
Events.COMPLETED, lambda _: pbar.log_message(
"Validation: %s" % pformat(evaluator.state.metrics)))
tb_logger = TensorboardLogger(log_dir=None)
tb_logger.attach(trainer,
log_handler=OutputHandler(tag="training",
metric_names=["loss"]),
event_name=Events.ITERATION_COMPLETED)
tb_logger.attach(trainer,
log_handler=OptimizerParamsHandler(optimizer),
event_name=Events.ITERATION_STARTED)
@evaluator.on(Events.COMPLETED) # Log evaluator metrics on tensorboard
def tb_log_metrics(engine):
for name in metrics.keys():
tb_logger.writer.add_scalar(name, engine.state.metrics[name],
trainer.state.iteration)
checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir,
'checkpoint',
save_interval=1,
n_saved=3)
trainer.add_event_handler(
Events.EPOCH_COMPLETED, checkpoint_handler,
{'mymodel': getattr(model, 'module', model)
}) # "getattr" take care of distributed encapsulation
torch.save(args, os.path.join(tb_logger.writer.log_dir, CONFIG_NAME))
# Run the training
trainer.run(train_loader, max_epochs=args.n_epochs)
# On the main process: close tensorboard logger and rename the last checkpoint for easy re-loading
if args.local_rank in [-1, 0] and args.n_epochs > 0:
os.rename(
checkpoint_handler._saved[-1][1][-1],
os.path.join(tb_logger.writer.log_dir, WEIGHTS_NAME)
) # TODO: PR in ignite to have better access to saved file paths (cleaner)
tb_logger.close()
