def __init__(self, corpus, base_path, max_epochs, evaluation_metric,
training_runs, optimization_value):
if (type(base_path) is unicode):
base_path = Path(base_path)
self.corpus = corpus
self.max_epochs = max_epochs
self.base_path = base_path
self.evaluation_metric = evaluation_metric
self.run = 1
self.training_runs = training_runs
self.optimization_value = optimization_value
self.param_selection_file = init_output_file(base_path,
u'param_selection.txt')
def __init__(self, corpus: Corpus, base_path: Union[(str, Path)],
max_epochs: int, evaluation_metric: EvaluationMetric,
training_runs: int, optimization_value: OptimizationValue):
if (type(base_path) is str):
base_path = Path(base_path)
self.corpus = corpus
self.max_epochs = max_epochs
self.base_path = base_path
self.evaluation_metric = evaluation_metric
self.run = 1
self.training_runs = training_runs
self.optimization_value = optimization_value
self.param_selection_file = init_output_file(base_path,
'param_selection.txt')
def train(
self,
base_path: Union[Path, str],
learning_rate: float = 0.1,
mini_batch_size: int = 32,
eval_mini_batch_size: int = None,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 3,
min_learning_rate: float = 0.0001,
train_with_dev: bool = False,
monitor_train: bool = False,
monitor_test: bool = False,
embeddings_storage_mode: str = "cpu",
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
num_workers: int = 6,
sampler=None,
**kwargs,
) -> dict:
"""
Trains any class that implements the flair.nn.Model interface.
:param base_path: Main path to which all output during training is logged and models are saved
:param learning_rate: Initial learning rate
:param mini_batch_size: Size of mini-batches during training
:param eval_mini_batch_size: Size of mini-batches during evaluation
:param max_epochs: Maximum number of epochs to train. Terminates training if this number is surpassed.
:param anneal_factor: The factor by which the learning rate is annealed
:param patience: Patience is the number of epochs with no improvement the Trainer waits
until annealing the learning rate
:param min_learning_rate: If the learning rate falls below this threshold, training terminates
:param train_with_dev: If True, training is performed using both train+dev data
:param monitor_train: If True, training data is evaluated at end of each epoch
:param monitor_test: If True, test data is evaluated at end of each epoch
:param embeddings_storage_mode: One of 'none' (all embeddings are deleted and freshly recomputed),
'cpu' (embeddings are stored on CPU) or 'gpu' (embeddings are stored on GPU)
:param checkpoint: If True, a full checkpoint is saved at end of each epoch
:param save_final_model: If True, final model is saved
:param anneal_with_restarts: If True, the last best model is restored when annealing the learning rate
:param shuffle: If True, data is shuffled during training
:param param_selection_mode: If True, testing is performed against dev data. Use this mode when doing
parameter selection.
:param num_workers: Number of workers in your data loader.
:param sampler: You can pass a data sampler here for special sampling of data.
:param kwargs: Other arguments for the Optimizer
:return:
"""
if self.use_tensorboard:
try:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter()
except:
log_line(log)
log.warning(
"ATTENTION! PyTorch >= 1.1.0 and pillow are required for TensorBoard support!"
)
log_line(log)
self.use_tensorboard = False
pass
if eval_mini_batch_size is None:
eval_mini_batch_size = mini_batch_size
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
log_handler = add_file_handler(log, base_path / "training.log")
log_line(log)
log.info(f'Model: "{self.model}"')
log_line(log)
log.info(f'Corpus: "{self.corpus}"')
log_line(log)
log.info("Parameters:")
log.info(f' - learning_rate: "{learning_rate}"')
log.info(f' - mini_batch_size: "{mini_batch_size}"')
log.info(f' - patience: "{patience}"')
log.info(f' - anneal_factor: "{anneal_factor}"')
log.info(f' - max_epochs: "{max_epochs}"')
log.info(f' - shuffle: "{shuffle}"')
log.info(f' - train_with_dev: "{train_with_dev}"')
log_line(log)
log.info(f'Model training base path: "{base_path}"')
log_line(log)
log.info(f"Device: {flair.device}")
log_line(log)
log.info(f"Embeddings storage mode: {embeddings_storage_mode}")
# determine what splits (train, dev, test) to evaluate and log
log_train = True if monitor_train else False
log_test = (True if (not param_selection_mode and self.corpus.test
and monitor_test) else False)
log_dev = True if not train_with_dev else False
# prepare loss logging file and set up header
loss_txt = init_output_file(base_path, "loss.tsv")
weight_extractor = WeightExtractor(base_path)
optimizer: torch.optim.Optimizer = self.optimizer(
self.model.parameters(), lr=learning_rate, **kwargs)
if self.optimizer_state is not None:
optimizer.load_state_dict(self.optimizer_state)
# minimize training loss if training with dev data, else maximize dev score
anneal_mode = "min" if train_with_dev else "max"
scheduler: ReduceLROnPlateau = ReduceLROnPlateau(
optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True,
)
if self.scheduler_state is not None:
scheduler.load_state_dict(self.scheduler_state)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data = ConcatDataset([self.corpus.train, self.corpus.dev])
if sampler is not None:
sampler = sampler(train_data)
shuffle = False
dev_score_history = []
dev_loss_history = []
train_loss_history = []
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for epoch in range(0 + self.epoch, max_epochs + self.epoch):
log_line(log)
# get new learning rate
for group in optimizer.param_groups:
learning_rate = group["lr"]
# reload last best model if annealing with restarts is enabled
if (learning_rate != previous_learning_rate
and anneal_with_restarts
and (base_path / "best-model.pt").exists()):
log.info("resetting to best model")
self.model.load(base_path / "best-model.pt")
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < min_learning_rate:
log_line(log)
log.info("learning rate too small - quitting training!")
log_line(log)
break
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
sampler=sampler,
)
self.model.train()
train_loss: float = 0
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int(total_number_of_batches / 10))
# process mini-batches
for batch_no, batch in enumerate(batch_loader):
loss = self.model.forward_loss(batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_batches += 1
train_loss += loss.item()
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(batch, embeddings_storage_mode)
if batch_no % modulo == 0:
log.info(
f"epoch {epoch + 1} - iter {batch_no}/{total_number_of_batches} - loss "
f"{train_loss / seen_batches:.8f}")
iteration = epoch * total_number_of_batches + batch_no
if not param_selection_mode:
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
train_loss /= seen_batches
self.model.eval()
log_line(log)
log.info(
f"EPOCH {epoch + 1} done: loss {train_loss:.4f} - lr {learning_rate:.4f}"
)
if self.use_tensorboard:
writer.add_scalar("train_loss", train_loss, epoch + 1)
# anneal against train loss if training with dev, otherwise anneal against dev score
current_score = train_loss
# evaluate on train / dev / test split depending on training settings
result_line: str = ""
if log_train:
train_eval_result, train_loss = self.model.evaluate(
DataLoader(
self.corpus.train,
batch_size=eval_mini_batch_size,
num_workers=num_workers,
),
embeddings_storage_mode=embeddings_storage_mode,
)
result_line += f"\t{train_eval_result.log_line}"
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.train,
embeddings_storage_mode)
if log_dev:
dev_eval_result, dev_loss = self.model.evaluate(
DataLoader(
self.corpus.dev,
batch_size=eval_mini_batch_size,
num_workers=num_workers,
),
embeddings_storage_mode=embeddings_storage_mode,
)
result_line += f"\t{dev_loss}\t{dev_eval_result.log_line}"
log.info(
f"DEV : loss {dev_loss} - score {dev_eval_result.main_score}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_score_history.append(dev_eval_result.main_score)
dev_loss_history.append(dev_loss)
current_score = dev_eval_result.main_score
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.dev, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("dev_loss", dev_loss, epoch + 1)
writer.add_scalar("dev_score",
dev_eval_result.main_score,
epoch + 1)
if log_test:
test_eval_result, test_loss = self.model.evaluate(
DataLoader(
self.corpus.test,
batch_size=eval_mini_batch_size,
num_workers=num_workers,
),
base_path / "test.tsv",
embeddings_storage_mode=embeddings_storage_mode,
)
result_line += f"\t{test_loss}\t{test_eval_result.log_line}"
log.info(
f"TEST : loss {test_loss} - score {test_eval_result.main_score}"
)
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.test, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("test_loss", test_loss, epoch + 1)
writer.add_scalar("test_score",
test_eval_result.main_score,
epoch + 1)
# determine learning rate annealing through scheduler
scheduler.step(current_score)
train_loss_history.append(train_loss)
# determine bad epoch number
try:
bad_epochs = scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
new_learning_rate = group["lr"]
if new_learning_rate != previous_learning_rate:
bad_epochs = patience + 1
# log bad epochs
log.info(f"BAD EPOCHS (no improvement): {bad_epochs}")
# output log file
with open(loss_txt, "a") as f:
# make headers on first epoch
if epoch == 0:
f.write(
f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS"
)
if log_train:
f.write("\tTRAIN_" + "\tTRAIN_".join(
train_eval_result.log_header.split("\t")))
if log_dev:
f.write("\tDEV_LOSS\tDEV_" + "\tDEV_".join(
dev_eval_result.log_header.split("\t")))
if log_test:
f.write("\tTEST_LOSS\tTEST_" + "\tTEST_".join(
test_eval_result.log_header.split("\t")))
f.write(
f"\n{epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t{train_loss}"
)
f.write(result_line)
# if checkpoint is enable, save model at each epoch
if checkpoint and not param_selection_mode:
self.model.save_checkpoint(
base_path / "checkpoint.pt",
optimizer.state_dict(),
scheduler.state_dict(),
epoch + 1,
train_loss,
)
# if we use dev data, remember best model based on dev evaluation score
if (not train_with_dev and not param_selection_mode
and current_score == scheduler.best):
self.model.save(base_path / "best-model.pt")
# if we do not use dev data for model selection, save final model
if save_final_model and not param_selection_mode:
self.model.save(base_path / "final-model.pt")
except KeyboardInterrupt:
log_line(log)
log.info("Exiting from training early.")
if self.use_tensorboard:
writer.close()
if not param_selection_mode:
log.info("Saving model ...")
self.model.save(base_path / "final-model.pt")
log.info("Done.")
# test best model if test data is present
if self.corpus.test:
final_score = self.final_test(base_path, eval_mini_batch_size,
num_workers)
else:
final_score = 0
log.info("Test data not provided setting final score to 0")
log.removeHandler(log_handler)
if self.use_tensorboard:
writer.close()
return {
"test_score": final_score,
"dev_score_history": dev_score_history,
"train_loss_history": train_loss_history,
"dev_loss_history": dev_loss_history,
}
from __future__ import absolute_import
def find_learning_rate(self,
base_path: Union[Path, str],
file_name: str = 'learning_rate.tsv',
start_learning_rate: float = 1e-7,
end_learning_rate: float = 10,
iterations: int = 100,
mini_batch_size: int = 32,
stop_early: bool = True,
smoothing_factor: float = 0.98,
**kwargs) -> Path:
best_loss = None
moving_avg_loss = 0
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
learning_rate_tsv = init_output_file(base_path, file_name)
with open(learning_rate_tsv, 'a') as f:
f.write('ITERATION\tTIMESTAMP\tLEARNING_RATE\tTRAIN_LOSS\n')
optimizer = self.optimizer(self.model.parameters(),
lr=start_learning_rate,
**kwargs)
train_data = self.corpus.train
random.shuffle(train_data)
batches = [
train_data[x:x + mini_batch_size]
for x in range(0, len(train_data), mini_batch_size)
][:iterations]
scheduler = ExpAnnealLR(optimizer, end_learning_rate, iterations)
model_state = self.model.state_dict()
model_device = next(self.model.parameters()).device
self.model.train()
for itr, batch in enumerate(batches):
loss = self.model.forward_loss(batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 5.0)
optimizer.step()
scheduler.step()
learning_rate = scheduler.get_lr()[0]
loss_item = loss.item()
if itr == 0:
best_loss = loss_item
else:
if smoothing_factor > 0:
moving_avg_loss = smoothing_factor * moving_avg_loss + (
1 - smoothing_factor) * loss_item
loss_item = moving_avg_loss / (1 -
smoothing_factor**(itr + 1))
if loss_item < best_loss:
best_loss = loss
if stop_early and (loss_item > 4 * best_loss or torch.isnan(loss)):
log_line(log)
log.info('loss diverged - stopping early!')
break
with open(learning_rate_tsv, 'a') as f:
f.write(
f'{itr}\t{datetime.datetime.now():%H:%M:%S}\t{learning_rate}\t{loss_item}\n'
)
self.model.load_state_dict(model_state)
self.model.to(model_device)
log_line(log)
log.info(f'learning rate finder finished - plot {learning_rate_tsv}')
log_line(log)
return Path(learning_rate_tsv)
def train(self,
base_path: Union[Path, str],
evaluation_metric: EvaluationMetric = EvaluationMetric.
MICRO_F1_SCORE,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
eval_mini_batch_size: int = None,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 3,
anneal_against_train_loss: bool = True,
train_with_dev: bool = False,
monitor_train: bool = False,
embeddings_in_memory: bool = True,
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
test_mode: bool = False,
param_selection_mode: bool = False,
**kwargs) -> dict:
if eval_mini_batch_size is None:
eval_mini_batch_size = mini_batch_size
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
add_file_handler(log, base_path / 'training.log')
log_line(log)
log.info(f'Evaluation method: {evaluation_metric.name}')
if not param_selection_mode:
loss_txt = init_output_file(base_path, 'loss.tsv')
with open(loss_txt, 'a') as f:
f.write(
f'EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS\t{Metric.tsv_header("TRAIN")}\tDEV_LOSS\t{Metric.tsv_header("DEV")}'
f'\tTEST_LOSS\t{Metric.tsv_header("TEST")}\n')
weight_extractor = WeightExtractor(base_path)
optimizer = self.optimizer(self.model.parameters(),
lr=learning_rate,
**kwargs)
if self.optimizer_state is not None:
optimizer.load_state_dict(self.optimizer_state)
# annealing scheduler
anneal_mode = 'min' if anneal_against_train_loss else 'max'
if isinstance(optimizer, (AdamW, SGDW)):
scheduler = ReduceLRWDOnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True)
else:
scheduler = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True)
if self.scheduler_state is not None:
scheduler.load_state_dict(self.scheduler_state)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data.extend(self.corpus.dev)
dev_score_history = []
dev_loss_history = []
train_loss_history = []
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for epoch in range(0 + self.epoch, max_epochs + self.epoch):
log_line(log)
try:
bad_epochs = scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
learning_rate = group['lr']
# reload last best model if annealing with restarts is enabled
if learning_rate != previous_learning_rate and anneal_with_restarts and \
(base_path / 'best-model.pt').exists():
log.info('resetting to best model')
self.model.load_from_file(base_path / 'best-model.pt')
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < 0.0001:
log_line(log)
log.info('learning rate too small - quitting training!')
log_line(log)
break
if not test_mode:
random.shuffle(train_data)
batches = [
train_data[x:x + mini_batch_size]
for x in range(0, len(train_data), mini_batch_size)
]
self.model.train()
train_loss: float = 0
seen_sentences = 0
modulo = max(1, int(len(batches) / 10))
for batch_no, batch in enumerate(batches):
loss = self.model.forward_loss(batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_sentences += len(batch)
train_loss += loss.item()
clear_embeddings(
batch,
also_clear_word_embeddings=not embeddings_in_memory)
if batch_no % modulo == 0:
log.info(
f'epoch {epoch + 1} - iter {batch_no}/{len(batches)} - loss '
f'{train_loss / seen_sentences:.8f}')
iteration = epoch * len(batches) + batch_no
if not param_selection_mode:
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
train_loss /= len(train_data)
self.model.eval()
log_line(log)
log.info(
f'EPOCH {epoch + 1} done: loss {train_loss:.4f} - lr {learning_rate:.4f} - bad epochs {bad_epochs}'
)
dev_metric = None
dev_loss = '_'
train_metric = None
if monitor_train:
train_metric, train_loss = self._calculate_evaluation_results_for(
'TRAIN', self.corpus.train, evaluation_metric,
embeddings_in_memory, eval_mini_batch_size)
if not train_with_dev:
dev_metric, dev_loss = self._calculate_evaluation_results_for(
'DEV', self.corpus.dev, evaluation_metric,
embeddings_in_memory, eval_mini_batch_size)
if not param_selection_mode:
test_metric, test_loss = self._calculate_evaluation_results_for(
'TEST', self.corpus.test, evaluation_metric,
embeddings_in_memory, eval_mini_batch_size,
base_path / 'test.tsv')
if not param_selection_mode:
with open(loss_txt, 'a') as f:
train_metric_str = train_metric.to_tsv(
) if train_metric is not None else Metric.to_empty_tsv(
)
dev_metric_str = dev_metric.to_tsv(
) if dev_metric is not None else Metric.to_empty_tsv()
test_metric_str = test_metric.to_tsv(
) if test_metric is not None else Metric.to_empty_tsv(
)
f.write(
f'{epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t'
f'{train_loss}\t{train_metric_str}\t{dev_loss}\t{dev_metric_str}\t_\t{test_metric_str}\n'
)
# calculate scores using dev data if available
dev_score = 0.
if not train_with_dev:
if evaluation_metric == EvaluationMetric.MACRO_ACCURACY:
dev_score = dev_metric.macro_avg_accuracy()
elif evaluation_metric == EvaluationMetric.MICRO_ACCURACY:
dev_score = dev_metric.micro_avg_accuracy()
elif evaluation_metric == EvaluationMetric.MACRO_F1_SCORE:
dev_score = dev_metric.macro_avg_f_score()
else:
dev_score = dev_metric.micro_avg_f_score()
# append dev score to score history
dev_score_history.append(dev_score)
dev_loss_history.append(dev_loss.item())
# anneal against train loss if training with dev, otherwise anneal against dev score
current_score = train_loss if anneal_against_train_loss else dev_score
scheduler.step(current_score)
train_loss_history.append(train_loss)
# if checkpoint is enable, save model at each epoch
if checkpoint and not param_selection_mode:
self.model.save_checkpoint(base_path / 'checkpoint.pt',
optimizer.state_dict(),
scheduler.state_dict(),
epoch + 1, train_loss)
# if we use dev data, remember best model based on dev evaluation score
if not train_with_dev and not param_selection_mode and current_score == scheduler.best:
self.model.save(base_path / 'best-model.pt')
# if we do not use dev data for model selection, save final model
if save_final_model and not param_selection_mode:
self.model.save(base_path / 'final-model.pt')
except KeyboardInterrupt:
log_line(log)
log.info('Exiting from training early.')
if not param_selection_mode:
log.info('Saving model ...')
self.model.save(base_path / 'final-model.pt')
log.info('Done.')
# test best model on test data
final_score = self.final_test(base_path, embeddings_in_memory,
evaluation_metric, eval_mini_batch_size)
return {
'test_score': final_score,
'dev_score_history': dev_score_history,
'train_loss_history': train_loss_history,
'dev_loss_history': dev_loss_history
}
def train(
self,
base_path: Union[Path, str],
evaluation_metric: EvaluationMetric = EvaluationMetric.MICRO_F1_SCORE,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
eval_mini_batch_size: int = None,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 3,
train_with_dev: bool = False,
monitor_train: bool = False,
embeddings_in_memory: bool = True,
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
num_workers: int = 8,
valid_with_misspellings: bool = True,
**kwargs,
) -> dict:
if eval_mini_batch_size is None:
eval_mini_batch_size = mini_batch_size
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
log_handler = add_file_handler(log, base_path / "training.log")
log_line(log)
log.info(f'Model: "{self.model}"')
log_line(log)
log.info(f'Corpus: "{self.corpus}"')
log_line(log)
log.info("Parameters:")
log.info(f' - learning_rate: "{learning_rate}"')
log.info(f' - mini_batch_size: "{mini_batch_size}"')
log.info(f' - patience: "{patience}"')
log.info(f' - anneal_factor: "{anneal_factor}"')
log.info(f' - max_epochs: "{max_epochs}"')
log.info(f' - shuffle: "{shuffle}"')
log.info(f' - train_with_dev: "{train_with_dev}"')
log.info(f' - valid_with_misspellings: "{valid_with_misspellings}"')
log.info("Model:")
log.info(f' - hidden_size: "{self.model.hidden_size}"')
log.info(f' - train_mode: "{self.model.train_mode}"')
log.info(f' - alpha: "{self.model.alpha}"')
log.info(f' - misspell_mode: "{self.model.misspell_mode}"')
log.info(f' - misspelling_rate: "{self.model.misspelling_rate_train}"')
log.info(f' - cmx_file: "{self.model.cmx_file_train}"')
log_line(log)
log.info(f'Model training base path: "{base_path}"')
log_line(log)
log.info(f"Evaluation method: {evaluation_metric.name}")
# determine what splits (train, dev, test) to evaluate and log
log_train = True if monitor_train else False
log_test = True if (not param_selection_mode
and self.corpus.test) else False
log_dev = True if not train_with_dev else False
log_test = not log_dev
eval_misspelling_rate = 0.05
log_suffix = lambda prefix, rate, cm, mode: f"{prefix} (misspell: cmx={cm})" if mode == MisspellingMode.ConfusionMatrixBased else f"{prefix} (misspell: rate={rate})"
loss_txt = init_output_file(base_path, "loss.tsv")
with open(loss_txt, "a") as f:
f.write(f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS")
dummy_result, _ = self.model.evaluate(
[Sentence("d", labels=["0.1"])],
eval_mini_batch_size,
embeddings_in_memory,
)
if log_train:
f.write("\tTRAIN_" +
"\tTRAIN_".join(dummy_result.log_header.split("\t")))
if log_dev:
f.write("\tDEV_LOSS\tDEV_" +
"\tDEV_".join(dummy_result.log_header.split("\t")))
if valid_with_misspellings:
suffix = log_suffix('DEV', eval_misspelling_rate,
self.model.cmx_file_train,
self.model.misspell_mode)
f.write(f"\t{suffix}" + f"_LOSS\t{suffix})_" +
f"\t{suffix}_".join(
dummy_result.log_header.split("\t")))
if log_test:
f.write("\tTEST_LOSS\tTEST_" +
"\tTEST_".join(dummy_result.log_header.split("\t")))
if valid_with_misspellings:
suffix = log_suffix('TEST', eval_misspelling_rate,
self.model.cmx_file_train,
self.model.misspell_mode)
f.write(f"\t{suffix}" + f"_LOSS\t{suffix})_" +
f"\t{suffix}_".join(
dummy_result.log_header.split("\t")))
weight_extractor = WeightExtractor(base_path)
optimizer = self.optimizer(self.model.parameters(),
lr=learning_rate,
**kwargs)
if self.optimizer_state is not None:
optimizer.load_state_dict(self.optimizer_state)
# minimize training loss if training with dev data, else maximize dev score
anneal_mode = "min" if train_with_dev else "max"
if isinstance(optimizer, (AdamW, SGDW)):
scheduler = ReduceLRWDOnPlateau(
optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True,
)
else:
scheduler = ReduceLROnPlateau(
optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True,
)
if self.scheduler_state is not None:
scheduler.load_state_dict(self.scheduler_state)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data = ConcatDataset([self.corpus.train, self.corpus.dev])
dev_clean_score_history = []
dev_noisy_score_history = []
dev_clean_loss_history = []
dev_noisy_loss_history = []
train_loss_history = []
complete_data = ConcatDataset(
[self.corpus.train, self.corpus.dev, self.corpus.test])
char_vocab = make_char_vocab(complete_data)
log.info(
f"Vocabulary of the corpus (#{len(char_vocab)}): {char_vocab}")
if self.model.misspell_mode == MisspellingMode.ConfusionMatrixBased:
cmx, lut = load_confusion_matrix(self.model.cmx_file_train)
cmx, lut = filter_cmx(cmx, lut, char_vocab)
else:
cmx, lut = None, {}
loss_params = {}
loss_params["verbose"] = False
loss_params["char_vocab"] = char_vocab
loss_params["cmx"] = cmx
loss_params["lut"] = lut
loss_params["embeddings_in_memory"] = embeddings_in_memory
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for epoch in range(0 + self.epoch, max_epochs + self.epoch):
log_line(log)
try:
bad_epochs = scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
learning_rate = group["lr"]
# reload last best model if annealing with restarts is enabled
if (learning_rate != previous_learning_rate
and anneal_with_restarts
and (base_path / "best-model.pt").exists()):
log.info("resetting to best model")
self.model.load(base_path / "best-model.pt")
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < 0.0001:
log_line(log)
log.info("learning rate too small - quitting training!")
log_line(log)
break
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
)
self.model.train()
train_loss: float = 0
train_auxilary_losses = {}
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int(total_number_of_batches / 10))
for batch_no, batch in enumerate(batch_loader):
loss, auxilary_losses = self.model.forward_loss(
batch, params=loss_params)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_batches += 1
train_loss += loss.item()
for k, v in auxilary_losses.items():
train_auxilary_losses[k] = train_auxilary_losses.get(
k, 0) + v
clear_embeddings(
batch,
also_clear_word_embeddings=not embeddings_in_memory)
if batch_no % modulo == 0:
msg = f"epoch {epoch + 1} - iter {batch_no}/{total_number_of_batches} - loss {train_loss / seen_batches:.6f}"
# note: this is the loss accumulated in the current epoch divided by the number of already seen batches
if len(train_auxilary_losses) > 0:
aux_losses_str = " ".join([
f"{key}={value / seen_batches:.6f}"
for (key,
value) in train_auxilary_losses.items()
])
msg += f" ({aux_losses_str})"
log.info(msg)
iteration = epoch * total_number_of_batches + batch_no
if not param_selection_mode:
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
train_loss /= seen_batches
for k, v in auxilary_losses.items():
train_auxilary_losses[k] /= seen_batches
self.model.eval()
log_line(log)
log.info(
f"EPOCH {epoch + 1} done: loss {train_loss:.6f} - lr {learning_rate:.4f} - bad epochs {bad_epochs}"
)
# anneal against train loss if training with dev, otherwise anneal against dev score
current_score = train_loss
with open(loss_txt, "a") as f:
f.write(
f"\n{epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t{train_loss}"
)
if log_train:
train_eval_result, train_loss = self.model.evaluate(
self.corpus.train,
eval_mini_batch_size,
embeddings_in_memory,
num_workers=num_workers,
)
f.write(f"\t{train_eval_result.log_line}")
if log_dev:
dev_eval_result_clean, dev_loss_clean = self.model.evaluate(
self.corpus.dev,
eval_mini_batch_size,
embeddings_in_memory,
num_workers=num_workers,
)
f.write(
f"\t{dev_loss_clean}\t{dev_eval_result_clean.log_line}"
)
log.info(
f"DEV : loss {dev_loss_clean:.6f} - score {dev_eval_result_clean.main_score:.4f}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_clean_score_history.append(
dev_eval_result_clean.main_score)
dev_clean_loss_history.append(dev_loss_clean)
if valid_with_misspellings:
# evaluate on misspellings
dev_eval_result_noisy, dev_loss_noisy = self.model.evaluate(
self.corpus.dev,
eval_mini_batch_size,
embeddings_in_memory,
num_workers=num_workers,
eval_mode=EvalMode.Misspellings,
misspell_mode=self.model.misspell_mode,
char_vocab=char_vocab,
cmx=cmx,
lut=lut,
misspelling_rate=eval_misspelling_rate,
)
f.write(
f"\t{dev_loss_noisy}\t{dev_eval_result_noisy.log_line}"
)
log.info(
f"{log_suffix('DEV', eval_misspelling_rate, self.model.cmx_file_train, self.model.misspell_mode)}"
+
f" : loss {dev_loss_noisy:.6f} - score {dev_eval_result_noisy.main_score:.4f}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_noisy_score_history.append(
dev_eval_result_noisy)
dev_noisy_loss_history.append(dev_loss_noisy)
current_score = (
dev_eval_result_clean.main_score +
dev_eval_result_noisy.main_score) / 2.0
else:
current_score = dev_eval_result_clean.main_score
if log_test:
test_eval_result_clean, test_loss_clean = self.model.evaluate(
self.corpus.test,
eval_mini_batch_size,
embeddings_in_memory,
base_path / f"test.tsv",
num_workers=num_workers,
)
f.write(
f"\t{test_loss_clean}\t{test_eval_result_clean.log_line}"
)
log.info(
f"TEST : loss {test_loss_clean:.6f} - score {test_eval_result_clean.main_score:.4f}"
)
if valid_with_misspellings:
# evaluate on misspellings
test_eval_result_noisy, test_loss_noisy = self.model.evaluate(
self.corpus.test,
eval_mini_batch_size,
embeddings_in_memory,
base_path / f"test.tsv",
num_workers=num_workers,
eval_mode=EvalMode.Misspellings,
misspell_mode=self.model.misspell_mode,
char_vocab=char_vocab,
cmx=cmx,
lut=lut,
misspelling_rate=eval_misspelling_rate,
)
f.write(
f"\t{test_loss_noisy}\t{test_eval_result_noisy.log_line}"
)
log.info(
f"{log_suffix('TEST', eval_misspelling_rate, self.model.cmx_file_train, self.model.misspell_mode)}"
+
f" : loss {test_loss_noisy:.6f} - score {test_eval_result_noisy.main_score:.4f}"
#f"TEST (misspell, rate={eval_misspelling_rate}) : loss {test_loss_noisy:.6f} - score {test_eval_result_noisy.main_score:.4f}"
)
scheduler.step(current_score)
train_loss_history.append(train_loss)
# if checkpoint is enable, save model at each epoch
if checkpoint and not param_selection_mode:
self.model.save_checkpoint(
base_path / "checkpoint.pt",
optimizer.state_dict(),
scheduler.state_dict(),
epoch + 1,
train_loss,
)
# if we use dev data, remember best model based on dev evaluation score
if (not train_with_dev and not param_selection_mode
and current_score == scheduler.best):
log.info("'best-model.pt' saved.")
self.model.save(base_path / "best-model.pt")
# if we do not use dev data for model selection, save final model
if save_final_model and not param_selection_mode:
self.model.save(base_path / "final-model.pt")
except KeyboardInterrupt:
log_line(log)
log.info("Exiting from training early.")
if not param_selection_mode:
log.info("Saving model ...")
self.model.save(base_path / "final-model.pt")
log.info("Done.")
# test best model if test data is present
if self.corpus.test:
final_score_clean = self.final_test(
base_path,
embeddings_in_memory,
evaluation_metric,
eval_mini_batch_size,
num_workers,
)
final_score_noisy = self.final_test(
base_path,
embeddings_in_memory,
evaluation_metric,
eval_mini_batch_size,
num_workers,
eval_mode=EvalMode.Misspellings,
misspell_mode=self.model.misspell_mode,
misspelling_rate=eval_misspelling_rate,
char_vocab=char_vocab,
cmx=cmx,
lut=lut,
)
else:
final_score_clean, final_score_noisy = 0, 0
log.info("Test data not provided setting final score to 0")
log.removeHandler(log_handler)
return {
"test_score_clean": final_score_clean,
"test_score_noisy": final_score_noisy,
"dev_clean_score_history": dev_clean_score_history,
"dev_noisy_score_history": dev_noisy_score_history,
"train_loss_history": train_loss_history,
"dev_clean_loss_history": dev_clean_loss_history,
"dev_noisy_loss_history": dev_noisy_loss_history,
}
def find_learning_rate(
self,
base_path: Union[Path, str],
file_name: str = "learning_rate.tsv",
start_learning_rate: float = 1e-7,
end_learning_rate: float = 10,
iterations: int = 100,
mini_batch_size: int = 32,
stop_early: bool = True,
smoothing_factor: float = 0.98,
**kwargs,
) -> Path:
best_loss = None
moving_avg_loss = 0
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
learning_rate_tsv = init_output_file(base_path, file_name)
with open(learning_rate_tsv, "a") as f:
f.write("ITERATION\tTIMESTAMP\tLEARNING_RATE\tTRAIN_LOSS\n")
optimizer = self.optimizer(self.model.parameters(),
lr=start_learning_rate,
**kwargs)
train_data = self.corpus.train
scheduler = ExpAnnealLR(optimizer, end_learning_rate, iterations)
model_state = self.model.state_dict()
self.model.train()
step = 0
while step < iterations:
batch_loader = DataLoader(train_data,
batch_size=mini_batch_size,
shuffle=True)
for batch in batch_loader:
step += 1
# forward pass
loss = self.model.forward_loss(batch)
# update optimizer and scheduler
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 5.0)
optimizer.step()
scheduler.step(step)
print(scheduler.get_lr())
learning_rate = scheduler.get_lr()[0]
loss_item = loss.item()
if step == 1:
best_loss = loss_item
else:
if smoothing_factor > 0:
moving_avg_loss = (smoothing_factor * moving_avg_loss +
(1 - smoothing_factor) * loss_item)
loss_item = moving_avg_loss / (1 - smoothing_factor**
(step + 1))
if loss_item < best_loss:
best_loss = loss
if step > iterations:
break
if stop_early and (loss_item > 4 * best_loss
or torch.isnan(loss)):
log_line(log)
log.info("loss diverged - stopping early!")
step = iterations
break
with open(str(learning_rate_tsv), "a") as f:
f.write(
f"{step}\t{datetime.datetime.now():%H:%M:%S}\t{learning_rate}\t{loss_item}\n"
)
self.model.load_state_dict(model_state)
self.model.to(flair.device)
log_line(log)
log.info(f"learning rate finder finished - plot {learning_rate_tsv}")
log_line(log)
return Path(learning_rate_tsv)
def train(self,
base_path: str,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
max_epochs: int = 50,
anneal_factor: float = 0.5,
patience: int = 5,
train_with_dev: bool = False,
embeddings_in_memory: bool = False,
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
eval_on_train: bool = True):
"""
Trains a text classification model using the training data of the corpus.
:param base_path: the directory to which any results should be written to
:param learning_rate: the learning rate
:param mini_batch_size: the mini batch size
:param max_epochs: the maximum number of epochs to train
:param anneal_factor: learning rate will be decreased by this factor
:param patience: number of 'bad' epochs before learning rate gets decreased
:param train_with_dev: boolean indicating, if the dev data set should be used for training or not
:param embeddings_in_memory: boolean indicating, if embeddings should be kept in memory or not
:param checkpoint: boolean indicating, whether the model should be save after every epoch or not
:param save_final_model: boolean indicating, whether the final model should be saved or not
:param anneal_with_restarts: boolean indicating, whether the best model should be reloaded once the learning
rate changed or not
:param eval_on_train: boolean value indicating, if evaluation metrics should be calculated on training data set
or not
"""
loss_txt = init_output_file(base_path, 'loss.tsv')
with open(loss_txt, 'a') as f:
f.write(
'EPOCH\tTIMESTAMP\tTRAIN_LOSS\t{}\tDEV_LOSS\t{}\tTEST_LOSS\t{}\n'
.format(Metric.tsv_header('TRAIN'), Metric.tsv_header('DEV'),
Metric.tsv_header('TEST')))
weight_extractor = WeightExtractor(base_path)
optimizer = torch.optim.SGD(self.model.parameters(), lr=learning_rate)
anneal_mode = 'min' if train_with_dev else 'max'
scheduler: ReduceLROnPlateau = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data.extend(self.corpus.dev)
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for epoch in range(max_epochs):
log.info('-' * 100)
bad_epochs = scheduler.num_bad_epochs
for group in optimizer.param_groups:
learning_rate = group['lr']
# reload last best model if annealing with restarts is enabled
if learning_rate != previous_learning_rate and anneal_with_restarts and \
os.path.exists(base_path + "/best-model.pt"):
log.info('Resetting to best model ...')
self.model.load_from_file(base_path + "/best-model.pt")
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < 0.001:
log.info('Learning rate too small - quitting training!')
break
if not self.test_mode:
random.shuffle(train_data)
self.model.train()
batches = [
self.corpus.train[x:x + mini_batch_size]
for x in range(0, len(self.corpus.train), mini_batch_size)
]
current_loss: float = 0
seen_sentences = 0
modulo = max(1, int(len(batches) / 10))
for batch_no, batch in enumerate(batches):
scores = self.model.forward(batch)
loss = self.model.calculate_loss(scores, batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_sentences += len(batch)
current_loss += loss.item()
clear_embeddings(
batch,
also_clear_word_embeddings=not embeddings_in_memory)
if batch_no % modulo == 0:
log.info(
"epoch {0} - iter {1}/{2} - loss {3:.8f}".format(
epoch + 1, batch_no, len(batches),
current_loss / seen_sentences))
iteration = epoch * len(batches) + batch_no
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
current_loss /= len(train_data)
self.model.eval()
# if checkpoint is enable, save model at each epoch
if checkpoint:
self.model.save(base_path + "/checkpoint.pt")
log.info('-' * 100)
log.info("EPOCH {0}: lr {1:.4f} - bad epochs {2}".format(
epoch + 1, learning_rate, bad_epochs))
dev_metric = train_metric = None
dev_loss = '_'
train_loss = current_loss
if eval_on_train:
train_metric, train_loss = self._calculate_evaluation_results_for(
'TRAIN', self.corpus.train, embeddings_in_memory,
mini_batch_size)
if not train_with_dev:
dev_metric, dev_loss = self._calculate_evaluation_results_for(
'DEV', self.corpus.dev, embeddings_in_memory,
mini_batch_size)
with open(loss_txt, 'a') as f:
train_metric_str = train_metric.to_tsv(
) if train_metric is not None else Metric.to_empty_tsv()
dev_metric_str = dev_metric.to_tsv(
) if dev_metric is not None else Metric.to_empty_tsv()
f.write('{}\t{:%H:%M:%S}\t{}\t{}\t{}\t{}\t{}\t{}\n'.format(
epoch, datetime.datetime.now(), train_loss,
train_metric_str, dev_loss, dev_metric_str, '_',
Metric.to_empty_tsv()))
# anneal against train loss if training with dev, otherwise anneal against dev score
scheduler.step(
current_loss) if train_with_dev else scheduler.step(
dev_metric.f_score())
current_score = dev_metric.f_score(
) if not train_with_dev else train_metric.f_score()
# if we use dev data, remember best model based on dev evaluation score
if not train_with_dev and current_score == scheduler.best:
self.model.save(base_path + "/best-model.pt")
if save_final_model:
self.model.save(base_path + "/final-model.pt")
log.info('-' * 100)
log.info('Testing using best model ...')
self.model.eval()
if os.path.exists(base_path + "/best-model.pt"):
self.model = TextClassifier.load_from_file(base_path +
"/best-model.pt")
test_metric, test_loss = self.evaluate(
self.corpus.test,
mini_batch_size=mini_batch_size,
eval_class_metrics=True,
embeddings_in_memory=embeddings_in_memory,
metric_name='TEST')
test_metric.print()
self.model.train()
log.info('-' * 100)
except KeyboardInterrupt:
log.info('-' * 100)
log.info('Exiting from training early.')
log.info('Saving model ...')
with open(base_path + "/final-model.pt", 'wb') as model_save_file:
torch.save(self.model, model_save_file, pickle_protocol=4)
model_save_file.close()
log.info('Done.')
def train(self,
base_path: str,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 2,
save_model: bool = True,
embeddings_in_memory: bool = True,
train_with_dev: bool = False):
"""
Trains the model using the training data of the corpus.
:param base_path: the directory to which any results should be written to
:param learning_rate: the learning rate
:param mini_batch_size: the mini batch size
:param max_epochs: the maximum number of epochs to train
:param save_model: boolean value indicating, whether the model should be saved or not
:param embeddings_in_memory: boolean value indicating, if embeddings should be kept in memory or not
:param train_with_dev: boolean value indicating, if the dev data set should be used for training or not
"""
loss_txt = init_output_file(base_path, 'loss.txt')
with open(loss_txt, 'a') as f:
f.write(
'EPOCH\tITERATION\tDEV_LOSS\tTRAIN_LOSS\tDEV_F_SCORE\tTRAIN_F_SCORE\tDEV_ACC\tTRAIN_ACC\n'
)
weights_txt = init_output_file(base_path, 'weights.txt')
weights_index = defaultdict(lambda: defaultdict(lambda: list()))
optimizer = torch.optim.SGD(self.model.parameters(), lr=learning_rate)
anneal_mode = 'min' if train_with_dev else 'max'
scheduler: ReduceLROnPlateau = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data.extend(self.corpus.dev)
# At any point you can hit Ctrl + C to break out of training early.
try:
# record overall best dev scores and best loss
best_score = 0
for epoch in range(max_epochs):
print('-' * 100)
if not self.test_mode:
random.shuffle(train_data)
batches = [
train_data[x:x + mini_batch_size]
for x in range(0, len(train_data), mini_batch_size)
]
current_loss: float = 0
seen_sentences = 0
modulo = max(1, int(len(batches) / 10))
self.model.train()
for batch_no, batch in enumerate(batches):
scores = self.model.forward(batch)
loss = self.model.calculate_loss(scores, batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_sentences += len(batch)
current_loss += loss.item()
if not embeddings_in_memory:
clear_embeddings(batch)
if batch_no % modulo == 0:
print("epoch {0} - iter {1}/{2} - loss {3:.8f}".format(
epoch + 1, batch_no, len(batches),
current_loss / seen_sentences))
iteration = epoch * len(batches) + batch_no
self._extract_weigths(iteration, weights_index,
weights_txt)
current_loss /= len(train_data)
# IMPORTANT: Switch to eval mode
self.model.eval()
print('-' * 100)
train_metrics, train_loss = self.evaluate(
self.corpus.train,
mini_batch_size=mini_batch_size,
embeddings_in_memory=embeddings_in_memory)
train_f_score = train_metrics['MICRO_AVG'].f_score()
train_acc = train_metrics['MICRO_AVG'].accuracy()
print(
"{0:<7} epoch {1} - loss {2:.8f} - f-score {3:.4f} - acc {4:.4f}"
.format('TRAIN:', epoch, train_loss, train_f_score,
train_acc))
dev_f_score = dev_acc = dev_loss = 0
if not train_with_dev:
dev_metrics, dev_loss = self.evaluate(
self.corpus.dev,
mini_batch_size=mini_batch_size,
embeddings_in_memory=embeddings_in_memory)
dev_f_score = dev_metrics['MICRO_AVG'].f_score()
dev_acc = dev_metrics['MICRO_AVG'].accuracy()
print(
"{0:<7} epoch {1} - loss {2:.8f} - f-score {3:.4f} - acc {4:.4f}"
.format('DEV:', epoch, dev_loss, dev_f_score, dev_acc))
with open(loss_txt, 'a') as f:
f.write('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\n'.format(
epoch, epoch * len(batches), dev_loss, train_loss,
dev_f_score, train_f_score, dev_acc, train_acc))
# IMPORTANT: Switch back to train mode
self.model.train()
# anneal against train loss if training with dev, otherwise anneal against dev score
scheduler.step(
current_loss) if train_with_dev else scheduler.step(
dev_f_score)
is_best_model_so_far: bool = False
current_score = dev_f_score if not train_with_dev else train_f_score
if current_score > best_score:
best_score = current_score
is_best_model_so_far = True
if is_best_model_so_far:
if save_model:
self.model.save(base_path + "/model.pt")
self.model.save(base_path + "/final-model.pt")
if save_model:
self.model = TextClassifier.load_from_file(base_path +
"/model.pt")
print('-' * 100)
print('testing...')
test_metrics, test_loss = self.evaluate(
self.corpus.test,
mini_batch_size=mini_batch_size,
eval_class_metrics=True,
embeddings_in_memory=embeddings_in_memory)
for metric in test_metrics.values():
metric.print()
print('-' * 100)
except KeyboardInterrupt:
print('-' * 89)
print('Exiting from training early')
print('saving model')
with open(base_path + "/final-model.pt", 'wb') as model_save_file:
torch.save(self.model, model_save_file, pickle_protocol=4)
model_save_file.close()
print('done')
def train(
self,
base_path: Union[Path, str],
evaluation_metric: EvaluationMetric = EvaluationMetric.MICRO_F1_SCORE,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
eval_mini_batch_size: int = None,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 3,
train_with_dev: bool = False,
monitor_train: bool = False,
embeddings_in_memory: bool = True,
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
num_workers: int = 8,
**kwargs,
) -> dict:
if eval_mini_batch_size is None:
eval_mini_batch_size = mini_batch_size
log.info(f'Model training base path: "{base_path}"')
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
add_file_handler(log, base_path / "training.log")
log_line(log)
log.info(f"Evaluation method: {evaluation_metric.name}")
# determine what splits (train, dev, test) to evaluate and log
log_train = True if monitor_train else False
log_test = True if (not param_selection_mode
and self.corpus.test) else False
log_dev = True if not train_with_dev else False
loss_txt = init_output_file(base_path, "loss.tsv")
with open(loss_txt, "a") as f:
f.write(f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS")
dummy_result, _ = self.model.evaluate(
[Sentence("d", labels=["0.1"])],
eval_mini_batch_size,
embeddings_in_memory,
)
if log_train:
f.write("\tTRAIN_" +
"\tTRAIN_".join(dummy_result.log_header.split("\t")))
if log_dev:
f.write("\tDEV_LOSS\tDEV_" +
"\tDEV_".join(dummy_result.log_header.split("\t")))
if log_test:
f.write("\tTEST_LOSS\tTEST_" +
"\tTEST_".join(dummy_result.log_header.split("\t")))
weight_extractor = WeightExtractor(base_path)
optimizer = self.optimizer(self.model.parameters(),
lr=learning_rate,
**kwargs)
if self.optimizer_state is not None:
optimizer.load_state_dict(self.optimizer_state)
# minimize training loss if training with dev data, else maximize dev score
anneal_mode = "min" if train_with_dev else "max"
if isinstance(optimizer, (AdamW, SGDW)):
scheduler = ReduceLRWDOnPlateau(
optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True,
)
else:
scheduler = ReduceLROnPlateau(
optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True,
)
if self.scheduler_state is not None:
scheduler.load_state_dict(self.scheduler_state)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data = ConcatDataset([self.corpus.train, self.corpus.dev])
dev_score_history = []
dev_loss_history = []
train_loss_history = []
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for epoch in range(0 + self.epoch, max_epochs + self.epoch):
log_line(log)
try:
bad_epochs = scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
learning_rate = group["lr"]
# reload last best model if annealing with restarts is enabled
if (learning_rate != previous_learning_rate
and anneal_with_restarts
and (base_path / "best-model.pt").exists()):
log.info("resetting to best model")
self.model.load(base_path / "best-model.pt")
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < 0.0001:
log_line(log)
log.info("learning rate too small - quitting training!")
log_line(log)
break
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
)
self.model.train()
train_loss: float = 0
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int(total_number_of_batches / 10))
for batch_no, batch in enumerate(batch_loader):
loss = self.model.forward_loss(batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_batches += 1
train_loss += loss.item()
clear_embeddings(
batch,
also_clear_word_embeddings=not embeddings_in_memory)
if batch_no % modulo == 0:
log.info(
f"epoch {epoch + 1} - iter {batch_no}/{total_number_of_batches} - loss "
f"{train_loss / seen_batches:.8f}")
iteration = epoch * total_number_of_batches + batch_no
if not param_selection_mode:
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
train_loss /= seen_batches
self.model.eval()
log_line(log)
log.info(
f"EPOCH {epoch + 1} done: loss {train_loss:.4f} - lr {learning_rate:.4f} - bad epochs {bad_epochs}"
)
# anneal against train loss if training with dev, otherwise anneal against dev score
current_score = train_loss
with open(loss_txt, "a") as f:
f.write(
f"\n{epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t{train_loss}"
)
if log_train:
train_eval_result, train_loss = self.model.evaluate(
self.corpus.train,
eval_mini_batch_size,
embeddings_in_memory,
num_workers=num_workers,
)
f.write(f"\t{train_eval_result.log_line}")
if log_dev:
dev_eval_result, dev_loss = self.model.evaluate(
self.corpus.dev,
eval_mini_batch_size,
embeddings_in_memory,
num_workers=num_workers,
)
f.write(f"\t{dev_loss}\t{dev_eval_result.log_line}")
log.info(
f"DEV : loss {dev_loss} - score {dev_eval_result.main_score}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_score_history.append(dev_eval_result.main_score)
dev_loss_history.append(dev_loss)
current_score = dev_eval_result.main_score
if log_test:
test_eval_result, test_loss = self.model.evaluate(
self.corpus.test,
eval_mini_batch_size,
embeddings_in_memory,
base_path / "test.tsv",
num_workers=num_workers,
)
f.write(f"\t{test_loss}\t{test_eval_result.log_line}")
log.info(
f"TEST : loss {test_loss} - score {test_eval_result.main_score}"
)
scheduler.step(current_score)
train_loss_history.append(train_loss)
# if checkpoint is enable, save model at each epoch
if checkpoint and not param_selection_mode:
self.model.save_checkpoint(
base_path / "checkpoint.pt",
optimizer.state_dict(),
scheduler.state_dict(),
epoch + 1,
train_loss,
)
# if we use dev data, remember best model based on dev evaluation score
if (not train_with_dev and not param_selection_mode
and current_score == scheduler.best):
self.model.save(base_path / "best-model.pt")
# if we do not use dev data for model selection, save final model
if save_final_model and not param_selection_mode:
self.model.save(base_path / "final-model.pt")
except KeyboardInterrupt:
log_line(log)
log.info("Exiting from training early.")
if not param_selection_mode:
log.info("Saving model ...")
self.model.save(base_path / "final-model.pt")
log.info("Done.")
# test best model if test data is present
if self.corpus.test:
final_score = self.final_test(
base_path,
embeddings_in_memory,
evaluation_metric,
eval_mini_batch_size,
num_workers,
)
else:
final_score = 0
log.info("Test data not provided setting final score to 0")
return {
"test_score": final_score,
"dev_score_history": dev_score_history,
"train_loss_history": train_loss_history,
"dev_loss_history": dev_loss_history,
}
def train(
self,
base_path: Union[Path, str],
learning_rate: float = 0.1,
mini_batch_size: int = 32,
mini_batch_chunk_size: int = None,
max_epochs: int = 100,
scheduler=AnnealOnPlateau,
anneal_factor: float = 0.5,
patience: int = 3,
initial_extra_patience=0,
min_learning_rate: float = 0.0001,
train_with_dev: bool = False,
monitor_train: bool = False,
monitor_test: bool = False,
embeddings_storage_mode: str = "cpu",
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
anneal_with_prestarts: bool = False,
batch_growth_annealing: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
num_workers: int = 6,
sampler=None,
use_amp: bool = False,
amp_opt_level: str = "O1",
eval_on_train_fraction=0.,
eval_on_train_shuffle=False,
valid_with_misspellings: bool = True,
corpus_name: str = "",
**kwargs,
) -> dict:
"""
Trains any class that implements the flair.nn.Model interface.
:param base_path: Main path to which all output during training is logged and models are saved
:param learning_rate: Initial learning rate
:param mini_batch_size: Size of mini-batches during training
:param mini_batch_chunk_size: If mini-batches are larger than this number, they get broken down into chunks of this size for processing purposes
:param max_epochs: Maximum number of epochs to train. Terminates training if this number is surpassed.
:param anneal_factor: The factor by which the learning rate is annealed
:param patience: Patience is the number of epochs with no improvement the Trainer waits until annealing the learning rate
:param min_learning_rate: If the learning rate falls below this threshold, training terminates
:param train_with_dev: If True, training is performed using both train+dev data
:param monitor_train: If True, training data is evaluated at end of each epoch
:param monitor_test: If True, test data is evaluated at end of each epoch
:param embeddings_storage_mode: One of 'none' (all embeddings are deleted and freshly recomputed),
'cpu' (embeddings are stored on CPU) or 'gpu' (embeddings are stored on GPU)
:param checkpoint: If True, a full checkpoint is saved at end of each epoch
:param save_final_model: If True, final model is saved
:param anneal_with_restarts: If True, the last best model is restored when annealing the learning rate
:param shuffle: If True, data is shuffled during training
:param param_selection_mode: If True, testing is performed against dev data. Use this mode when doing parameter selection.
:param num_workers: Number of workers in your data loader.
:param sampler: You can pass a data sampler here for special sampling of data.
:param eval_on_train_fraction: the fraction of train data to do the evaluation on,
if 0. the evaluation is not performed on fraction of training data,
if 'dev' the size is determined from dev set size
:param eval_on_train_shuffle: if True the train data fraction is determined on the start of training
and kept fixed during training, otherwise it's sampled at beginning of each epoch
:param valid_with_misspellings: use a combination of the original loss and the loss computed using the misspelled sentences for validation
:param kwargs: Other arguments for the Optimizer
:return:
"""
if self.use_tensorboard:
try:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter()
except:
log_line(log)
log.warning(
"ATTENTION! PyTorch >= 1.1.0 and pillow are required for TensorBoard support!"
)
log_line(log)
self.use_tensorboard = False
pass
if use_amp:
if sys.version_info < (3, 0):
raise RuntimeError(
"Apex currently only supports Python 3. Aborting.")
if amp is None:
raise RuntimeError(
"Failed to import apex. Please install apex from https://www.github.com/nvidia/apex "
"to enable mixed-precision training.")
if mini_batch_chunk_size is None:
mini_batch_chunk_size = mini_batch_size
if learning_rate < min_learning_rate:
min_learning_rate = learning_rate / 10
initial_learning_rate = learning_rate
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
log_handler = add_file_handler(log, base_path / "training.log")
log_line(log)
log.info(f'Model: "{self.model}"')
log_line(log)
log.info(f'Corpus: "{self.corpus}"')
log_line(log)
log.info("Parameters:")
log.info(f' - learning_rate: "{learning_rate}"')
log.info(f' - mini_batch_size: "{mini_batch_size}"')
log.info(f' - patience: "{patience}"')
log.info(f' - anneal_factor: "{anneal_factor}"')
log.info(f' - max_epochs: "{max_epochs}"')
log.info(f' - shuffle: "{shuffle}"')
log.info(f' - train_with_dev: "{train_with_dev}"')
log.info(f' - batch_growth_annealing: "{batch_growth_annealing}"')
log.info(f' - mixed precision training: "{use_amp}"')
log.info(f' - valid_with_misspellings: "{valid_with_misspellings}"')
log.info("Model:")
log.info(f' - hidden_size: "{self.model.hidden_size}"')
log.info(f' - train_mode: "{self.model.train_mode}"')
log.info(f' - misspell_mode: "{self.model.misspell_mode}"')
log.info(f' - alpha: "{self.model.alpha}"')
log.info(f' - beta: "{self.model.beta}"')
if self.model.misspell_mode == MisspellingMode.Seq2Seq:
log.info(f' - errgen_model: "{self.model.errgen_model_train}"')
log.info(f' - errgen_mode: "{self.model.errgen_mode_train}"')
from pysia.utils import is_generation_mode, is_correction_mode
if is_generation_mode(self.model.errgen_mode_train):
log.info(f' - errgen_temp: "{self.model.errgen_temp_train}"')
log.info(f' - errgen_topk: "{self.model.errgen_topk_train}"')
elif is_correction_mode(self.model.errgen_mode_train):
log.info(f' - errgen_nbest: "{self.model.errgen_nbest_train}"')
log.info(
f' - errgen_beam_size: "{self.model.errgen_beam_size_train}"'
)
elif self.model.misspell_mode in [MisspellingMode.Random]:
log.info(
f' - misspelling_rate: "{self.model.misspelling_rate_train}"')
elif self.model.misspell_mode in [
MisspellingMode.ConfusionMatrixBased
]:
log.info(f' - cmx_file: "{self.model.cmx_file_train}"')
elif self.model.misspell_mode in [MisspellingMode.Typos]:
log.info(f' - typos_file: "{self.model.typos_file_train}"')
log.info(
f' - misspelling_rate: "{self.model.misspelling_rate_train}"')
log_line(log)
log.info(f'Model training base path: "{base_path}"')
log_line(log)
log.info(f"Device: {flair.device}")
log_line(log)
log.info(f"Embeddings storage mode: {embeddings_storage_mode}")
# determine what splits (train, dev, test) to evaluate and log
log_train = True if monitor_train else False
log_test = (True if (not param_selection_mode and self.corpus.test
and monitor_test) else False)
log_dev = True if not train_with_dev else False
log_train_part = (True if (eval_on_train_fraction == "dev"
or eval_on_train_fraction > 0.0) else False)
if log_train_part:
train_part_size = (len(
self.corpus.dev) if eval_on_train_fraction == "dev" else int(
len(self.corpus.train) * eval_on_train_fraction))
assert train_part_size > 0
if not eval_on_train_shuffle:
train_part_indices = list(range(train_part_size))
train_part = torch.utils.data.dataset.Subset(
self.corpus.train, train_part_indices)
log_test = not log_dev
eval_misspelling_rate = 0.05
eval_misspelling_mode = MisspellingMode.Random
log_suffix = lambda prefix, rate, cm, mode: f"{prefix} (misspell: cmx={cm})" if mode == MisspellingMode.ConfusionMatrixBased else f"{prefix} (misspell: rate={rate})"
# prepare loss logging file and set up header
loss_txt = init_output_file(base_path, "loss.tsv")
weight_extractor = WeightExtractor(base_path)
optimizer: torch.optim.Optimizer = self.optimizer(
self.model.parameters(), lr=learning_rate, **kwargs)
if use_amp:
self.model, optimizer = amp.initialize(self.model,
optimizer,
opt_level=amp_opt_level)
# minimize training loss if training with dev data, else maximize dev score
anneal_mode = "min" if train_with_dev else "max"
lr_scheduler = scheduler(
optimizer,
factor=anneal_factor,
patience=patience,
initial_extra_patience=initial_extra_patience,
mode=anneal_mode,
verbose=True,
)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data = ConcatDataset([self.corpus.train, self.corpus.dev])
# initialize sampler if provided
if sampler is not None:
# init with default values if only class is provided
if inspect.isclass(sampler):
sampler = sampler()
# set dataset to sample from
sampler.set_dataset(train_data)
shuffle = False
dev_clean_score_history = []
dev_noisy_score_history = []
dev_clean_loss_history = []
dev_noisy_loss_history = []
train_loss_history = []
micro_batch_size = mini_batch_chunk_size
complete_data = ConcatDataset(
[self.corpus.train, self.corpus.dev, self.corpus.test])
char_vocab = make_char_vocab(complete_data)
log.info(
f"Vocabulary of the corpus (#{len(char_vocab)}): {char_vocab}")
cmx, lut, typos = None, {}, {}
if self.model.misspell_mode == MisspellingMode.ConfusionMatrixBased:
cmx, lut = load_confusion_matrix(self.model.cmx_file_train)
cmx, lut = filter_cmx(cmx, lut, char_vocab)
elif self.model.misspell_mode == MisspellingMode.Typos:
typos = load_typos(self.model.typos_file_train, char_vocab, False)
if self.model.misspell_mode == MisspellingMode.Seq2Seq:
translator, opt = init_translator(
self.model.errgen_model_train,
self.model.errgen_mode_train,
log,
temp=self.model.errgen_temp_train,
topk=self.model.errgen_topk_train,
nbest=self.model.errgen_nbest_train,
beam_size=self.model.errgen_beam_size_train,
shard_size=20000,
batch_size=256,
verbose=True)
else:
translator, opt = None, None
loss_params = {}
loss_params["verbose"] = False
loss_params["char_vocab"] = char_vocab
loss_params["cmx"] = cmx
loss_params["lut"] = lut
loss_params["typos"] = typos
loss_params["translator"] = translator
loss_params["opt"] = opt
loss_params["translation_mode"] = self.model.errgen_mode_train
loss_params["embeddings_storage_mode"] = embeddings_storage_mode
if self.model.train_mode == TrainingMode.Combined and self.model.beta > 0.0:
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
sampler=sampler,
)
sum_sent_len, cnt_sent = 0, 0
for batch_no, batch in enumerate(batch_loader):
for sent in batch:
sum_sent_len += len(sent)
cnt_sent += len(batch)
mean_tokens_per_batch = float(sum_sent_len) / float(cnt_sent)
loss_params["mean_tokens_per_batch"] = mean_tokens_per_batch
log.info(f"mean_tokens_per_batch = {mean_tokens_per_batch:.4f}")
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for self.epoch in range(self.epoch + 1, max_epochs + 1):
log_line(log)
if anneal_with_prestarts:
last_epoch_model_state_dict = copy.deepcopy(
self.model.state_dict())
if eval_on_train_shuffle:
train_part_indices = list(range(self.corpus.train))
random.shuffle(train_part_indices)
train_part_indices = train_part_indices[:train_part_size]
train_part = torch.utils.data.dataset.Subset(
self.corpus.train, train_part_indices)
# get new learning rate
for group in optimizer.param_groups:
learning_rate = group["lr"]
if learning_rate != previous_learning_rate and batch_growth_annealing:
mini_batch_size *= 2
# reload last best model if annealing with restarts is enabled
if ((anneal_with_restarts or anneal_with_prestarts)
and learning_rate != previous_learning_rate
and (base_path / "best-model.pt").exists()):
if anneal_with_restarts:
log.info("resetting to best model")
self.model.load_state_dict(
self.model.load(base_path /
"best-model.pt").state_dict())
if anneal_with_prestarts:
log.info("resetting to pre-best model")
self.model.load_state_dict(
self.model.load(base_path /
"pre-best-model.pt").state_dict())
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < min_learning_rate:
log_line(log)
log.info("learning rate too small - quitting training!")
log_line(log)
break
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
sampler=sampler,
)
self.model.train()
train_loss: float = 0
train_auxilary_losses = {}
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int(total_number_of_batches / 10))
# process mini-batches
batch_time = 0
for batch_no, batch in enumerate(batch_loader):
start_time = time.time()
# zero the gradients on the model and optimizer
self.model.zero_grad()
optimizer.zero_grad()
# if necessary, make batch_steps
batch_steps = [batch]
if len(batch) > micro_batch_size:
batch_steps = [
batch[x:x + micro_batch_size]
for x in range(0, len(batch), micro_batch_size)
]
# forward and backward for batch
for batch_step in batch_steps:
# forward pass
loss, auxilary_losses = self.model.forward_loss(
batch_step, params=loss_params)
# Backward
if use_amp:
with amp.scale_loss(loss,
optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# do the optimizer step
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_batches += 1
train_loss += loss.item()
for k, v in auxilary_losses.items():
train_auxilary_losses[k] = train_auxilary_losses.get(
k, 0) + v
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(batch, embeddings_storage_mode)
batch_time += time.time() - start_time
if seen_batches % modulo == 0:
msg = f"epoch {self.epoch} - iter {seen_batches}/{total_number_of_batches} - loss {train_loss / seen_batches:.6f} - samples/sec: {mini_batch_size * modulo / batch_time:.2f}"
# note: this is the loss accumulated in the current epoch divided by the number of already seen batches
if len(train_auxilary_losses) > 0:
accuracies = [
(key, value)
for (key,
value) in train_auxilary_losses.items()
if key.startswith("acc_")
]
counts = [
(key, value)
for (key,
value) in train_auxilary_losses.items()
if key.startswith("cnt_")
or key.startswith("sum_")
]
losses = [
(key, value)
for (key,
value) in train_auxilary_losses.items()
if key.startswith("loss_")
]
aux_losses_str = ""
if len(losses) > 0:
aux_losses_str = " ".join([
f"{key}={value / seen_batches:.6f}"
for (key, value) in losses
])
if len(accuracies) > 0:
if len(aux_losses_str) > 0:
aux_losses_str += " "
aux_losses_str += " ".join([
f"{key}={value / seen_batches:.2f}%"
for (key, value) in accuracies
])
if len(counts) > 0:
if len(aux_losses_str) > 0:
aux_losses_str += " "
aux_losses_str += " ".join([
f"{key}={value / seen_batches:.2f}"
for (key, value) in counts
])
msg += f" ({aux_losses_str})"
log.info(msg)
batch_time = 0
iteration = self.epoch * total_number_of_batches + batch_no
if not param_selection_mode:
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
train_loss /= seen_batches
for k, v in auxilary_losses.items():
train_auxilary_losses[k] /= seen_batches
self.model.eval()
log_line(log)
log.info(
f"EPOCH {self.epoch} done: loss {train_loss:.4f} - lr {learning_rate:.4f}"
)
if self.use_tensorboard:
writer.add_scalar("train_loss", train_loss, self.epoch)
# anneal against train loss if training with dev, otherwise anneal against dev score
current_score = train_loss
# evaluate on train / dev / test split depending on training settings
result_line: str = ""
if log_train:
train_eval_result, train_loss = self.model.evaluate(
DataLoader(
self.corpus.train,
batch_size=mini_batch_chunk_size,
num_workers=num_workers,
),
embeddings_storage_mode=embeddings_storage_mode,
eval_dict_name=corpus_name,
)
result_line += f"\t{train_eval_result.log_line}"
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.train,
embeddings_storage_mode)
if log_train_part:
train_part_eval_result, train_part_loss = self.model.evaluate(
DataLoader(
train_part,
batch_size=mini_batch_chunk_size,
num_workers=num_workers,
),
embeddings_storage_mode=embeddings_storage_mode,
eval_dict_name=corpus_name,
)
result_line += (
f"\t{train_part_loss}\t{train_part_eval_result.log_line}"
)
log.info(
f"TRAIN_SPLIT : loss {train_part_loss} - score {round(train_part_eval_result.main_score, 4)}"
)
if log_dev:
dev_eval_result_clean, dev_loss_clean = self.model.evaluate(
DataLoader(
self.corpus.dev,
batch_size=mini_batch_chunk_size,
num_workers=num_workers,
),
embeddings_storage_mode=embeddings_storage_mode,
eval_dict_name=corpus_name,
)
result_line += f"\t{dev_loss_clean}\t{dev_eval_result_clean.log_line}"
log.info(
f"DEV : loss {dev_loss_clean} - score {round(dev_eval_result_clean.main_score, 4)}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_clean_score_history.append(
dev_eval_result_clean.main_score)
dev_clean_loss_history.append(dev_loss_clean.item())
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.dev, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("dev_clean_loss", dev_loss_clean,
self.epoch)
writer.add_scalar("dev_clean_score",
dev_eval_result_clean.main_score,
self.epoch)
# evaluate on misspellings
if valid_with_misspellings:
dev_eval_result_noisy, dev_loss_noisy = self.model.evaluate(
DataLoader(
self.corpus.dev,
batch_size=mini_batch_chunk_size,
num_workers=num_workers,
),
embeddings_storage_mode=embeddings_storage_mode,
eval_mode=EvalMode.Misspellings,
misspell_mode=eval_misspelling_mode,
char_vocab=char_vocab,
cmx=cmx,
lut=lut,
typos=typos,
misspelling_rate=eval_misspelling_rate,
eval_dict_name=corpus_name,
)
result_line += f"\t{dev_loss_noisy}\t{dev_eval_result_noisy.log_line}"
log.info(
f"{log_suffix('DEV', eval_misspelling_rate, '', eval_misspelling_mode)}"
+
f" : loss {dev_loss_noisy} - score {round(dev_eval_result_noisy.main_score, 4)}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_noisy_score_history.append(dev_eval_result_noisy)
dev_noisy_loss_history.append(dev_loss_noisy.item())
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.dev,
embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("dev_noisy_loss", dev_loss_noisy,
self.epoch)
writer.add_scalar("dev_noisy_score",
dev_eval_result_noisy.main_score,
self.epoch)
if valid_with_misspellings:
current_score = (
dev_eval_result_clean.main_score +
dev_eval_result_noisy.main_score) / 2.0
dev_loss = (dev_loss_clean + dev_loss_noisy) / 2.0
else:
current_score = dev_eval_result_clean.main_score
dev_loss = dev_loss_clean
# else: current_score = train_loss
if log_test:
test_eval_result_clean, test_loss_clean = self.model.evaluate(
DataLoader(
self.corpus.test,
batch_size=mini_batch_chunk_size,
num_workers=num_workers,
),
base_path / "test.tsv",
embeddings_storage_mode=embeddings_storage_mode,
eval_dict_name=corpus_name,
)
result_line += f"\t{test_loss_clean}\t{test_eval_result_clean.log_line}"
log.info(
f"TEST : loss {test_loss_clean} - score {round(test_eval_result_clean.main_score, 4)}"
)
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.test, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("test_clean_loss", test_loss_clean,
self.epoch)
writer.add_scalar("test_clean_score",
test_eval_result_clean.main_score,
self.epoch)
if valid_with_misspellings:
# evaluate on misspellings
test_eval_result_noisy, test_loss_noisy = self.model.evaluate(
DataLoader(
self.corpus.test,
batch_size=mini_batch_chunk_size,
num_workers=num_workers,
),
base_path / f"test.tsv",
embeddings_storage_mode=embeddings_storage_mode,
eval_mode=EvalMode.Misspellings,
misspell_mode=eval_misspelling_mode,
char_vocab=char_vocab,
cmx=cmx,
lut=lut,
typos=typos,
misspelling_rate=eval_misspelling_rate,
eval_dict_name=corpus_name,
)
result_line += f"\t{test_loss_noisy}\t{test_eval_result_noisy.log_line}"
log.info(
f"{log_suffix('TEST', eval_misspelling_rate, '', eval_misspelling_mode)}"
+
f" : loss {test_loss_noisy} - score {round(test_eval_result_noisy.main_score, 4)}"
)
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.test,
embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("test_noisy_loss",
test_loss_noisy, self.epoch)
writer.add_scalar(
"test_noisy_score",
test_eval_result_noisy.main_score, self.epoch)
# determine learning rate annealing through scheduler. Use auxiliary metric for AnnealOnPlateau
if not train_with_dev and isinstance(lr_scheduler,
AnnealOnPlateau):
lr_scheduler.step(current_score, dev_loss)
else:
lr_scheduler.step(current_score)
train_loss_history.append(train_loss)
# determine bad epoch number
try:
bad_epochs = lr_scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
new_learning_rate = group["lr"]
if new_learning_rate != previous_learning_rate:
bad_epochs = patience + 1
if previous_learning_rate == initial_learning_rate:
bad_epochs += initial_extra_patience
# log bad epochs
log.info(f"BAD EPOCHS (no improvement): {bad_epochs}")
# output log file
with open(loss_txt, "a") as f:
# make headers on first epoch
if self.epoch == 1:
f.write(
f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS"
)
if log_train:
f.write("\tTRAIN_" + "\tTRAIN_".join(
train_eval_result.log_header.split("\t")))
if log_train_part:
f.write("\tTRAIN_PART_LOSS\tTRAIN_PART_" +
"\tTRAIN_PART_".join(
train_part_eval_result.log_header.
split("\t")))
if log_dev:
f.write("\tDEV_LOSS\tDEV_" + "\tDEV_".join(
dev_eval_result_clean.log_header.split("\t")))
if log_test:
f.write("\tTEST_LOSS\tTEST_" + "\tTEST_".join(
test_eval_result_clean.log_header.split("\t")))
f.write(
f"\n{self.epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t{train_loss}"
)
f.write(result_line)
# if checkpoint is enabled, save model at each epoch
if checkpoint and not param_selection_mode:
self.save_checkpoint(base_path / "checkpoint.pt")
# if we use dev data, remember best model based on dev evaluation score
if ((not train_with_dev or anneal_with_restarts
or anneal_with_prestarts) and not param_selection_mode
and current_score == lr_scheduler.best
and bad_epochs == 0):
log.info("saving best model")
self.model.save(base_path / "best-model.pt")
if anneal_with_prestarts:
current_state_dict = self.model.state_dict()
self.model.load_state_dict(last_epoch_model_state_dict)
self.model.save(base_path / "pre-best-model.pt")
self.model.load_state_dict(current_state_dict)
# if we do not use dev data for model selection, save final model
if save_final_model and not param_selection_mode:
self.model.save(base_path / "final-model.pt")
except KeyboardInterrupt:
log_line(log)
log.info("Exiting from training early.")
if self.use_tensorboard:
writer.close()
if not param_selection_mode:
log.info("Saving model ...")
self.model.save(base_path / "final-model.pt")
log.info("Done.")
# test best model if test data is present
if self.corpus.test:
final_score_clean = self.final_test(base_path,
mini_batch_chunk_size,
num_workers,
embeddings_storage_mode,
corpus_name=corpus_name)
final_score_noisy = self.final_test(
base_path,
mini_batch_chunk_size,
num_workers,
embeddings_storage_mode,
eval_mode=EvalMode.Misspellings,
misspell_mode=eval_misspelling_mode,
misspelling_rate=eval_misspelling_rate,
char_vocab=char_vocab,
cmx=cmx,
lut=lut,
typos=typos,
corpus_name=corpus_name)
else:
final_score_clean, final_score_noisy = 0, 0
log.info("Test data not provided setting final score to 0")
log.removeHandler(log_handler)
if self.use_tensorboard:
writer.close()
return {
"test_score_clean": final_score_clean,
"test_score_noisy": final_score_noisy,
"dev_clean_score_history": dev_clean_score_history,
"dev_noisy_score_history": dev_noisy_score_history,
"train_loss_history": train_loss_history,
"dev_clean_loss_history": dev_clean_loss_history,
"dev_noisy_loss_history": dev_noisy_loss_history,
}
def train(
self,
base_path: Union[Path, str],
learning_rate: float = 0.1,
mini_batch_size: int = 32,
mini_batch_chunk_size: int = None,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 3,
min_learning_rate: float = 0.0001,
train_with_dev: bool = False,
monitor_train: bool = False,
monitor_test: bool = False,
embeddings_storage_mode: str = "cpu",
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
batch_growth_annealing: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
num_workers: int = 6,
sampler=None,
use_amp: bool = False,
amp_opt_level: str = "O1",
eval_on_train_fraction=0.0,
eval_on_train_shuffle=False,
gamma: float = 1.0,
**kwargs,
) -> dict:
if mini_batch_chunk_size is None:
mini_batch_chunk_size = mini_batch_size
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
log_line(log)
log.info(f'Model1: "{self.model1}"')
log.info(f'Model2: "{self.model2}"')
log_line(log)
log.info(f'Corpus: "{self.corpus}"')
log_line(log)
log.info(f'Model training base path: "{base_path}"')
log_line(log)
log.info(f"Device: {flair.device}")
log_line(log)
log.info(f"Embeddings storage mode: {embeddings_storage_mode}")
loss_txt = init_output_file(base_path, "loss.tsv")
learning_rate1 = learning_rate2 = learning_rate
optimizer1: torch.optim.Optimizer = self.optimizer(
self.model1.parameters(), lr=learning_rate1, **kwargs
)
optimizer2: torch.optim.Optimizer = self.optimizer(
self.model2.parameters(), lr=learning_rate2, **kwargs
)
anneal_mode = "min" if train_with_dev else "max"
scheduler1: ReduceLROnPlateau = ReduceLROnPlateau(
optimizer1,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True,
)
scheduler2: ReduceLROnPlateau = ReduceLROnPlateau(
optimizer2,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True,
)
train_data = self.corpus.train
dev_score_history = []
dev_loss_history = []
train_loss_history = []
try:
previous_learning_rate1 = learning_rate1
# previous_learning_rate2 = learning_rate2
for self.epoch in range(self.epoch + 1, max_epochs + 1):
log_line(log)
# get new learning rate
for group in optimizer1.param_groups:
learning_rate1 = group["lr"]
for group in optimizer2.param_groups:
learning_rate2 = group["lr"]
previous_learning_rate1 = learning_rate1
# previous_learning_rate2 = learning_rate2
# stop training if learning rate becomes too small
if learning_rate1 < min_learning_rate:
log_line(log)
log.info("learning rate (1) too small - quitting training!")
log_line(log)
break
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
sampler=sampler,
)
self.model1.train()
self.model2.train()
train_loss: float = 0
train_loss1: float = 0
train_loss2: float = 0
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int(total_number_of_batches / 10))
# process mini-batches
batch_time = 0
for batch_no, batch in enumerate(batch_loader):
start_time = time.time()
# zero the gradients on the model and optimizer
self.model1.zero_grad()
self.model2.zero_grad()
optimizer1.zero_grad()
optimizer2.zero_grad()
loss1 = self.model1.forward_loss(batch)
loss2 = self.model2.forward_loss(batch, self.model1.encoder_final)
loss = loss1 + gamma * loss2
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model1.parameters(), 5.0)
torch.nn.utils.clip_grad_norm_(self.model2.parameters(), 5.0)
optimizer1.step()
optimizer2.step()
seen_batches += 1
train_loss += loss.item()
train_loss1 += loss1.item()
train_loss2 += loss2.item()
store_embeddings(batch, embeddings_storage_mode)
batch_time += time.time() - start_time
if batch_no % modulo == 0:
log.info(
f"epoch {self.epoch} - iter {batch_no}/{total_number_of_batches} - "
f"loss {train_loss / seen_batches:.8f} - "
f"samples/sec: {mini_batch_size * modulo / batch_time:.2f}"
)
batch_time = 0
train_loss /= seen_batches
train_loss1 /= seen_batches
train_loss2 /= seen_batches
self.model1.eval()
self.model2.eval()
log_line(log)
log.info(
f"EPOCH {self.epoch} done: "
f"loss {train_loss:.5f} - "
f"loss1 {train_loss1:.5f} - "
f"loss2 {train_loss2:.5f} - "
f"gamma {gamma:.2f} - "
f"lr1 {learning_rate1:.5f} - "
f"lr2 {learning_rate2:.5f}"
)
result_line: str = ""
dev_eval_result, dev_loss = self.model1.evaluate(
DataLoader(
self.corpus.dev,
batch_size=mini_batch_chunk_size,
num_workers=num_workers,
),
embedding_storage_mode=embeddings_storage_mode,
)
result_line += f"\t{dev_loss}\t{dev_eval_result.log_line}"
log.info(f"DEV : loss {dev_loss} - score {dev_eval_result.main_score}")
# calculate scores using dev data if available
# append dev score to score history
dev_score_history.append(dev_eval_result.main_score)
dev_loss_history.append(dev_loss)
current_score = dev_eval_result.main_score
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.dev, embeddings_storage_mode)
# determine learning rate annealing through scheduler
scheduler1.step(current_score)
scheduler2.step(current_score)
train_loss_history.append(train_loss)
# determine bad epoch number
try:
bad_epochs = scheduler1.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer1.param_groups:
new_learning_rate1 = group["lr"]
if new_learning_rate1 != previous_learning_rate1:
bad_epochs = patience + 1
# log bad epochs
log.info(f"BAD EPOCHS (no improvement): {bad_epochs}")
# output log file
with open(loss_txt, "a") as f: # make headers on first epoch
if self.epoch == 1:
f.write(
f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS"
)
f.write(
"\tDEV_LOSS\tDEV_"
+ "\tDEV_".join(dev_eval_result.log_header.split("\t"))
)
f.write(
f"\n{self.epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t{train_loss}"
)
f.write(result_line)
# if we use dev data, remember best model based on dev evaluation score
if current_score == scheduler1.best:
log.info(f"BEST SO FAR: {scheduler1.best}")
self.model1.save(base_path / "best-model1.pt")
self.model2.save(base_path / "best-model2.pt")
except KeyboardInterrupt:
log_line(log)
log.info("Exiting from training early.")
return {
"dev_score_history": dev_score_history,
"train_loss_history": train_loss_history,
"dev_loss_history": dev_loss_history,
}
def train(self,
base_path: str,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 2,
save_model: bool = True,
embeddings_in_memory: bool = True,
train_with_dev: bool = False,
use_tensorboard: bool = False):
"""
Trains the model using the training data of the corpus.
:param base_path: the directory to which any results should be written to
:param learning_rate: the learning rate
:param mini_batch_size: the mini batch size
:param max_epochs: the maximum number of epochs to train
:param save_model: boolean value indicating, whether the model should be saved or not
:param embeddings_in_memory: boolean value indicating, if embeddings should be kept in memory or not
:param train_with_dev: boolean value indicating, if the dev data set should be used for training or not
"""
if use_tensorboard:
try:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter()
except:
log_line(log)
log.warning(
"ATTENTION! PyTorch >= 1.1.0 and pillow are required for TensorBoard support!"
)
log_line(log)
self.use_tensorboard = False
pass
loss_txt = init_output_file_in(base_path, 'loss.tsv')
training_log = init_output_file_in(base_path, 'training_log.txt')
with open(loss_txt, 'a') as f:
f.write(
f"EPOCH\tTIMESTAMP\tLEARNING_RATE\tTRAIN_LOSS\tDEV_LOSS\tDEV_PRECISION\tDEV_RECALL\tDEV_F1\tDEV_ACC\n"
#f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS\tDEV_LOSS\tDEV_PRECISION\tDEV_RECALL\tDEV_F1\tTRAIN_PRECISION\tTRAIN_RECALL\tTRAIN_F1\tDEV_ACC\tTRAIN_ACC\n"
)
f.close()
# with open(loss_txt, 'a') as f:
# f.write('EPOCH\tITERATION\tDEV_LOSS\tTRAIN_LOSS\tDEV_F_SCORE\tTRAIN_F_SCORE\tDEV_ACC\tTRAIN_ACC\n')
weights_txt = init_output_file(base_path, 'weights.txt')
weights_index = defaultdict(lambda: defaultdict(lambda: list()))
optimizer = torch.optim.SGD(self.model.parameters(), lr=learning_rate)
anneal_mode = 'min' if train_with_dev else 'max'
scheduler: ReduceLROnPlateau = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data.extend(self.corpus.dev)
# At any point you can hit Ctrl + C to break out of training early.
try:
# record overall best dev scores and best loss
best_score = 0
for epoch in range(max_epochs):
print('-' * 100)
if not self.test_mode:
random.shuffle(train_data)
batches = [
train_data[x:x + mini_batch_size]
for x in range(0, len(train_data), mini_batch_size)
]
current_loss: float = 0
seen_sentences = 0
modulo = max(1, int(len(batches) / 10))
self.model.train()
for batch_no, batch in enumerate(batches):
scores = self.model.forward(batch)
loss = self.model.calculate_loss(scores, batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_sentences += len(batch)
current_loss += loss.item()
if not embeddings_in_memory:
clear_embeddings(batch)
if batch_no % modulo == 0:
print("epoch {0} - iter {1}/{2} - loss {3:.8f}".format(
epoch + 1, batch_no, len(batches),
current_loss / seen_sentences))
iteration = epoch * len(batches) + batch_no
self._extract_weigths(iteration, weights_index,
weights_txt)
current_loss /= len(train_data)
# IMPORTANT: Switch to eval mode
self.model.eval()
print('-' * 100)
#train_metrics, train_loss = self.evaluate(self.corpus.train, mini_batch_size=mini_batch_size,
# embeddings_in_memory=embeddings_in_memory)
#train_f_score = train_metrics['MICRO_AVG'].f_score()
#train_acc = train_metrics['MICRO_AVG'].accuracy()
#print("{0:<7} epoch {1} - loss {2:.8f} - f-score {3:.4f} - acc {4:.4f}".format(
# 'TRAIN:', epoch, train_loss, train_f_score, train_acc))
dev_presicion = dev_recall = dev_f_score = dev_acc = dev_loss = 0
if not train_with_dev:
dev_metrics, dev_loss = self.evaluate(
self.corpus.dev,
mini_batch_size=mini_batch_size,
embeddings_in_memory=embeddings_in_memory)
dev_precision = dev_metrics['MICRO_AVG'].precision()
dev_recall = dev_metrics['MICRO_AVG'].recall()
dev_f_score = dev_metrics['MICRO_AVG'].f_score()
dev_acc = dev_metrics['MICRO_AVG'].accuracy()
print(
"{0:<7} epoch {1} - loss {2:.8f} - f-score {3:.4f} - acc {4:.4f}"
.format('DEV:', epoch, dev_loss, dev_f_score, dev_acc))
with open(loss_txt, 'a') as f:
f.write(
f"{epoch}\t{datetime.datetime.now():%H:%M:%S}\t{learning_rate:.4f}\t{current_loss}\t{dev_loss}\t{dev_presicion}\t{dev_recall}\t{dev_f_score}\t{dev_acc}\n"
)
f.close()
if use_tensorboard:
writer.add_scalar("dev_loss", dev_loss, epoch)
writer.add_scalar("dev_score", dev_f_score, epoch)
writer.add_scalar("train_loss", current_loss, epoch)
#writer.add_scalar("train_score", train_f_score, epoch)
# f.write('{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\n'.format(
# epoch, epoch * len(batches), dev_loss, train_loss, dev_f_score, train_f_score, dev_acc, train_acc))
# IMPORTANT: Switch back to train mode
self.model.train()
# anneal against train loss if training with dev, otherwise anneal against dev score
scheduler.step(
current_loss) if train_with_dev else scheduler.step(
dev_f_score)
is_best_model_so_far: bool = False
current_score = dev_f_score if not train_with_dev else train_f_score
if current_score > best_score:
best_score = current_score
is_best_model_so_far = True
if is_best_model_so_far:
if save_model:
self.model.save(base_path + "/model.pt")
self.model.save(base_path + "/final-model.pt")
if save_model:
self.model = TextClassifier.load_from_file(base_path +
"/model.pt")
print('-' * 100)
print('testing...')
test_metrics, test_loss = self.evaluate(
self.corpus.test,
mini_batch_size=mini_batch_size,
eval_class_metrics=True,
embeddings_in_memory=embeddings_in_memory)
for metric in test_metrics.values():
metric.print()
with open(training_log, 'a') as c:
for metric in test_metrics.values():
print(metric.name + "\t" + "False-Negative: " +
str(metric._fn) + "\t" + "False-Positive: " +
str(metric._fp) + "\t" + "True-Negative: " +
str(metric._tn) + "\t" + "True-Positive: " +
str(metric._tp))
c.write(metric.name + "\t" + "False-Negative: " +
str(metric._fn) + "\t" + "False-Positive: " +
str(metric._fp) + "\t" + "True-Negative: " +
str(metric._tn) + "\t" + "True-Positive: " +
str(metric._tp) + "\n")
c.close()
print('-' * 100)
if use_tensorboard:
writer.close()
except KeyboardInterrupt:
print('-' * 89)
print('Exiting from training early')
print('saving model')
with open(base_path + "/final-model.pt", 'wb') as model_save_file:
torch.save(self.model, model_save_file, pickle_protocol=4)
model_save_file.close()
print('done')
from pathlib import Path
def find_learning_rate(self,
base_path: Union[(Path, str)],
file_name: str = 'learning_rate.tsv',
start_learning_rate: float = 1e-07,
end_learning_rate: float = 10,
iterations: int = 100,
mini_batch_size: int = 32,
stop_early: bool = True,
smoothing_factor: float = 0.98,
**kwargs) -> Path:
best_loss = None
moving_avg_loss = 0
if (type(base_path) is str):
base_path = Path(base_path)
learning_rate_tsv = init_output_file(base_path, file_name)
with open(learning_rate_tsv, 'a') as f:
f.write('ITERATION\tTIMESTAMP\tLEARNING_RATE\tTRAIN_LOSS\n')
optimizer = self.optimizer(self.model.parameters(),
lr=start_learning_rate,
**kwargs)
train_data = self.corpus.train
batch_loader = DataLoader(train_data,
batch_size=mini_batch_size,
shuffle=True)
scheduler = ExpAnnealLR(optimizer, end_learning_rate, iterations)
model_state = self.model.state_dict()
model_device = next(self.model.parameters()).device
self.model.train()
for (itr, batch) in enumerate(batch_loader):
loss = self.model.forward_loss(batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 5.0)
optimizer.step()
scheduler.step(1)
learning_rate = scheduler.get_lr()[0]
loss_item = loss.item()
if (itr == 0):
best_loss = loss_item
else:
if (smoothing_factor > 0):
moving_avg_loss = ((smoothing_factor * moving_avg_loss) +
((1 - smoothing_factor) * loss_item))
loss_item = (moving_avg_loss /
(1 - (smoothing_factor**(itr + 1))))
if (loss_item < best_loss):
best_loss = loss
if (stop_early and ((loss_item >
(4 * best_loss)) or torch.isnan(loss))):
log_line(log)
log.info('loss diverged - stopping early!')
break
if (itr > iterations):
break
with open(str(learning_rate_tsv), 'a') as f:
f.write(''.join([
'{}'.format(itr), '\t',
'{:%H:%M:%S}'.format(datetime.datetime.now()), '\t',
'{}'.format(learning_rate), '\t', '{}'.format(loss_item),
'\n'
]))
self.model.load_state_dict(model_state)
self.model.to(model_device)
log_line(log)
log.info(''.join([
'learning rate finder finished - plot ',
'{}'.format(learning_rate_tsv)
]))
log_line(log)
return Path(learning_rate_tsv)
def train(self,
base_path: Union[(Path, str)],
learning_rate: float = 0.1,
mini_batch_size: int = 32,
eval_mini_batch_size: int = None,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 3,
min_learning_rate: float = 0.0001,
train_with_dev: bool = False,
monitor_train: bool = False,
monitor_test: bool = False,
embeddings_storage_mode: str = 'cpu',
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
num_workers: int = 6,
sampler=None,
use_amp: bool = False,
amp_opt_level: str = 'O1',
**kwargs) -> dict:
"\n Trains any class that implements the flair.nn.Model interface.\n :param base_path: Main path to which all output during training is logged and models are saved\n :param learning_rate: Initial learning rate\n :param mini_batch_size: Size of mini-batches during training\n :param eval_mini_batch_size: Size of mini-batches during evaluation\n :param max_epochs: Maximum number of epochs to train. Terminates training if this number is surpassed.\n :param anneal_factor: The factor by which the learning rate is annealed\n :param patience: Patience is the number of epochs with no improvement the Trainer waits\n until annealing the learning rate\n :param min_learning_rate: If the learning rate falls below this threshold, training terminates\n :param train_with_dev: If True, training is performed using both train+dev data\n :param monitor_train: If True, training data is evaluated at end of each epoch\n :param monitor_test: If True, test data is evaluated at end of each epoch\n :param embeddings_storage_mode: One of 'none' (all embeddings are deleted and freshly recomputed),\n 'cpu' (embeddings are stored on CPU) or 'gpu' (embeddings are stored on GPU)\n :param checkpoint: If True, a full checkpoint is saved at end of each epoch\n :param save_final_model: If True, final model is saved\n :param anneal_with_restarts: If True, the last best model is restored when annealing the learning rate\n :param shuffle: If True, data is shuffled during training\n :param param_selection_mode: If True, testing is performed against dev data. Use this mode when doing\n parameter selection.\n :param num_workers: Number of workers in your data loader.\n :param sampler: You can pass a data sampler here for special sampling of data.\n :param kwargs: Other arguments for the Optimizer\n :return:\n "
if self.use_tensorboard:
try:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter()
except:
log_line(log)
log.warning(
'ATTENTION! PyTorch >= 1.1.0 and pillow are required for TensorBoard support!'
)
log_line(log)
self.use_tensorboard = False
pass
if use_amp:
if (sys.version_info < (3, 0)):
raise RuntimeError(
'Apex currently only supports Python 3. Aborting.')
if (amp is None):
raise RuntimeError(
'Failed to import apex. Please install apex from https://www.github.com/nvidia/apex to enable mixed-precision training.'
)
if (eval_mini_batch_size is None):
eval_mini_batch_size = mini_batch_size
if (type(base_path) is str):
base_path = Path(base_path)
log_handler = add_file_handler(log, (base_path / 'training.log'))
log_line(log)
log.info(''.join(['Model: "', '{}'.format(self.model), '"']))
log_line(log)
log.info(''.join(['Corpus: "', '{}'.format(self.corpus), '"']))
log_line(log)
log.info('Parameters:')
log.info(''.join(
[' - learning_rate: "', '{}'.format(learning_rate), '"']))
log.info(''.join(
[' - mini_batch_size: "', '{}'.format(mini_batch_size), '"']))
log.info(''.join([' - patience: "', '{}'.format(patience), '"']))
log.info(''.join(
[' - anneal_factor: "', '{}'.format(anneal_factor), '"']))
log.info(''.join([' - max_epochs: "', '{}'.format(max_epochs), '"']))
log.info(''.join([' - shuffle: "', '{}'.format(shuffle), '"']))
log.info(''.join(
[' - train_with_dev: "', '{}'.format(train_with_dev), '"']))
log_line(log)
log.info(''.join(
['Model training base path: "', '{}'.format(base_path), '"']))
log_line(log)
log.info(''.join(['Device: ', '{}'.format(flair.device)]))
log_line(log)
log.info(''.join([
'Embeddings storage mode: ', '{}'.format(embeddings_storage_mode)
]))
log_train = (True if monitor_train else False)
log_test = (True if ((not param_selection_mode) and self.corpus.test
and monitor_test) else False)
log_dev = (True if (not train_with_dev) else False)
loss_txt = init_output_file(base_path, 'loss.tsv')
weight_extractor = WeightExtractor(base_path)
optimizer = self.optimizer(self.model.parameters(),
lr=learning_rate,
**kwargs)
if (self.optimizer_state is not None):
optimizer.load_state_dict(self.optimizer_state)
if use_amp:
(self.model, optimizer) = amp.initialize(self.model,
optimizer,
opt_level=amp_opt_level)
anneal_mode = ('min' if train_with_dev else 'max')
scheduler = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True)
if (self.scheduler_state is not None):
scheduler.load_state_dict(self.scheduler_state)
train_data = self.corpus.train
if train_with_dev:
train_data = ConcatDataset([self.corpus.train, self.corpus.dev])
if (sampler is not None):
sampler = sampler(train_data)
shuffle = False
dev_score_history = []
dev_loss_history = []
train_loss_history = []
try:
previous_learning_rate = learning_rate
for epoch in range((0 + self.epoch), (max_epochs + self.epoch)):
log_line(log)
for group in optimizer.param_groups:
learning_rate = group['lr']
if ((learning_rate != previous_learning_rate)
and anneal_with_restarts
and (base_path / 'best-model.pt').exists()):
log.info('resetting to best model')
self.model.load((base_path / 'best-model.pt'))
previous_learning_rate = learning_rate
if (learning_rate < min_learning_rate):
log_line(log)
log.info('learning rate too small - quitting training!')
log_line(log)
break
batch_loader = DataLoader(train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
sampler=sampler)
self.model.train()
train_loss = 0
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int((total_number_of_batches / 10)))
batch_time = 0
for (batch_no, batch) in enumerate(batch_loader):
start_time = time.time()
loss = self.model.forward_loss(batch)
optimizer.zero_grad()
if use_amp:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_batches += 1
train_loss += loss.item()
store_embeddings(batch, embeddings_storage_mode)
batch_time += (time.time() - start_time)
if ((batch_no % modulo) == 0):
log.info(''.join([
'epoch ', '{}'.format((epoch + 1)), ' - iter ',
'{}'.format(batch_no), '/',
'{}'.format(total_number_of_batches), ' - loss ',
'{:.8f}'.format((train_loss / seen_batches)),
' - samples/sec: ', '{:.2f}'.format(
((mini_batch_size * modulo) / batch_time))
]))
batch_time = 0
iteration = ((epoch * total_number_of_batches) +
batch_no)
if (not param_selection_mode):
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
train_loss /= seen_batches
self.model.eval()
log_line(log)
log.info(''.join([
'EPOCH ', '{}'.format((epoch + 1)), ' done: loss ',
'{:.4f}'.format(train_loss), ' - lr ',
'{:.4f}'.format(learning_rate)
]))
if self.use_tensorboard:
writer.add_scalar('train_loss', train_loss, (epoch + 1))
current_score = train_loss
result_line = ''
if log_train:
(train_eval_result, train_loss) = self.model.evaluate(
DataLoader(self.corpus.train,
batch_size=eval_mini_batch_size,
num_workers=num_workers),
embeddings_storage_mode=embeddings_storage_mode)
result_line += ''.join(
['\t', '{}'.format(train_eval_result.log_line)])
store_embeddings(self.corpus.train,
embeddings_storage_mode)
if log_dev:
(dev_eval_result, dev_loss) = self.model.evaluate(
DataLoader(self.corpus.dev,
batch_size=eval_mini_batch_size,
num_workers=num_workers),
embeddings_storage_mode=embeddings_storage_mode)
result_line += ''.join([
'\t', '{}'.format(dev_loss), '\t',
'{}'.format(dev_eval_result.log_line)
])
log.info(''.join([
'DEV : loss ', '{}'.format(dev_loss), ' - score ',
'{}'.format(dev_eval_result.main_score)
]))
dev_score_history.append(dev_eval_result.main_score)
dev_loss_history.append(dev_loss)
current_score = dev_eval_result.main_score
store_embeddings(self.corpus.dev, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar('dev_loss', dev_loss, (epoch + 1))
writer.add_scalar('dev_score',
dev_eval_result.main_score,
(epoch + 1))
if log_test:
(test_eval_result, test_loss) = self.model.evaluate(
DataLoader(self.corpus.test,
batch_size=eval_mini_batch_size,
num_workers=num_workers),
(base_path / 'test.tsv'),
embeddings_storage_mode=embeddings_storage_mode)
result_line += ''.join([
'\t', '{}'.format(test_loss), '\t',
'{}'.format(test_eval_result.log_line)
])
log.info(''.join([
'TEST : loss ', '{}'.format(test_loss), ' - score ',
'{}'.format(test_eval_result.main_score)
]))
store_embeddings(self.corpus.test, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar('test_loss', test_loss, (epoch + 1))
writer.add_scalar('test_score',
test_eval_result.main_score,
(epoch + 1))
scheduler.step(current_score)
train_loss_history.append(train_loss)
try:
bad_epochs = scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
new_learning_rate = group['lr']
if (new_learning_rate != previous_learning_rate):
bad_epochs = (patience + 1)
log.info(''.join(
['BAD EPOCHS (no improvement): ',
'{}'.format(bad_epochs)]))
with open(loss_txt, 'a') as f:
if (epoch == 0):
f.write(
'EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS'
)
if log_train:
f.write(('\tTRAIN_' + '\tTRAIN_'.join(
train_eval_result.log_header.split('\t'))))
if log_dev:
f.write(('\tDEV_LOSS\tDEV_' + '\tDEV_'.join(
dev_eval_result.log_header.split('\t'))))
if log_test:
f.write(('\tTEST_LOSS\tTEST_' + '\tTEST_'.join(
test_eval_result.log_header.split('\t'))))
f.write(''.join([
'\n', '{}'.format(epoch), '\t',
'{:%H:%M:%S}'.format(datetime.datetime.now()), '\t',
'{}'.format(bad_epochs), '\t',
'{:.4f}'.format(learning_rate), '\t',
'{}'.format(train_loss)
]))
f.write(result_line)
if (checkpoint and (not param_selection_mode)):
self.model.save_checkpoint((base_path / 'checkpoint.pt'),
optimizer.state_dict(),
scheduler.state_dict(),
(epoch + 1), train_loss)
if ((not train_with_dev) and (not param_selection_mode)
and (current_score == scheduler.best)):
self.model.save((base_path / 'best-model.pt'))
if (save_final_model and (not param_selection_mode)):
self.model.save((base_path / 'final-model.pt'))
except KeyboardInterrupt:
log_line(log)
log.info('Exiting from training early.')
if self.use_tensorboard:
writer.close()
if (not param_selection_mode):
log.info('Saving model ...')
self.model.save((base_path / 'final-model.pt'))
log.info('Done.')
if self.corpus.test:
final_score = self.final_test(base_path, eval_mini_batch_size,
num_workers)
else:
final_score = 0
log.info('Test data not provided setting final score to 0')
log.removeHandler(log_handler)
if self.use_tensorboard:
writer.close()
return {
'test_score': final_score,
'dev_score_history': dev_score_history,
'train_loss_history': train_loss_history,
'dev_loss_history': dev_loss_history,
}
def train(
self,
base_path: Union[Path, str],
learning_rate: float = 0.1,
mini_batch_size: int = 32,
mini_batch_chunk_size: Optional[int] = None,
max_epochs: int = 100,
train_with_dev: bool = False,
train_with_test: bool = False,
monitor_train: bool = False,
monitor_test: bool = False,
main_evaluation_metric: Tuple[str, str] = ("micro avg", 'f1-score'),
scheduler=AnnealOnPlateau,
anneal_factor: float = 0.5,
patience: int = 3,
min_learning_rate: float = 0.0001,
initial_extra_patience: int = 0,
optimizer: torch.optim.Optimizer = SGD,
cycle_momentum: bool = False,
warmup_fraction: float = 0.1,
embeddings_storage_mode: str = "cpu",
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
anneal_with_prestarts: bool = False,
anneal_against_dev_loss: bool = False,
batch_growth_annealing: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
write_weights: bool = False,
num_workers: int = 6,
sampler=None,
use_amp: bool = False,
amp_opt_level: str = "O1",
eval_on_train_fraction: float = 0.0,
eval_on_train_shuffle: bool = False,
save_model_each_k_epochs: int = 0,
tensorboard_comment: str = '',
use_swa: bool = False,
use_final_model_for_eval: bool = False,
gold_label_dictionary_for_eval: Optional[Dictionary] = None,
create_file_logs: bool = True,
create_loss_file: bool = True,
epoch: int = 0,
use_tensorboard: bool = False,
tensorboard_log_dir=None,
metrics_for_tensorboard=[],
optimizer_state_dict: Optional = None,
scheduler_state_dict: Optional = None,
save_optimizer_state: bool = False,
**kwargs,
) -> dict:
"""
Trains any class that implements the flair.nn.Model interface.
:param base_path: Main path to which all output during training is logged and models are saved
:param learning_rate: Initial learning rate (or max, if scheduler is OneCycleLR)
:param mini_batch_size: Size of mini-batches during training
:param mini_batch_chunk_size: If mini-batches are larger than this number, they get broken down into chunks of this size for processing purposes
:param max_epochs: Maximum number of epochs to train. Terminates training if this number is surpassed.
:param scheduler: The learning rate scheduler to use
:param checkpoint: If True, a full checkpoint is saved at end of each epoch
:param cycle_momentum: If scheduler is OneCycleLR, whether the scheduler should cycle also the momentum
:param anneal_factor: The factor by which the learning rate is annealed
:param patience: Patience is the number of epochs with no improvement the Trainer waits
until annealing the learning rate
:param min_learning_rate: If the learning rate falls below this threshold, training terminates
:param warmup_fraction: Fraction of warmup steps if the scheduler is LinearSchedulerWithWarmup
:param train_with_dev: If True, the data from dev split is added to the training data
:param train_with_test: If True, the data from test split is added to the training data
:param monitor_train: If True, training data is evaluated at end of each epoch
:param monitor_test: If True, test data is evaluated at end of each epoch
:param embeddings_storage_mode: One of 'none' (all embeddings are deleted and freshly recomputed),
'cpu' (embeddings are stored on CPU) or 'gpu' (embeddings are stored on GPU)
:param save_final_model: If True, final model is saved
:param anneal_with_restarts: If True, the last best model is restored when annealing the learning rate
:param shuffle: If True, data is shuffled during training
:param param_selection_mode: If True, testing is performed against dev data. Use this mode when doing
parameter selection.
:param num_workers: Number of workers in your data loader.
:param sampler: You can pass a data sampler here for special sampling of data.
:param eval_on_train_fraction: the fraction of train data to do the evaluation on,
if 0. the evaluation is not performed on fraction of training data,
if 'dev' the size is determined from dev set size
:param eval_on_train_shuffle: if True the train data fraction is determined on the start of training
and kept fixed during training, otherwise it's sampled at beginning of each epoch
:param save_model_each_k_epochs: Each k epochs, a model state will be written out. If set to '5', a model will
be saved each 5 epochs. Default is 0 which means no model saving.
:param main_evaluation_metric: Type of metric to use for best model tracking and learning rate scheduling (if dev data is available, otherwise loss will be used), currently only applicable for text_classification_model
:param tensorboard_comment: Comment to use for tensorboard logging
:param create_file_logs: If True, the logs will also be stored in a file 'training.log' in the model folder
:param create_loss_file: If True, the loss will be writen to a file 'loss.tsv' in the model folder
:param optimizer: The optimizer to use (typically SGD or Adam)
:param epoch: The starting epoch (normally 0 but could be higher if you continue training model)
:param use_tensorboard: If True, writes out tensorboard information
:param tensorboard_log_dir: Directory into which tensorboard log files will be written
:param metrics_for_tensorboard: List of tuples that specify which metrics (in addition to the main_score) shall be plotted in tensorboard, could be [("macro avg", 'f1-score'), ("macro avg", 'precision')] for example
:param kwargs: Other arguments for the Optimizer
:return:
"""
# create a model card for this model with Flair and PyTorch version
model_card = {'flair_version': flair.__version__, 'pytorch_version': torch.__version__}
# also record Transformers version if library is loaded
try:
import transformers
model_card['transformers_version'] = transformers.__version__
except:
pass
# remember all parameters used in train() call
local_variables = locals()
training_parameters = {}
for parameter in signature(self.train).parameters:
training_parameters[parameter] = local_variables[parameter]
model_card['training_parameters'] = training_parameters
# add model card to model
self.model.model_card = model_card
if use_tensorboard:
try:
from torch.utils.tensorboard import SummaryWriter
if tensorboard_log_dir is not None and not os.path.exists(tensorboard_log_dir):
os.mkdir(tensorboard_log_dir)
writer = SummaryWriter(log_dir=tensorboard_log_dir, comment=tensorboard_comment)
log.info(f"tensorboard logging path is {tensorboard_log_dir}")
except:
log_line(log)
log.warning("ATTENTION! PyTorch >= 1.1.0 and pillow are required for TensorBoard support!")
log_line(log)
use_tensorboard = False
pass
if use_amp:
if sys.version_info < (3, 0):
raise RuntimeError("Apex currently only supports Python 3. Aborting.")
if amp is None:
raise RuntimeError(
"Failed to import apex. Please install apex from https://www.github.com/nvidia/apex "
"to enable mixed-precision training."
)
if mini_batch_chunk_size is None:
mini_batch_chunk_size = mini_batch_size
if learning_rate < min_learning_rate:
min_learning_rate = learning_rate / 10
initial_learning_rate = learning_rate
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
base_path.mkdir(exist_ok=True, parents=True)
if create_file_logs:
log_handler = add_file_handler(log, base_path / "training.log")
else:
log_handler = None
log_line(log)
log.info(f'Model: "{self.model}"')
log_line(log)
log.info(f'Corpus: "{self.corpus}"')
log_line(log)
log.info("Parameters:")
log.info(f' - learning_rate: "{learning_rate}"')
log.info(f' - mini_batch_size: "{mini_batch_size}"')
log.info(f' - patience: "{patience}"')
log.info(f' - anneal_factor: "{anneal_factor}"')
log.info(f' - max_epochs: "{max_epochs}"')
log.info(f' - shuffle: "{shuffle}"')
log.info(f' - train_with_dev: "{train_with_dev}"')
log.info(f' - batch_growth_annealing: "{batch_growth_annealing}"')
log_line(log)
log.info(f'Model training base path: "{base_path}"')
log_line(log)
log.info(f"Device: {flair.device}")
log_line(log)
log.info(f"Embeddings storage mode: {embeddings_storage_mode}")
if isinstance(self.model, SequenceTagger) and self.model.weight_dict and self.model.use_crf:
log_line(log)
log.warning(f'WARNING: Specified class weights will not take effect when using CRF')
# check for previously saved best models in the current training folder and delete them
self.check_for_and_delete_previous_best_models(base_path)
# determine what splits (train, dev, test) to evaluate and log
log_train = True if monitor_train else False
log_test = True if (not param_selection_mode and self.corpus.test and monitor_test) else False
log_dev = False if train_with_dev or not self.corpus.dev else True
log_train_part = True if (eval_on_train_fraction == "dev" or eval_on_train_fraction > 0.0) else False
if log_train_part:
train_part_size = len(self.corpus.dev) if eval_on_train_fraction == "dev" \
else int(len(self.corpus.train) * eval_on_train_fraction)
assert train_part_size > 0
if not eval_on_train_shuffle:
train_part_indices = list(range(train_part_size))
train_part = torch.utils.data.dataset.Subset(self.corpus.train, train_part_indices)
# prepare loss logging file and set up header
loss_txt = init_output_file(base_path, "loss.tsv") if create_loss_file else None
weight_extractor = WeightExtractor(base_path)
# if optimizer class is passed, instantiate:
if inspect.isclass(optimizer):
optimizer: torch.optim.Optimizer = optimizer(self.model.parameters(), lr=learning_rate, **kwargs)
if use_swa:
import torchcontrib
optimizer = torchcontrib.optim.SWA(optimizer, swa_start=10, swa_freq=5, swa_lr=learning_rate)
if use_amp:
self.model, optimizer = amp.initialize(
self.model, optimizer, opt_level=amp_opt_level
)
# load existing optimizer state dictionary if it exists
if optimizer_state_dict:
optimizer.load_state_dict(optimizer_state_dict)
# minimize training loss if training with dev data, else maximize dev score
anneal_mode = "min" if train_with_dev or anneal_against_dev_loss else "max"
best_validation_score = 100000000000 if train_with_dev or anneal_against_dev_loss else 0.
dataset_size = len(self.corpus.train)
if train_with_dev:
dataset_size += len(self.corpus.dev)
# if scheduler is passed as a class, instantiate
if inspect.isclass(scheduler):
if scheduler == OneCycleLR:
scheduler = OneCycleLR(optimizer,
max_lr=learning_rate,
steps_per_epoch=dataset_size // mini_batch_size + 1,
epochs=max_epochs - epoch,
# if we load a checkpoint, we have already trained for epoch
pct_start=0.0,
cycle_momentum=cycle_momentum)
elif scheduler == LinearSchedulerWithWarmup:
steps_per_epoch = (dataset_size + mini_batch_size - 1) / mini_batch_size
num_train_steps = int(steps_per_epoch * max_epochs)
num_warmup_steps = int(num_train_steps * warmup_fraction)
scheduler = LinearSchedulerWithWarmup(optimizer,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps)
else:
scheduler = scheduler(
optimizer,
factor=anneal_factor,
patience=patience,
initial_extra_patience=initial_extra_patience,
mode=anneal_mode,
verbose=True,
)
# load existing scheduler state dictionary if it exists
if scheduler_state_dict:
scheduler.load_state_dict(scheduler_state_dict)
# update optimizer and scheduler in model card
model_card['training_parameters']['optimizer'] = optimizer
model_card['training_parameters']['scheduler'] = scheduler
if isinstance(scheduler, OneCycleLR) and batch_growth_annealing:
raise ValueError("Batch growth with OneCycle policy is not implemented.")
train_data = self.corpus.train
# if training also uses dev/train data, include in training set
if train_with_dev or train_with_test:
parts = [self.corpus.train]
if train_with_dev: parts.append(self.corpus.dev)
if train_with_test: parts.append(self.corpus.test)
train_data = ConcatDataset(parts)
# initialize sampler if provided
if sampler is not None:
# init with default values if only class is provided
if inspect.isclass(sampler):
sampler = sampler()
# set dataset to sample from
sampler.set_dataset(train_data)
shuffle = False
dev_score_history = []
dev_loss_history = []
train_loss_history = []
micro_batch_size = mini_batch_chunk_size
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
momentum = 0
for group in optimizer.param_groups:
if "momentum" in group:
momentum = group["momentum"]
for epoch in range(epoch + 1, max_epochs + 1):
log_line(log)
# update epoch in model card
self.model.model_card['training_parameters']['epoch'] = epoch
if anneal_with_prestarts:
last_epoch_model_state_dict = copy.deepcopy(self.model.state_dict())
if eval_on_train_shuffle:
train_part_indices = list(range(self.corpus.train))
random.shuffle(train_part_indices)
train_part_indices = train_part_indices[:train_part_size]
train_part = torch.utils.data.dataset.Subset(self.corpus.train, train_part_indices)
# get new learning rate
for group in optimizer.param_groups:
learning_rate = group["lr"]
if learning_rate != previous_learning_rate and batch_growth_annealing:
mini_batch_size *= 2
# reload last best model if annealing with restarts is enabled
if (
(anneal_with_restarts or anneal_with_prestarts)
and learning_rate != previous_learning_rate
and os.path.exists(base_path / "best-model.pt")
):
if anneal_with_restarts:
log.info("resetting to best model")
self.model.load_state_dict(
self.model.load(base_path / "best-model.pt").state_dict()
)
if anneal_with_prestarts:
log.info("resetting to pre-best model")
self.model.load_state_dict(
self.model.load(base_path / "pre-best-model.pt").state_dict()
)
previous_learning_rate = learning_rate
if use_tensorboard:
writer.add_scalar("learning_rate", learning_rate, epoch)
# stop training if learning rate becomes too small
if ((not isinstance(scheduler, (OneCycleLR, LinearSchedulerWithWarmup)) and
learning_rate < min_learning_rate)):
log_line(log)
log.info("learning rate too small - quitting training!")
log_line(log)
break
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle if epoch > 1 else False, # never shuffle the first epoch
num_workers=num_workers,
sampler=sampler,
)
self.model.train()
train_loss: float = 0
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int(total_number_of_batches / 10))
# process mini-batches
batch_time = 0
average_over = 0
for batch_no, batch in enumerate(batch_loader):
start_time = time.time()
# zero the gradients on the model and optimizer
self.model.zero_grad()
optimizer.zero_grad()
# if necessary, make batch_steps
batch_steps = [batch]
if len(batch) > micro_batch_size:
batch_steps = [batch[x: x + micro_batch_size] for x in range(0, len(batch), micro_batch_size)]
# forward and backward for batch
for batch_step in batch_steps:
# forward pass
loss = self.model.forward_loss(batch_step)
if isinstance(loss, Tuple):
average_over += loss[1]
loss = loss[0]
# Backward
if use_amp:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
train_loss += loss.item()
# do the optimizer step
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 5.0)
optimizer.step()
# do the scheduler step if one-cycle or linear decay
if isinstance(scheduler, (OneCycleLR, LinearSchedulerWithWarmup)):
scheduler.step()
# get new learning rate
for group in optimizer.param_groups:
learning_rate = group["lr"]
if "momentum" in group:
momentum = group["momentum"]
if "betas" in group:
momentum, _ = group["betas"]
seen_batches += 1
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(batch, embeddings_storage_mode)
batch_time += time.time() - start_time
if seen_batches % modulo == 0:
momentum_info = f' - momentum: {momentum:.4f}' if cycle_momentum else ''
intermittent_loss = train_loss / average_over if average_over > 0 else train_loss / seen_batches
log.info(
f"epoch {epoch} - iter {seen_batches}/{total_number_of_batches} - loss "
f"{intermittent_loss:.8f} - samples/sec: {mini_batch_size * modulo / batch_time:.2f}"
f" - lr: {learning_rate:.6f}{momentum_info}"
)
batch_time = 0
iteration = epoch * total_number_of_batches + batch_no
if not param_selection_mode and write_weights:
weight_extractor.extract_weights(self.model.state_dict(), iteration)
if average_over != 0:
train_loss /= average_over
self.model.eval()
log_line(log)
log.info(f"EPOCH {epoch} done: loss {train_loss:.4f} - lr {learning_rate:.7f}")
if use_tensorboard:
writer.add_scalar("train_loss", train_loss, epoch)
# evaluate on train / dev / test split depending on training settings
result_line: str = ""
if log_train:
train_eval_result = self.model.evaluate(
self.corpus.train,
gold_label_type=self.model.label_type,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
embedding_storage_mode=embeddings_storage_mode,
main_evaluation_metric=main_evaluation_metric,
gold_label_dictionary=gold_label_dictionary_for_eval,
)
result_line += f"\t{train_eval_result.log_line}"
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.train, embeddings_storage_mode)
if log_train_part:
train_part_eval_result = self.model.evaluate(
train_part,
gold_label_type=self.model.label_type,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
embedding_storage_mode=embeddings_storage_mode,
main_evaluation_metric=main_evaluation_metric,
gold_label_dictionary=gold_label_dictionary_for_eval,
)
result_line += f"\t{train_part_eval_result.loss}\t{train_part_eval_result.log_line}"
log.info(
f"TRAIN_SPLIT : loss {train_part_eval_result.loss} - {main_evaluation_metric[1]} ({main_evaluation_metric[0]}) {round(train_part_eval_result.main_score, 4)}"
)
if use_tensorboard:
for (metric_class_avg_type, metric_type) in metrics_for_tensorboard:
writer.add_scalar(
f"train_{metric_class_avg_type}_{metric_type}",
train_part_eval_result.classification_report[metric_class_avg_type][metric_type], epoch
)
if log_dev:
dev_eval_result = self.model.evaluate(
self.corpus.dev,
gold_label_type=self.model.label_type,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
out_path=base_path / "dev.tsv",
embedding_storage_mode=embeddings_storage_mode,
main_evaluation_metric=main_evaluation_metric,
gold_label_dictionary=gold_label_dictionary_for_eval,
)
result_line += f"\t{dev_eval_result.loss}\t{dev_eval_result.log_line}"
log.info(
f"DEV : loss {dev_eval_result.loss} - {main_evaluation_metric[1]} ({main_evaluation_metric[0]}) {round(dev_eval_result.main_score, 4)}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_score_history.append(dev_eval_result.main_score)
dev_loss_history.append(dev_eval_result.loss)
dev_score = dev_eval_result.main_score
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.dev, embeddings_storage_mode)
if use_tensorboard:
writer.add_scalar("dev_loss", dev_eval_result.loss, epoch)
writer.add_scalar("dev_score", dev_eval_result.main_score, epoch)
for (metric_class_avg_type, metric_type) in metrics_for_tensorboard:
writer.add_scalar(
f"dev_{metric_class_avg_type}_{metric_type}",
dev_eval_result.classification_report[metric_class_avg_type][metric_type], epoch
)
if log_test:
test_eval_result = self.model.evaluate(
self.corpus.test,
gold_label_type=self.model.label_type,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
out_path=base_path / "test.tsv",
embedding_storage_mode=embeddings_storage_mode,
main_evaluation_metric=main_evaluation_metric,
gold_label_dictionary=gold_label_dictionary_for_eval,
)
result_line += f"\t{test_eval_result.loss}\t{test_eval_result.log_line}"
log.info(
f"TEST : loss {test_eval_result.loss} - {main_evaluation_metric[1]} ({main_evaluation_metric[0]}) {round(test_eval_result.main_score, 4)}"
)
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.test, embeddings_storage_mode)
if use_tensorboard:
writer.add_scalar("test_loss", test_eval_result.loss, epoch)
writer.add_scalar("test_score", test_eval_result.main_score, epoch)
for (metric_class_avg_type, metric_type) in metrics_for_tensorboard:
writer.add_scalar(
f"test_{metric_class_avg_type}_{metric_type}",
test_eval_result.classification_report[metric_class_avg_type][metric_type], epoch
)
# determine if this is the best model or if we need to anneal
current_epoch_has_best_model_so_far = False
# default mode: anneal against dev score
if not train_with_dev and not anneal_against_dev_loss:
if dev_score > best_validation_score:
current_epoch_has_best_model_so_far = True
best_validation_score = dev_score
if isinstance(scheduler, AnnealOnPlateau):
scheduler.step(dev_score, dev_eval_result.loss)
# alternative: anneal against dev loss
if not train_with_dev and anneal_against_dev_loss:
if dev_eval_result.loss < best_validation_score:
current_epoch_has_best_model_so_far = True
best_validation_score = dev_eval_result.loss
if isinstance(scheduler, AnnealOnPlateau):
scheduler.step(dev_eval_result.loss)
# alternative: anneal against train loss
if train_with_dev:
if train_loss < best_validation_score:
current_epoch_has_best_model_so_far = True
best_validation_score = train_loss
if isinstance(scheduler, AnnealOnPlateau):
scheduler.step(train_loss)
train_loss_history.append(train_loss)
# determine bad epoch number
try:
bad_epochs = scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
new_learning_rate = group["lr"]
if new_learning_rate != previous_learning_rate:
bad_epochs = patience + 1
if previous_learning_rate == initial_learning_rate: bad_epochs += initial_extra_patience
# log bad epochs
log.info(f"BAD EPOCHS (no improvement): {bad_epochs}")
if create_loss_file:
# output log file
with open(loss_txt, "a") as f:
# make headers on first epoch
if epoch == 1:
f.write(f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS")
if log_train:
f.write("\tTRAIN_" + "\tTRAIN_".join(train_eval_result.log_header.split("\t")))
if log_train_part:
f.write("\tTRAIN_PART_LOSS\tTRAIN_PART_" + "\tTRAIN_PART_".join(
train_part_eval_result.log_header.split("\t")))
if log_dev:
f.write("\tDEV_LOSS\tDEV_" + "\tDEV_".join(dev_eval_result.log_header.split("\t")))
if log_test:
f.write("\tTEST_LOSS\tTEST_" + "\tTEST_".join(test_eval_result.log_header.split("\t")))
f.write(
f"\n{epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t{train_loss}"
)
f.write(result_line)
# if checkpoint is enabled, save model at each epoch
if checkpoint and not param_selection_mode:
self.model.save(base_path / "checkpoint.pt", checkpoint=True)
# Check whether to save best model
if (
(not train_with_dev or anneal_with_restarts or anneal_with_prestarts)
and not param_selection_mode
and current_epoch_has_best_model_so_far
and not use_final_model_for_eval
):
log.info("saving best model")
self.model.save(base_path / "best-model.pt", checkpoint=save_optimizer_state)
if anneal_with_prestarts:
current_state_dict = self.model.state_dict()
self.model.load_state_dict(last_epoch_model_state_dict)
self.model.save(base_path / "pre-best-model.pt")
self.model.load_state_dict(current_state_dict)
if save_model_each_k_epochs > 0 and not epoch % save_model_each_k_epochs:
print("saving model of current epoch")
model_name = "model_epoch_" + str(epoch) + ".pt"
self.model.save(base_path / model_name, checkpoint=save_optimizer_state)
if use_swa:
optimizer.swap_swa_sgd()
# if we do not use dev data for model selection, save final model
if save_final_model and not param_selection_mode:
self.model.save(base_path / "final-model.pt", checkpoint=save_optimizer_state)
except KeyboardInterrupt:
log_line(log)
log.info("Exiting from training early.")
if use_tensorboard:
writer.close()
if not param_selection_mode:
log.info("Saving model ...")
self.model.save(base_path / "final-model.pt", checkpoint=save_optimizer_state)
log.info("Done.")
# test best model if test data is present
if self.corpus.test and not train_with_test:
final_score = self.final_test(
base_path=base_path,
eval_mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
main_evaluation_metric=main_evaluation_metric,
gold_label_dictionary_for_eval=gold_label_dictionary_for_eval,
)
else:
final_score = 0
log.info("Test data not provided setting final score to 0")
if create_file_logs:
log_handler.close()
log.removeHandler(log_handler)
if use_tensorboard:
writer.close()
return {
"test_score": final_score,
"dev_score_history": dev_score_history,
"train_loss_history": train_loss_history,
"dev_loss_history": dev_loss_history,
}
def train(
self,
base_path: Union[Path, str],
learning_rate: float = 0.1,
mini_batch_size: int = 32,
mini_batch_chunk_size: int = None,
max_epochs: int = 100,
scheduler=AnnealOnPlateau,
cycle_momentum: bool = False,
anneal_factor: float = 0.5,
patience: int = 3,
initial_extra_patience=0,
min_learning_rate: float = 0.0001,
train_with_dev: bool = False,
train_with_test: bool = False,
monitor_train: bool = False,
monitor_test: bool = False,
embeddings_storage_mode: str = "cpu",
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
anneal_with_prestarts: bool = False,
batch_growth_annealing: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
write_weights: bool = False,
num_workers: int = 6,
sampler=None,
use_amp: bool = False,
amp_opt_level: str = "O1",
eval_on_train_fraction=0.0,
eval_on_train_shuffle=False,
save_model_at_each_epoch=False,
**kwargs,
) -> dict:
"""
Trains any class that implements the flair.nn.Model interface.
:param base_path: Main path to which all output during training is logged and models are saved
:param learning_rate: Initial learning rate (or max, if scheduler is OneCycleLR)
:param mini_batch_size: Size of mini-batches during training
:param mini_batch_chunk_size: If mini-batches are larger than this number, they get broken down into chunks of this size for processing purposes
:param max_epochs: Maximum number of epochs to train. Terminates training if this number is surpassed.
:param scheduler: The learning rate scheduler to use
:param cycle_momentum: If scheduler is OneCycleLR, whether the scheduler should cycle also the momentum
:param anneal_factor: The factor by which the learning rate is annealed
:param patience: Patience is the number of epochs with no improvement the Trainer waits
until annealing the learning rate
:param min_learning_rate: If the learning rate falls below this threshold, training terminates
:param train_with_dev: If True, training is performed using both train+dev data
:param monitor_train: If True, training data is evaluated at end of each epoch
:param monitor_test: If True, test data is evaluated at end of each epoch
:param embeddings_storage_mode: One of 'none' (all embeddings are deleted and freshly recomputed),
'cpu' (embeddings are stored on CPU) or 'gpu' (embeddings are stored on GPU)
:param checkpoint: If True, a full checkpoint is saved at end of each epoch
:param save_final_model: If True, final model is saved
:param anneal_with_restarts: If True, the last best model is restored when annealing the learning rate
:param shuffle: If True, data is shuffled during training
:param param_selection_mode: If True, testing is performed against dev data. Use this mode when doing
parameter selection.
:param num_workers: Number of workers in your data loader.
:param sampler: You can pass a data sampler here for special sampling of data.
:param eval_on_train_fraction: the fraction of train data to do the evaluation on,
if 0. the evaluation is not performed on fraction of training data,
if 'dev' the size is determined from dev set size
:param eval_on_train_shuffle: if True the train data fraction is determined on the start of training
and kept fixed during training, otherwise it's sampled at beginning of each epoch
:param save_model_at_each_epoch: If True, at each epoch the thus far trained model will be saved
:param kwargs: Other arguments for the Optimizer
:return:
"""
if self.use_tensorboard:
try:
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter()
except:
log_line(log)
log.warning(
"ATTENTION! PyTorch >= 1.1.0 and pillow are required for TensorBoard support!"
)
log_line(log)
self.use_tensorboard = False
pass
if use_amp:
if sys.version_info < (3, 0):
raise RuntimeError(
"Apex currently only supports Python 3. Aborting.")
if amp is None:
raise RuntimeError(
"Failed to import apex. Please install apex from https://www.github.com/nvidia/apex "
"to enable mixed-precision training.")
if mini_batch_chunk_size is None:
mini_batch_chunk_size = mini_batch_size
if learning_rate < min_learning_rate:
min_learning_rate = learning_rate / 10
initial_learning_rate = learning_rate
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
log_handler = add_file_handler(log, base_path / "training.log")
log_line(log)
log.info(f'Model: "{self.model}"')
log_line(log)
log.info(f'Corpus: "{self.corpus}"')
log_line(log)
log.info("Parameters:")
log.info(f' - learning_rate: "{learning_rate}"')
log.info(f' - mini_batch_size: "{mini_batch_size}"')
log.info(f' - patience: "{patience}"')
log.info(f' - anneal_factor: "{anneal_factor}"')
log.info(f' - max_epochs: "{max_epochs}"')
log.info(f' - shuffle: "{shuffle}"')
log.info(f' - train_with_dev: "{train_with_dev}"')
log.info(f' - batch_growth_annealing: "{batch_growth_annealing}"')
log_line(log)
log.info(f'Model training base path: "{base_path}"')
log_line(log)
log.info(f"Device: {flair.device}")
log_line(log)
log.info(f"Embeddings storage mode: {embeddings_storage_mode}")
if isinstance(self.model, SequenceTagger
) and self.model.weight_dict and self.model.use_crf:
log_line(log)
log.warning(
f'WARNING: Specified class weights will not take effect when using CRF'
)
# determine what splits (train, dev, test) to evaluate and log
log_train = True if monitor_train else False
log_test = (True if (not param_selection_mode and self.corpus.test
and monitor_test) else False)
log_dev = False if train_with_dev or not self.corpus.dev else True
log_train_part = (True if (eval_on_train_fraction == "dev"
or eval_on_train_fraction > 0.0) else False)
if log_train_part:
train_part_size = (len(
self.corpus.dev) if eval_on_train_fraction == "dev" else int(
len(self.corpus.train) * eval_on_train_fraction))
assert train_part_size > 0
if not eval_on_train_shuffle:
train_part_indices = list(range(train_part_size))
train_part = torch.utils.data.dataset.Subset(
self.corpus.train, train_part_indices)
# prepare loss logging file and set up header
loss_txt = init_output_file(base_path, "loss.tsv")
weight_extractor = WeightExtractor(base_path)
optimizer: torch.optim.Optimizer = self.optimizer(
self.model.parameters(), lr=learning_rate, **kwargs)
if use_amp:
self.model, optimizer = amp.initialize(self.model,
optimizer,
opt_level=amp_opt_level)
# minimize training loss if training with dev data, else maximize dev score
anneal_mode = "min" if train_with_dev else "max"
if scheduler == OneCycleLR:
dataset_size = len(self.corpus.train)
if train_with_dev:
dataset_size += len(self.corpus.dev)
lr_scheduler = OneCycleLR(
optimizer,
max_lr=learning_rate,
steps_per_epoch=dataset_size // mini_batch_size + 1,
epochs=max_epochs - self.
epoch, # if we load a checkpoint, we have already trained for self.epoch
pct_start=0.0,
cycle_momentum=cycle_momentum)
else:
lr_scheduler = scheduler(
optimizer,
factor=anneal_factor,
patience=patience,
initial_extra_patience=initial_extra_patience,
mode=anneal_mode,
verbose=True,
)
if (isinstance(lr_scheduler, OneCycleLR) and batch_growth_annealing):
raise ValueError(
"Batch growth with OneCycle policy is not implemented.")
train_data = self.corpus.train
# if training also uses dev/train data, include in training set
if train_with_dev or train_with_test:
parts = [self.corpus.train]
if train_with_dev: parts.append(self.corpus.dev)
if train_with_test: parts.append(self.corpus.test)
train_data = ConcatDataset(parts)
# initialize sampler if provided
if sampler is not None:
# init with default values if only class is provided
if inspect.isclass(sampler):
sampler = sampler()
# set dataset to sample from
sampler.set_dataset(train_data)
shuffle = False
dev_score_history = []
dev_loss_history = []
train_loss_history = []
micro_batch_size = mini_batch_chunk_size
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
momentum = 0
for group in optimizer.param_groups:
if "momentum" in group:
momentum = group["momentum"]
for self.epoch in range(self.epoch + 1, max_epochs + 1):
log_line(log)
if anneal_with_prestarts:
last_epoch_model_state_dict = copy.deepcopy(
self.model.state_dict())
if eval_on_train_shuffle:
train_part_indices = list(range(self.corpus.train))
random.shuffle(train_part_indices)
train_part_indices = train_part_indices[:train_part_size]
train_part = torch.utils.data.dataset.Subset(
self.corpus.train, train_part_indices)
# get new learning rate
for group in optimizer.param_groups:
learning_rate = group["lr"]
if learning_rate != previous_learning_rate and batch_growth_annealing:
mini_batch_size *= 2
# reload last best model if annealing with restarts is enabled
if ((anneal_with_restarts or anneal_with_prestarts)
and learning_rate != previous_learning_rate
and (base_path / "best-model.pt").exists()):
if anneal_with_restarts:
log.info("resetting to best model")
self.model.load_state_dict(
self.model.load(base_path /
"best-model.pt").state_dict())
if anneal_with_prestarts:
log.info("resetting to pre-best model")
self.model.load_state_dict(
self.model.load(base_path /
"pre-best-model.pt").state_dict())
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if (not isinstance(lr_scheduler, OneCycleLR)
) and learning_rate < min_learning_rate:
log_line(log)
log.info("learning rate too small - quitting training!")
log_line(log)
break
batch_loader = DataLoader(
train_data,
batch_size=mini_batch_size,
shuffle=shuffle,
num_workers=num_workers,
sampler=sampler,
)
self.model.train()
train_loss: float = 0
seen_batches = 0
total_number_of_batches = len(batch_loader)
modulo = max(1, int(total_number_of_batches / 10))
# process mini-batches
batch_time = 0
for batch_no, batch in enumerate(batch_loader):
start_time = time.time()
# zero the gradients on the model and optimizer
self.model.zero_grad()
optimizer.zero_grad()
# if necessary, make batch_steps
batch_steps = [batch]
if len(batch) > micro_batch_size:
batch_steps = [
batch[x:x + micro_batch_size]
for x in range(0, len(batch), micro_batch_size)
]
# forward and backward for batch
for batch_step in batch_steps:
# forward pass
loss = self.model.forward_loss(batch_step)
# Backward
if use_amp:
with amp.scale_loss(loss,
optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# do the optimizer step
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
# do the scheduler step if one-cycle
if isinstance(lr_scheduler, OneCycleLR):
lr_scheduler.step()
# get new learning rate
for group in optimizer.param_groups:
learning_rate = group["lr"]
if "momentum" in group:
momentum = group["momentum"]
seen_batches += 1
train_loss += loss.item()
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(batch, embeddings_storage_mode)
batch_time += time.time() - start_time
if seen_batches % modulo == 0:
momentum_info = f' - momentum: {momentum:.4f}' if cycle_momentum else ''
log.info(
f"epoch {self.epoch} - iter {seen_batches}/{total_number_of_batches} - loss "
f"{train_loss / seen_batches:.8f} - samples/sec: {mini_batch_size * modulo / batch_time:.2f}"
f" - lr: {learning_rate:.6f}{momentum_info}")
batch_time = 0
iteration = self.epoch * total_number_of_batches + batch_no
if not param_selection_mode and write_weights:
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
train_loss /= seen_batches
self.model.eval()
log_line(log)
log.info(
f"EPOCH {self.epoch} done: loss {train_loss:.4f} - lr {learning_rate:.7f}"
)
if self.use_tensorboard:
writer.add_scalar("train_loss", train_loss, self.epoch)
# anneal against train loss if training with dev, otherwise anneal against dev score
current_score = train_loss
# evaluate on train / dev / test split depending on training settings
result_line: str = ""
if log_train:
train_eval_result, train_loss = self.model.evaluate(
self.corpus.train,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
embedding_storage_mode=embeddings_storage_mode,
)
result_line += f"\t{train_eval_result.log_line}"
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.train,
embeddings_storage_mode)
if log_train_part:
train_part_eval_result, train_part_loss = self.model.evaluate(
train_part,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
embedding_storage_mode=embeddings_storage_mode,
)
result_line += (
f"\t{train_part_loss}\t{train_part_eval_result.log_line}"
)
log.info(
f"TRAIN_SPLIT : loss {train_part_loss} - score {round(train_part_eval_result.main_score, 4)}"
)
if log_dev:
dev_eval_result, dev_loss = self.model.evaluate(
self.corpus.dev,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
out_path=base_path / "dev.tsv",
embedding_storage_mode=embeddings_storage_mode,
)
result_line += f"\t{dev_loss}\t{dev_eval_result.log_line}"
log.info(
f"DEV : loss {dev_loss} - score {round(dev_eval_result.main_score, 4)}"
)
# calculate scores using dev data if available
# append dev score to score history
dev_score_history.append(dev_eval_result.main_score)
dev_loss_history.append(dev_loss.item())
current_score = dev_eval_result.main_score
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.dev, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("dev_loss", dev_loss, self.epoch)
writer.add_scalar("dev_score",
dev_eval_result.main_score,
self.epoch)
if log_test:
test_eval_result, test_loss = self.model.evaluate(
self.corpus.test,
mini_batch_size=mini_batch_chunk_size,
num_workers=num_workers,
out_path=base_path / "test.tsv",
embedding_storage_mode=embeddings_storage_mode,
)
result_line += f"\t{test_loss}\t{test_eval_result.log_line}"
log.info(
f"TEST : loss {test_loss} - score {round(test_eval_result.main_score, 4)}"
)
# depending on memory mode, embeddings are moved to CPU, GPU or deleted
store_embeddings(self.corpus.test, embeddings_storage_mode)
if self.use_tensorboard:
writer.add_scalar("test_loss", test_loss, self.epoch)
writer.add_scalar("test_score",
test_eval_result.main_score,
self.epoch)
# determine learning rate annealing through scheduler. Use auxiliary metric for AnnealOnPlateau
if log_dev and isinstance(lr_scheduler, AnnealOnPlateau):
lr_scheduler.step(current_score, dev_loss)
elif not isinstance(lr_scheduler, OneCycleLR):
lr_scheduler.step(current_score)
train_loss_history.append(train_loss)
# determine bad epoch number
try:
bad_epochs = lr_scheduler.num_bad_epochs
except:
bad_epochs = 0
for group in optimizer.param_groups:
new_learning_rate = group["lr"]
if new_learning_rate != previous_learning_rate:
bad_epochs = patience + 1
if previous_learning_rate == initial_learning_rate:
bad_epochs += initial_extra_patience
# log bad epochs
log.info(f"BAD EPOCHS (no improvement): {bad_epochs}")
# output log file
with open(loss_txt, "a") as f:
# make headers on first epoch
if self.epoch == 1:
f.write(
f"EPOCH\tTIMESTAMP\tBAD_EPOCHS\tLEARNING_RATE\tTRAIN_LOSS"
)
if log_train:
f.write("\tTRAIN_" + "\tTRAIN_".join(
train_eval_result.log_header.split("\t")))
if log_train_part:
f.write("\tTRAIN_PART_LOSS\tTRAIN_PART_" +
"\tTRAIN_PART_".join(
train_part_eval_result.log_header.
split("\t")))
if log_dev:
f.write("\tDEV_LOSS\tDEV_" + "\tDEV_".join(
dev_eval_result.log_header.split("\t")))
if log_test:
f.write("\tTEST_LOSS\tTEST_" + "\tTEST_".join(
test_eval_result.log_header.split("\t")))
f.write(
f"\n{self.epoch}\t{datetime.datetime.now():%H:%M:%S}\t{bad_epochs}\t{learning_rate:.4f}\t{train_loss}"
)
f.write(result_line)
# if checkpoint is enabled, save model at each epoch
if checkpoint and not param_selection_mode:
self.save_checkpoint(base_path / "checkpoint.pt")
# if we use dev data, remember best model based on dev evaluation score
if ((not train_with_dev or anneal_with_restarts
or anneal_with_prestarts) and not param_selection_mode
and not isinstance(lr_scheduler, OneCycleLR)
and current_score == lr_scheduler.best
and bad_epochs == 0):
print("saving best model")
self.model.save(base_path / "best-model.pt")
if anneal_with_prestarts:
current_state_dict = self.model.state_dict()
self.model.load_state_dict(last_epoch_model_state_dict)
self.model.save(base_path / "pre-best-model.pt")
self.model.load_state_dict(current_state_dict)
if save_model_at_each_epoch:
print("saving model of current epoch")
model_name = "model_epoch_" + str(self.epoch) + ".pt"
self.model.save(base_path / model_name)
# if we do not use dev data for model selection, save final model
if save_final_model and not param_selection_mode:
self.model.save(base_path / "final-model.pt")
except KeyboardInterrupt:
log_line(log)
log.info("Exiting from training early.")
if self.use_tensorboard:
writer.close()
if not param_selection_mode:
log.info("Saving model ...")
self.model.save(base_path / "final-model.pt")
log.info("Done.")
# test best model if test data is present
if self.corpus.test and not train_with_test:
final_score = self.final_test(base_path, mini_batch_chunk_size,
num_workers)
else:
final_score = 0
log.info("Test data not provided setting final score to 0")
log.removeHandler(log_handler)
if self.use_tensorboard:
writer.close()
return {
"test_score": final_score,
"dev_score_history": dev_score_history,
"train_loss_history": train_loss_history,
"dev_loss_history": dev_loss_history,
}
def find_learning_rate(
self,
base_path: Union[Path, str],
optimizer,
mini_batch_size: int = 32,
start_learning_rate: float = 1e-7,
end_learning_rate: float = 10,
iterations: int = 1000,
stop_early: bool = True,
file_name: str = "learning_rate.tsv",
**kwargs,
) -> Path:
best_loss = None
# cast string to Path
if type(base_path) is str:
base_path = Path(base_path)
base_path.mkdir(exist_ok=True, parents=True)
learning_rate_tsv = init_output_file(base_path, file_name)
with open(learning_rate_tsv, "a") as f:
f.write("ITERATION\tTIMESTAMP\tLEARNING_RATE\tTRAIN_LOSS\n")
optimizer = optimizer(self.model.parameters(), lr=start_learning_rate, **kwargs)
train_data = self.corpus.train
scheduler = ExpAnnealLR(optimizer, end_learning_rate, iterations)
model_state = self.model.state_dict()
self.model.train()
step = 0
loss_list = []
average_loss_list = []
while step < iterations:
batch_loader = DataLoader(train_data, batch_size=mini_batch_size, shuffle=True)
for batch in batch_loader:
step += 1
# forward pass
loss = self.model.forward_loss(batch)
if isinstance(loss, Tuple):
loss = loss[0]
# update optimizer and scheduler
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 5.0)
optimizer.step()
scheduler.step()
learning_rate = scheduler.get_lr()[0]
# append current loss to list of losses for all iterations
loss_list.append(loss.item())
# compute averaged loss
import statistics
moving_avg_loss = statistics.mean(loss_list)
average_loss_list.append(moving_avg_loss)
if len(average_loss_list) > 10:
drop = average_loss_list[-10] - moving_avg_loss
else:
drop = 0.
if not best_loss or moving_avg_loss < best_loss:
best_loss = moving_avg_loss
if step > iterations:
break
if stop_early and (moving_avg_loss > 4 * best_loss or torch.isnan(loss)):
log_line(log)
log.info("loss diverged - stopping early!")
step = iterations
break
with open(str(learning_rate_tsv), "a") as f:
f.write(f"{step}\t{learning_rate}\t{loss.item()}\t{moving_avg_loss}\t{drop}\n")
self.model.load_state_dict(model_state)
self.model.to(flair.device)
log_line(log)
log.info(f"learning rate finder finished - plot {learning_rate_tsv}")
log_line(log)
return Path(learning_rate_tsv)
def train(
self,
base_path: str,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 4,
train_with_dev: bool = False,
embeddings_in_memory: bool = True,
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
):
evaluation_method = 'F1'
if self.model.tag_type in ['pos', 'upos']:
evaluation_method = 'accuracy'
log.info('Evaluation method: {}'.format(evaluation_method))
loss_txt = init_output_file(base_path, 'loss.tsv')
with open(loss_txt, 'a') as f:
f.write(
'EPOCH\tTIMESTAMP\tTRAIN_LOSS\t{}\tDEV_LOSS\t{}\tTEST_LOSS\t{}\n'
.format(Metric.tsv_header('TRAIN'), Metric.tsv_header('DEV'),
Metric.tsv_header('TEST')))
weight_extractor = WeightExtractor(base_path)
optimizer = torch.optim.SGD(self.model.parameters(), lr=learning_rate)
# annealing scheduler
anneal_mode = 'min' if train_with_dev else 'max'
scheduler = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data.extend(self.corpus.dev)
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for epoch in range(0, max_epochs):
log.info('-' * 100)
bad_epochs = scheduler.num_bad_epochs
for group in optimizer.param_groups:
learning_rate = group['lr']
# reload last best model if annealing with restarts is enabled
if learning_rate != previous_learning_rate and anneal_with_restarts and \
os.path.exists(base_path + "/best-model.pt"):
log.info('resetting to best model')
self.model.load_from_file(base_path + "/best-model.pt")
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < 0.001:
log.info('learning rate too small - quitting training!')
break
if not self.test_mode:
random.shuffle(train_data)
batches = [
train_data[x:x + mini_batch_size]
for x in range(0, len(train_data), mini_batch_size)
]
self.model.train()
current_loss: float = 0
seen_sentences = 0
modulo = max(1, int(len(batches) / 10))
for batch_no, batch in enumerate(batches):
batch: List[Sentence] = batch
optimizer.zero_grad()
# Step 4. Compute the loss, gradients, and update the parameters by calling optimizer.step()
loss = self.model.neg_log_likelihood(batch)
current_loss += loss.item()
seen_sentences += len(batch)
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
if not embeddings_in_memory:
self.clear_embeddings_in_batch(batch)
if batch_no % modulo == 0:
log.info(
"epoch {0} - iter {1}/{2} - loss {3:.8f}".format(
epoch + 1, batch_no, len(batches),
current_loss / seen_sentences))
iteration = epoch * len(batches) + batch_no
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
current_loss /= len(train_data)
# switch to eval mode
self.model.eval()
# if checkpointing is enable, save model at each epoch
if checkpoint:
self.model.save(base_path + "/checkpoint.pt")
log.info('-' * 100)
dev_score = dev_metric = None
if not train_with_dev:
dev_score, dev_metric = self.evaluate(
self.corpus.dev,
base_path,
evaluation_method=evaluation_method,
embeddings_in_memory=embeddings_in_memory)
test_score, test_metric = self.evaluate(
self.corpus.test,
base_path,
evaluation_method=evaluation_method,
embeddings_in_memory=embeddings_in_memory)
# anneal against train loss if training with dev, otherwise anneal against dev score
scheduler.step(
current_loss) if train_with_dev else scheduler.step(
dev_score)
# logging info
log.info("EPOCH {0}: lr {1:.4f} - bad epochs {2}".format(
epoch + 1, learning_rate, bad_epochs))
if not train_with_dev:
log.info(
"{0:<4}: f-score {1:.4f} - acc {2:.4f} - tp {3} - fp {4} - fn {5} - tn {6}"
.format('DEV', dev_metric.f_score(),
dev_metric.accuracy(), dev_metric._tp,
dev_metric._fp, dev_metric._fn,
dev_metric._tn))
log.info(
"{0:<4}: f-score {1:.4f} - acc {2:.4f} - tp {3} - fp {4} - fn {5} - tn {6}"
.format('TEST', test_metric.f_score(),
test_metric.accuracy(), test_metric._tp,
test_metric._fp, test_metric._fn, test_metric._tn))
with open(loss_txt, 'a') as f:
dev_metric_str = dev_metric.to_tsv(
) if dev_metric is not None else Metric.to_empty_tsv()
f.write('{}\t{:%H:%M:%S}\t{}\t{}\t{}\t{}\t{}\t{}\n'.format(
epoch, datetime.datetime.now(), '_',
Metric.to_empty_tsv(), '_', dev_metric_str, '_',
test_metric.to_tsv()))
# if we use dev data, remember best model based on dev evaluation score
if not train_with_dev and dev_score == scheduler.best:
self.model.save(base_path + "/best-model.pt")
# if we do not use dev data for model selection, save final model
if save_final_model:
self.model.save(base_path + "/final-model.pt")
except KeyboardInterrupt:
log.info('-' * 100)
log.info('Exiting from training early.')
log.info('Saving model ...')
self.model.save(base_path + "/final-model.pt")
log.info('Done.')
def prepare_data(
self,
base_path: Union[Path, str],
learning_rate: float = 0.1,
mini_batch_size: int = 32,
eval_mini_batch_size: int = None,
anneal_factor: float = 0.5,
patience: int = 3,
min_learning_rate: float = 0.0001,
train_with_dev: bool = False,
monitor_train: bool = False,
monitor_test: bool = False,
embedding_storage_mode: str = "cpu",
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
shuffle: bool = True,
param_selection_mode: bool = False,
num_workers: int = 6,
**kwargs,
) -> dict:
"""
Trains any class that implements the flair.nn.Model interface.
:param base_path: Main path to which all output during training is logged and models are saved
:param learning_rate: Initial learning rate
:param mini_batch_size: Size of mini-batches during training
:param eval_mini_batch_size: Size of mini-batches during evaluation
:param max_epochs: Maximum number of epochs to train. Terminates training if this number is surpassed.
:param anneal_factor: The factor by which the learning rate is annealed
:param patience: Patience is the number of epochs with no improvement the Trainer waits
until annealing the learning rate
:param min_learning_rate: If the learning rate falls below this threshold, training terminates
:param train_with_dev: If True, training is performed using both train+dev data
:param monitor_train: If True, training data is evaluated at end of each epoch
:param monitor_test: If True, test data is evaluated at end of each epoch
:param embedding_storage_mode: One of 'none' (all embeddings are deleted and freshly recomputed),
'cpu' (embeddings are stored on CPU) or 'gpu' (embeddings are stored on GPU)
:param checkpoint: If True, a full checkpoint is saved at end of each epoch
:param save_final_model: If True, final model is saved
:param anneal_with_restarts: If True, the last best model is restored when annealing the learning rate
:param shuffle: If True, data is shuffled during training
:param param_selection_mode: If True, testing is performed against dev data. Use this mode when doing
parameter selection.
:param num_workers: Number of workers in your data loader.
:param sampler: You can pass a data sampler here for special sampling of data.
:param kwargs: Other arguments for the Optimizer
:return:
"""
self.shuffle = shuffle
self.embedding_storage_mode = embedding_storage_mode
self.checkpoint = checkpoint
self.save_final_model = save_final_model
self.anneal_with_restarts = anneal_with_restarts
self.num_workers = num_workers
self.mini_batch_size = mini_batch_size
if eval_mini_batch_size is None:
self.eval_mini_batch_size = mini_batch_size
else:
self.eval_mini_batch_size = eval_mini_batch_size
# cast string to Path
if type(base_path) is str:
self.base_path = Path(base_path)
self.log_handler = add_file_handler(log,
self.base_path / "training.log")
if self.display_name is not None:
log_line(log)
log.info(f'Model: {self.display_name}')
log_line(log)
log.info(f'Model: "{self.model}"')
log_line(log)
log.info(f'Corpus: "{self.corpus}"')
log_line(log)
log.info("Parameters:")
log.info(f' - learning_rate: "{learning_rate}"')
log.info(f' - mini_batch_size: "{mini_batch_size}"')
log.info(f' - patience: "{patience}"')
log.info(f' - anneal_factor: "{anneal_factor}"')
log.info(f' - max_epochs: "{self.max_epochs}"')
log.info(f' - shuffle: "{shuffle}"')
log.info(f' - train_with_dev: "{train_with_dev}"')
log_line(log)
log.info(f'Model training base path: "{base_path}"')
log_line(log)
log.info(f"Device: {flair.device}")
log_line(log)
log.info(f"Embedding storage mode: {embedding_storage_mode}")
# determine what splits (train, dev, test) to evaluate and log
self.monitor_train = monitor_train
self.monitor_test = monitor_test
self.param_selection_mode = param_selection_mode
self.train_with_dev = train_with_dev
self.log_train = True if self.monitor_train else False
self.log_test = (True if
(not self.param_selection_mode and self.corpus.test
and self.monitor_test) else False)
self.log_dev = True if not self.train_with_dev else False
# prepare loss logging file and set up header
self.loss_txt = init_output_file(self.base_path, "loss.tsv")
self.weight_extractor = WeightExtractor(self.base_path)
self.learning_rate = learning_rate
self.min_learning_rate = min_learning_rate
self.previous_learning_rate = learning_rate
self.optimizer: torch.optim.Optimizer = self.optimizer_type(
self.model.parameters(), lr=self.learning_rate, **kwargs)
if self.optimizer_state is not None:
optimizer.load_state_dict(self.optimizer_state)
# minimize training loss if training with dev data, else maximize dev score
self.anneal_mode = "min" if self.train_with_dev else "max"
self.anneal_factor = anneal_factor
self.patience = patience
self.scheduler: ReduceLROnPlateau = ReduceLROnPlateau(
self.optimizer,
factor=self.anneal_factor,
patience=self.patience,
mode=self.anneal_mode,
verbose=True,
)
if self.scheduler_state is not None:
self.scheduler.load_state_dict(self.scheduler_state)
self.train_data = self.corpus.train
# if training also uses dev data, include in training set
if self.train_with_dev:
self.train_data = ConcatDataset(
[self.corpus.train, self.corpus.dev])
self.dev_score_history = []
self.dev_loss_history = []
self.train_loss_history = []
