def final_test(self,
base_path: Path,
eval_mini_batch_size: int,
num_workers: int = 8):
log_line(log)
log.info("Testing using best model ...")
final_score = 0.0
for id, self.model in self.models.items():
self.model.eval()
if (base_path / f"{id}-best-model.pt").exists():
self.model = self.model.load(base_path / f"{id}-best-model.pt")
test_results, test_loss = self.model.evaluate(
DataLoader(
self.corpus.get_test(id),
batch_size=eval_mini_batch_size,
num_workers=num_workers,
),
out_path=base_path / f"{id}-test.tsv",
embedding_storage_mode="none",
)
test_results: Result = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
# get and return the final test score of best model
final_score += test_results.main_score
return final_score / len(self.models)
def final_test(self, base_path: Path, embeddings_in_memory: bool,
evaluation_metric: EvaluationMetric,
eval_mini_batch_size: int):
log_line(log)
log.info('Testing using best model ...')
self.model.eval()
if (base_path / 'best-model.pt').exists():
self.model = TextRegressor.load_from_file(base_path /
'best-model.pt')
test_metric, test_loss = self._evaluate_text_regressor(
self.model,
self.corpus.test,
eval_mini_batch_size=eval_mini_batch_size,
embeddings_in_memory=embeddings_in_memory)
log.info(f'AVG: mse: {test_metric.mean_squared_error():.4f} - '
f'mae: {test_metric.mean_absolute_error():.4f} - '
f'pearson: {test_metric.pearsonr():.4f} - '
f'spearman: {test_metric.spearmanr():.4f}')
log_line(log)
return test_metric.mean_squared_error()
def train(self, max_epochs: int = 100, **kwargs):
log.info(f"Prepare data loader")
for trainer, base_path, lr in zip(self.trainer_list,
self.base_path_list,
self.learning_rate_list):
trainer.max_epochs = max_epochs
trainer.prepare_data(base_path, lr, **kwargs)
if self.pretrain_list is not None:
for i, epochs_pretrain in enumerate(self.pretrain_list):
log_line(log)
log.info(f"Pretrain Trainer {i} for {epochs_pretrain} epochs")
for epoch in range(0, epochs_pretrain):
trainer = self.trainer_list[i]
trainer.cur_epoch = epoch
trainer.prepare_epoch()
for batch_no, batch in enumerate(trainer.batch_loader):
trainer.train_batch(batch_no, batch)
trainer.eval_after_epoch()
log_line(log)
log.info(f"Start multi-task training")
for epoch in range(0, max_epochs):
quit = False
for trainer in self.trainer_list:
trainer.cur_epoch = epoch
trainer.prepare_epoch()
if trainer.learning_rate < trainer.min_learning_rate:
quit = True
if quit:
log.info('Quitting')
break
num_batches = min([len(t.batch_loader) for t in self.trainer_list])
iters = [iter(t.batch_loader) for t in self.trainer_list]
batch_no = 0
while batch_no < num_batches:
for i, trainer in enumerate(self.trainer_list):
batch = next(iters[i])
trainer.train_batch(batch_no, batch)
batch_no += 1
for trainer in self.trainer_list:
trainer.eval_after_epoch()
for trainer in self.trainer_list:
trainer.make_final_predictions()
def optimize(self, space: SearchSpace, max_evals=100):
search_space = space.search_space
best = fmin(
self._objective, search_space, algo=tpe.suggest, max_evals=max_evals
)
log_line(log)
log.info("Optimizing parameter configuration done.")
log.info("Best parameter configuration found:")
for k, v in best.items():
log.info(f"\t{k}: {v}")
log_line(log)
with open(self.param_selection_file, "a") as f:
f.write("best parameter combination\n")
for k, v in best.items():
if isinstance(v, Tuple):
v = ",".join([str(x) for x in v])
f.write(f"\t{k}: {str(v)}\n")
def train(self, max_epochs: int = 100, **kwargs):
self.model.set_output(self.model_mode)
self.max_epochs = max_epochs
self.prepare_data(**kwargs)
for epoch in range(0 + self.epoch, max_epochs + self.epoch):
self.cur_epoch = epoch
self.prepare_epoch()
if self.learning_rate < self.min_learning_rate:
log_line(log)
log.info("Quitting Training as one Model finished")
log_line(log)
break
for batch_no, batch in enumerate(self.batch_loader):
self.train_batch(batch_no, batch)
self.eval_after_epoch()
self.make_final_predictions()
def optimize(self, space: SearchSpace, max_evals=100):
search_space = space.search_space
best = fmin(self._objective,
search_space,
algo=tpe.suggest,
max_evals=max_evals)
log_line(log)
log.info('Optimizing parameter configuration done.')
log.info('Best parameter configuration found:')
for (k, v) in best.items():
log.info(''.join(['\t', '{}'.format(k), ': ', '{}'.format(v)]))
log_line(log)
with open(self.param_selection_file, 'a') as f:
f.write('best parameter combination\n')
for (k, v) in best.items():
if isinstance(v, Tuple):
v = ','.join([str(x) for x in v])
f.write(''.join(
['\t', '{}'.format(k), ': ', '{}'.format(str(v)), '\n']))
def final_test(self,
base_path: Path,
eval_mini_batch_size: int,
num_workers: int = 8):
log_line(log)
log.info('Testing using best model ...')
self.model.eval()
if (base_path / 'best-model.pt').exists():
self.model = self.model.load((base_path / 'best-model.pt'))
(test_results, test_loss) = self.model.evaluate(
DataLoader(self.corpus.test,
batch_size=eval_mini_batch_size,
num_workers=num_workers),
out_path=(base_path / 'test.tsv'),
embeddings_storage_mode='none')
test_results = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
if (type(self.corpus) is MultiCorpus):
for subcorpus in self.corpus.corpora:
log_line(log)
self.model.evaluate(
DataLoader(subcorpus.test,
batch_size=eval_mini_batch_size,
num_workers=num_workers),
out_path=(
base_path /
''.join(['{}'.format(subcorpus.name), '-test.tsv'])),
embeddings_storage_mode='none')
final_score = test_results.main_score
return final_score
def predict(
self,
sentences: Union[List[Sentence], Sentence],
mini_batch_size: int = 32,
num_workers: int = 8,
print_tree: bool = False,
embedding_storage_mode="none",
) -> None:
"""
Predict arcs and tags for Dependency Parser task
:param sentences: a Sentence or a List of Sentence
:param mini_batch_size: mini batch size to use
:param print_tree: set to True to print dependency parser of sentence as tree shape
:param embedding_storage_mode: default is 'none' which is always best. Only set to 'cpu' or 'gpu' if
you wish to not only predict, but also keep the generated embeddings in CPU or GPU memory respectively.
'gpu' to store embeddings in GPU memory.
"""
if not isinstance(sentences, list):
sentences = [sentences]
sentence_dataset = FlairDatapointDataset(sentences)
data_loader = DataLoader(sentence_dataset, batch_size=mini_batch_size, num_workers=num_workers)
for batch in data_loader:
with torch.no_grad():
score_arc, score_rel = self.forward(batch)
arc_prediction, relation_prediction = self._obtain_labels_(score_arc, score_rel)
for sentnce_index, (sentence, sent_tags, sent_arcs) in enumerate(
zip(batch, relation_prediction, arc_prediction)
):
for token_index, (token, tag, head_id) in enumerate(zip(sentence.tokens, sent_tags, sent_arcs)):
token.add_tag(self.tag_type, tag, score_rel[sentnce_index][token_index])
token.head_id = int(head_id)
if print_tree:
tree_printer(sentence, self.tag_type)
log_line(log)
store_embeddings(batch, storage_mode=embedding_storage_mode)
def final_test(self, base_path: Path, embeddings_in_memory: bool,
evaluation_metric: EvaluationMetric,
eval_mini_batch_size: int):
log_line(log)
log.info('Testing using best model ...')
self.model.eval()
if (base_path / 'best-model.pt').exists():
if isinstance(self.model, TextClassifier):
self.model = TextClassifier.load_from_file(base_path /
'best-model.pt')
if isinstance(self.model, SequenceTagger):
self.model = SequenceTagger.load_from_file(base_path /
'best-model.pt')
test_metric, test_loss = self.evaluate(
self.model,
self.corpus.test,
eval_mini_batch_size=eval_mini_batch_size,
embeddings_in_memory=embeddings_in_memory)
log.info(
f'MICRO_AVG: acc {test_metric.micro_avg_accuracy()} - f1-score {test_metric.micro_avg_f_score()}'
)
log.info(
f'MACRO_AVG: acc {test_metric.macro_avg_accuracy()} - f1-score {test_metric.macro_avg_f_score()}'
)
for class_name in test_metric.get_classes():
log.info(
f'{class_name:<10} tp: {test_metric.get_tp(class_name)} - fp: {test_metric.get_fp(class_name)} - '
f'fn: {test_metric.get_fn(class_name)} - tn: {test_metric.get_tn(class_name)} - precision: '
f'{test_metric.precision(class_name):.4f} - recall: {test_metric.recall(class_name):.4f} - '
f'accuracy: {test_metric.accuracy(class_name):.4f} - f1-score: '
f'{test_metric.f_score(class_name):.4f}')
log_line(log)
# if we are training over multiple datasets, do evaluation for each
if type(self.corpus) is MultiCorpus:
for subcorpus in self.corpus.corpora:
log_line(log)
self._calculate_evaluation_results_for(subcorpus.name,
subcorpus.test,
evaluation_metric,
embeddings_in_memory,
eval_mini_batch_size,
base_path / 'test.tsv')
# get and return the final test score of best model
if evaluation_metric == EvaluationMetric.MACRO_ACCURACY:
final_score = test_metric.macro_avg_accuracy()
elif evaluation_metric == EvaluationMetric.MICRO_ACCURACY:
final_score = test_metric.micro_avg_accuracy()
elif evaluation_metric == EvaluationMetric.MACRO_F1_SCORE:
final_score = test_metric.macro_avg_f_score()
else:
final_score = test_metric.micro_avg_f_score()
return final_score
def final_test(
self,
base_path: Path,
embeddings_in_memory: bool,
evaluation_metric: EvaluationMetric,
eval_mini_batch_size: int,
num_workers: int = 8,
eval_mode: EvalMode = EvalMode.Standard,
misspell_mode: MisspellingMode = MisspellingMode.Random,
misspelling_rate: float = 0.0,
char_vocab: set = {},
cmx=None,
lut={},
):
log_line(log)
log.info("Testing using best model ...")
self.model.eval()
if (base_path / "best-model.pt").exists():
self.model = self.model.load(base_path / "best-model.pt")
test_results, test_loss = self.model.evaluate(
self.corpus.test,
eval_mini_batch_size=eval_mini_batch_size,
embeddings_in_memory=embeddings_in_memory,
out_path=base_path / "test.tsv",
num_workers=num_workers,
eval_mode=eval_mode,
misspell_mode=misspell_mode,
misspelling_rate=misspelling_rate,
char_vocab=char_vocab,
cmx=cmx,
lut=lut,
)
test_results: Result = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
# if we are training over multiple datasets, do evaluation for each
if type(self.corpus) is MultiCorpus:
for subcorpus in self.corpus.corpora:
log_line(log)
self.model.evaluate(
subcorpus.test,
eval_mini_batch_size,
embeddings_in_memory,
base_path / f"{subcorpus.name}-test.tsv",
eval_mode=eval_mode,
misspelling_rate=misspelling_rate,
char_vocab=char_vocab,
)
# get and return the final test score of best model
final_score = test_results.main_score
return final_score
def final_test(
self,
base_path: Union[Path, str],
eval_mini_batch_size: int,
main_evaluation_metric: Tuple[str, str],
num_workers: int = 8,
gold_label_dictionary_for_eval: Optional[Dictionary] = None
):
if type(base_path) is str:
base_path = Path(base_path)
base_path.mkdir(exist_ok=True, parents=True)
log_line(log)
self.model.eval()
if (base_path / "best-model.pt").exists():
self.model.load_state_dict(self.model.load(base_path / "best-model.pt").state_dict())
else:
log.info("Testing using last state of model ...")
test_results = self.model.evaluate(
self.corpus.test,
gold_label_type=self.model.label_type,
mini_batch_size=eval_mini_batch_size,
num_workers=num_workers,
out_path=base_path / "test.tsv",
embedding_storage_mode="none",
main_evaluation_metric=main_evaluation_metric,
gold_label_dictionary=gold_label_dictionary_for_eval,
)
test_results: Result = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
# if we are training over multiple datasets, do evaluation for each
if type(self.corpus) is MultiCorpus:
for subcorpus in self.corpus.corpora:
log_line(log)
if subcorpus.test:
subcorpus_results = self.model.evaluate(
subcorpus.test,
gold_label_type=self.model.label_type,
mini_batch_size=eval_mini_batch_size,
num_workers=num_workers,
out_path=base_path / f"{subcorpus.name}-test.tsv",
embedding_storage_mode="none",
main_evaluation_metric=main_evaluation_metric
)
log.info(subcorpus.name)
log.info(subcorpus_results.log_line)
# get and return the final test score of best model
final_score = test_results.main_score
return final_score
def final_test(
self,
base_path: Union[Path, str],
eval_mini_batch_size: int,
num_workers: int = 8,
main_score_type: str = None,
):
if type(base_path) is str:
base_path = Path(base_path)
log_line(log)
self.model.eval()
if (os.path.exists(self.get_best_model_path(base_path, check_model_existance=True))):
log.info("Testing using best model ...")
self.model = self.model.load(self.get_best_model_path(base_path, check_model_existance=True))
else:
log.info("Testing using last state of model ...")
test_results, test_loss = self.model.evaluate(
self.corpus.test,
mini_batch_size=eval_mini_batch_size,
num_workers=num_workers,
out_path=base_path / "test.tsv",
embedding_storage_mode="none",
main_score_type=main_score_type
)
test_results: Result = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
# if we are training over multiple datasets, do evaluation for each
if type(self.corpus) is MultiCorpus:
for subcorpus in self.corpus.corpora:
log_line(log)
if subcorpus.test:
subcorpus_results, subcorpus_loss = self.model.evaluate(
subcorpus.test,
mini_batch_size=eval_mini_batch_size,
num_workers=num_workers,
out_path=base_path / f"{subcorpus.name}-test.tsv",
embedding_storage_mode="none",
main_score_type=main_score_type
)
log.info(subcorpus.name)
log.info(subcorpus_results.log_line)
# get and return the final test score of best model
final_score = test_results.main_score
return final_score
def final_test(self,
base_path: Path,
eval_mini_batch_size: int,
num_workers: int = 8):
log_line(log)
log.info("Testing using best model ...")
if (base_path / "best-model.pt").exists():
cls = type(self.model)
del self.model
self.model = cls.load(base_path / "best-model.pt")
self.model.eval()
test_results, test_loss = self.model.evaluate(
DataLoader(
self.corpus.test,
batch_size=eval_mini_batch_size,
num_workers=num_workers,
),
out_path=base_path / "test.tsv",
embedding_storage_mode="none",
)
test_results: Result = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
# if we are training over multiple datasets, do evaluation for each
if type(self.corpus) is MultiCorpus:
for subcorpus in self.corpus.corpora:
log_line(log)
self.model.evaluate(
DataLoader(
subcorpus.test,
batch_size=eval_mini_batch_size,
num_workers=num_workers,
),
out_path=base_path / f"{subcorpus.name}-test.tsv",
embedding_storage_mode="none",
)
# get and return the final test score of best model
final_score = test_results.main_score
return final_score
def final_test(self,
base_path: Union[Path, str],
eval_mini_batch_size: int,
num_workers: int = 8):
if type(base_path) is str:
base_path = Path(base_path)
log_line(log)
log.info("Testing using best model ...")
self.model.eval()
torch.jit.script(self.model)
#if (base_path / "best-model.pt").exists():
# self.model = self.model.load(base_path / "best-model.pt")
test_results, test_loss = self.model.evaluate(
self.corpus.test,
mini_batch_size=eval_mini_batch_size,
num_workers=num_workers,
out_path=base_path / "test.tsv",
embedding_storage_mode="none",
)
test_results: Result = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
# if we are training over multiple datasets, do evaluation for each
if type(self.corpus) is MultiCorpus:
for subcorpus in self.corpus.corpora:
log_line(log)
self.model.evaluate(
subcorpus.test,
mini_batch_size=eval_mini_batch_size,
num_workers=num_workers,
out_path=base_path / f"{subcorpus.name}-test.tsv",
embedding_storage_mode="none",
)
# get and return the final test score of best model
final_score = test_results.main_score
print("FINAL RESULT", test_loss, final_score)
return final_score
def final_test(
self,
base_path: Path,
embeddings_in_memory: bool,
evaluation_metric: EvaluationMetric,
eval_mini_batch_size: int,
num_workers: int = 8,
):
log_line(log)
log.info("Testing using best model ...")
self.model.eval()
if (base_path / "best-model.pt").exists():
self.model = self.model.load(base_path / "best-model.pt")
test_results, test_loss = self.model.evaluate(
self.corpus.test,
eval_mini_batch_size=eval_mini_batch_size,
embeddings_in_memory=embeddings_in_memory,
out_path=base_path / "test.tsv",
num_workers=num_workers,
)
test_results: Result = test_results
log.info(test_results.log_line)
log.info(test_results.detailed_results)
log_line(log)
# if we are training over multiple datasets, do evaluation for each
if type(self.corpus) is MultiCorpus:
for subcorpus in self.corpus.corpora:
log_line(log)
self.model.evaluate(
subcorpus.test,
eval_mini_batch_size,
embeddings_in_memory,
base_path / f"{subcorpus.name}-test.tsv",
)
# get and return the final test score of best model
final_score = test_results.main_score
return final_score
def _setup(self, config):
args = config["args"]
config.pop("args", None)
self.params = config
self.corpus = args["corpus"]
torch.manual_seed(args["seed"])
if args.cuda:
torch.cuda.manual_seed(args["seed"])
model = args["_set_up_model"](self.params)
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: "{config["learning_rate"]}"')
# log.info(f' - mini_batch_size: "{config["mini_batch_size"]}"')
# log.info(f' - patience: "{config["patience"]}"')
# log.info(f' - anneal_factor: "{config["anneal_factor"]}"')
# Check this once more
weight_extractor = None # WeightExtractor()
training_params = {
key: config[key]
for key in config if key in TRAINING_PARAMETERS
}
model_trainer_parameters = {
key: config[key]
for key in config if key in MODEL_TRAINER_PARAMETERS
}
# This should be enough for initializing all the parameters for the trainer
self.trainer = ModelTrainer(model, self.corpus,
model_trainer_parameters)
def prepare_epoch(self):
log_line(log)
for group in self.optimizer.param_groups:
self.learning_rate = group["lr"]
# reload last best model if annealing with restarts is enabled
if (self.learning_rate != self.previous_learning_rate
and self.anneal_with_restarts
and (base_path / "best-model.pt").exists()):
log.info("resetting to best model")
self.model.load(base_path / "best-model.pt")
self.previous_learning_rate = self.learning_rate
# stop training if learning rate becomes too small
if self.learning_rate < self.min_learning_rate:
log_line(log)
log.info("learning rate too small - quitting training!")
log_line(log)
return
self.batch_loader = DataLoader(
self.train_data,
batch_size=self.mini_batch_size,
shuffle=self.shuffle,
num_workers=self.num_workers,
)
self.model.train()
self.model.set_output(self.model_mode)
self.train_loss: float = 0
self.seen_batches = 0
self.total_number_of_batches = len(self.batch_loader)
self.modulo = max(1, int(self.total_number_of_batches / 10))
self.batch_time = 0
from __future__ import absolute_import
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,
}
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,
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],
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 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 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: 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 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,
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 _objective(self, params: dict):
log_line(log)
log.info(f"Evaluation run: {self.run}")
log.info(f"Evaluating parameter combination:")
for k, v in params.items():
if isinstance(v, Tuple):
v = ",".join([str(x) for x in v])
log.info(f"\t{k}: {str(v)}")
log_line(log)
for sent in self.corpus.get_all_sentences():
sent.clear_embeddings()
scores = []
vars = []
for i in range(0, self.training_runs):
log_line(log)
log.info(f"Training run: {i + 1}")
model = self._set_up_model(params)
training_params = {
key: params[key] for key in params if key in TRAINING_PARAMETERS
}
model_trainer_parameters = {
key: params[key] for key in params if key in MODEL_TRAINER_PARAMETERS
}
trainer: ModelTrainer = ModelTrainer(
model, self.corpus, **model_trainer_parameters
)
result = trainer.train(
self.base_path,
evaluation_metric=self.evaluation_metric,
max_epochs=self.max_epochs,
param_selection_mode=True,
**training_params,
)
# take the average over the last three scores of training
if self.optimization_value == OptimizationValue.DEV_LOSS:
curr_scores = result["dev_loss_history"][-3:]
else:
curr_scores = list(
map(lambda s: 1 - s, result["dev_score_history"][-3:])
)
score = sum(curr_scores) / float(len(curr_scores))
var = np.var(curr_scores)
scores.append(score)
vars.append(var)
# take average over the scores from the different training runs
final_score = sum(scores) / float(len(scores))
final_var = sum(vars) / float(len(vars))
test_score = result["test_score"]
log_line(log)
log.info(f"Done evaluating parameter combination:")
for k, v in params.items():
if isinstance(v, Tuple):
v = ",".join([str(x) for x in v])
log.info(f"\t{k}: {v}")
log.info(f"{self.optimization_value.value}: {final_score}")
log.info(f"variance: {final_var}")
log.info(f"test_score: {test_score}\n")
log_line(log)
with open(self.param_selection_file, "a") as f:
f.write(f"evaluation run {self.run}\n")
for k, v in params.items():
if isinstance(v, Tuple):
v = ",".join([str(x) for x in v])
f.write(f"\t{k}: {str(v)}\n")
f.write(f"{self.optimization_value.value}: {final_score}\n")
f.write(f"variance: {final_var}\n")
f.write(f"test_score: {test_score}\n")
f.write("-" * 100 + "\n")
self.run += 1
return {"status": "ok", "loss": final_score, "loss_variance": final_var}
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 _objective(self, params):
log_line(log)
log.info(u''.join([u'Evaluation run: ', u'{}'.format(self.run)]))
log.info(u''.join([u'Evaluating parameter combination:']))
for (k, v) in params.items():
if isinstance(v, Tuple):
v = u','.join([unicode(x) for x in v])
log.info(u''.join(
[u'\t', u'{}'.format(k), u': ', u'{}'.format(unicode(v))]))
log_line(log)
for sent in self.corpus.get_all_sentences():
sent.clear_embeddings()
scores = []
vars = []
for i in range(0, self.training_runs):
log_line(log)
log.info(u''.join([u'Training run: ', u'{}'.format((i + 1))]))
model = self._set_up_model(params)
training_params = {
key: params[key]
for key in params if (key in TRAINING_PARAMETERS)
}
model_trainer_parameters = {
key: params[key]
for key in params if (key in MODEL_TRAINER_PARAMETERS)
}
trainer = ModelTrainer(model, self.corpus,
**model_trainer_parameters)
result = trainer.train(self.base_path,
evaluation_metric=self.evaluation_metric,
max_epochs=self.max_epochs,
param_selection_mode=True,
**training_params)
if (self.optimization_value == OptimizationValue.DEV_LOSS):
curr_scores = result[u'dev_loss_history'][(-3):]
else:
curr_scores = list(
map((lambda s: (1 - s)),
result[u'dev_score_history'][(-3):]))
score = (sum(curr_scores) / float(len(curr_scores)))
var = np.var(curr_scores)
scores.append(score)
vars.append(var)
final_score = (sum(scores) / float(len(scores)))
final_var = (sum(vars) / float(len(vars)))
test_score = result[u'test_score']
log_line(log)
log.info(u''.join([u'Done evaluating parameter combination:']))
for (k, v) in params.items():
if isinstance(v, Tuple):
v = u','.join([unicode(x) for x in v])
log.info(u''.join([u'\t', u'{}'.format(k), u': ',
u'{}'.format(v)]))
log.info(u''.join([
u'{}'.format(self.optimization_value.value), u': ',
u'{}'.format(final_score)
]))
log.info(u''.join([u'variance: ', u'{}'.format(final_var)]))
log.info(u''.join([u'test_score: ', u'{}'.format(test_score), u'\n']))
log_line(log)
with open(self.param_selection_file, u'a') as f:
f.write(u''.join(
[u'evaluation run ', u'{}'.format(self.run), u'\n']))
for (k, v) in params.items():
if isinstance(v, Tuple):
v = u','.join([unicode(x) for x in v])
f.write(u''.join([
u'\t', u'{}'.format(k), u': ', u'{}'.format(unicode(v)),
u'\n'
]))
f.write(u''.join([
u'{}'.format(self.optimization_value.value), u': ',
u'{}'.format(final_score), u'\n'
]))
f.write(u''.join([u'variance: ', u'{}'.format(final_var), u'\n']))
f.write(u''.join(
[u'test_score: ', u'{}'.format(test_score), u'\n']))
f.write(((u'-' * 100) + u'\n'))
self.run += 1
return {
u'status': u'ok',
u'loss': final_score,
u'loss_variance': final_var,
}
