def _evaluate_with_span_F1(self, data_loader, embedding_storage_mode,
mini_batch_size, out_path):
eval_loss = 0
batch_no: int = 0
metric = Metric("Evaluation", beta=self.beta)
lines: List[str] = []
y_true = []
y_pred = []
for batch in data_loader:
# predict for batch
loss = self.predict(batch,
embedding_storage_mode=embedding_storage_mode,
mini_batch_size=mini_batch_size,
label_name='predicted',
return_loss=True)
eval_loss += loss
batch_no += 1
for sentence in batch:
# make list of gold tags
gold_spans = sentence.get_spans(self.tag_type)
gold_tags = [(span.tag, repr(span)) for span in gold_spans]
# make list of predicted tags
predicted_spans = sentence.get_spans("predicted")
predicted_tags = [(span.tag, repr(span))
for span in predicted_spans]
# check for true positives, false positives and false negatives
for tag, prediction in predicted_tags:
if (tag, prediction) in gold_tags:
metric.add_tp(tag)
else:
metric.add_fp(tag)
for tag, gold in gold_tags:
if (tag, gold) not in predicted_tags:
metric.add_fn(tag)
tags_gold = []
tags_pred = []
# also write to file in BIO format to use old conlleval script
if out_path:
for token in sentence:
# check if in gold spans
gold_tag = 'O'
for span in gold_spans:
if token in span:
gold_tag = 'B-' + span.tag if token == span[
0] else 'I-' + span.tag
tags_gold.append(gold_tag)
predicted_tag = 'O'
# check if in predicted spans
for span in predicted_spans:
if token in span:
predicted_tag = 'B-' + span.tag if token == span[
0] else 'I-' + span.tag
tags_pred.append(predicted_tag)
lines.append(
f'{token.text} {gold_tag} {predicted_tag}\n')
lines.append('\n')
y_true.append(tags_gold)
y_pred.append(tags_pred)
if out_path:
with open(Path(out_path), "w", encoding="utf-8") as outfile:
outfile.write("".join(lines))
eval_loss /= batch_no
detailed_result = (
"\nResults:"
f"\n- F1-score (micro) {metric.micro_avg_f_score():.4f}"
f"\n- F1-score (macro) {metric.macro_avg_f_score():.4f}"
'\n\nBy class:')
for class_name in metric.get_classes():
detailed_result += (
f"\n{class_name:<10} tp: {metric.get_tp(class_name)} - fp: {metric.get_fp(class_name)} - "
f"fn: {metric.get_fn(class_name)} - precision: "
f"{metric.precision(class_name):.4f} - recall: {metric.recall(class_name):.4f} - "
f"f1-score: "
f"{metric.f_score(class_name):.4f}")
result = Result(
main_score=metric.micro_avg_f_score(),
log_line=
f"{metric.precision():.4f}\t{metric.recall():.4f}\t{metric.micro_avg_f_score():.4f}",
log_header="PRECISION\tRECALL\tF1",
detailed_results=detailed_result,
)
return result, eval_loss
def evaluate(self,
evaluation: List[Sentence],
out_path=None,
evaluation_method: str = 'F1',
eval_batch_size: int = 32,
embeddings_in_memory: bool = True):
batch_no: int = 0
batches = [
evaluation[x:x + eval_batch_size]
for x in range(0, len(evaluation), eval_batch_size)
]
metric = Metric('')
lines: List[str] = []
for batch in batches:
batch_no += 1
scores, tag_seq = self.model._predict_scores_batch(batch)
predicted_ids = tag_seq
all_tokens = []
for sentence in batch:
all_tokens.extend(sentence.tokens)
for (token, score, predicted_id) in zip(all_tokens, scores,
predicted_ids):
token: Token = token
# get the predicted tag
predicted_value = self.model.tag_dictionary.get_item_for_index(
predicted_id)
token.add_tag('predicted', predicted_value, score)
for sentence in batch:
# add predicted tags
for token in sentence.tokens:
predicted_tag: Label = token.get_tag('predicted')
# append both to file for evaluation
eval_line = '{} {} {}\n'.format(
token.text,
token.get_tag(self.model.tag_type).value,
predicted_tag.value)
lines.append(eval_line)
lines.append('\n')
# make list of gold tags
gold_tags = [
str(tag) for tag in sentence.get_spans(self.model.tag_type)
]
# make list of predicted tags
predicted_tags = [
str(tag) for tag in sentence.get_spans('predicted')
]
# check for true positives, false positives and false negatives
for prediction in predicted_tags:
if prediction in gold_tags:
metric.tp()
else:
metric.fp()
for gold in gold_tags:
if gold not in predicted_tags:
metric.fn()
if not embeddings_in_memory:
self.clear_embeddings_in_batch(batch)
if out_path is not None:
test_tsv = os.path.join(out_path, "test.tsv")
with open(test_tsv, "w", encoding='utf-8') as outfile:
outfile.write(''.join(lines))
if evaluation_method == 'accuracy':
score = metric.accuracy()
return score, metric
if evaluation_method == 'F1':
score = metric.f_score()
return score, metric
def evaluate(
self,
data_loader: DataLoader,
out_path: Path = None,
embedding_storage_mode: str = "none",
) -> (Result, float):
with torch.no_grad():
eval_loss = 0
metric = Metric("Evaluation")
lines: List[str] = []
batch_count: int = 0
for batch in data_loader:
batch_count += 1
labels, loss = self.forward_labels_and_loss(batch)
eval_loss += loss
sentences_for_batch = [
sent.to_plain_string() for sent in batch
]
confidences_for_batch = [[
label.score for label in sent_labels
] for sent_labels in labels]
predictions_for_batch = [[
label.value for label in sent_labels
] for sent_labels in labels]
true_values_for_batch = [
sentence.get_label_names() for sentence in batch
]
available_labels = self.label_dictionary.get_items()
for sentence, confidence, prediction, true_value in zip(
sentences_for_batch,
confidences_for_batch,
predictions_for_batch,
true_values_for_batch,
):
eval_line = "{}\t{}\t{}\t{}\n".format(
sentence, true_value, prediction, confidence)
lines.append(eval_line)
for predictions_for_sentence, true_values_for_sentence in zip(
predictions_for_batch, true_values_for_batch):
for label in available_labels:
if (label in predictions_for_sentence
and label in true_values_for_sentence):
metric.add_tp(label)
elif (label in predictions_for_sentence
and label not in true_values_for_sentence):
metric.add_fp(label)
elif (label not in predictions_for_sentence
and label in true_values_for_sentence):
metric.add_fn(label)
elif (label not in predictions_for_sentence
and label not in true_values_for_sentence):
metric.add_tn(label)
store_embeddings(batch, embedding_storage_mode)
eval_loss /= batch_count
detailed_result = (
f"\nMICRO_AVG: acc {metric.micro_avg_accuracy()} - f1-score {metric.micro_avg_f_score()}"
f"\nMACRO_AVG: acc {metric.macro_avg_accuracy()} - f1-score {metric.macro_avg_f_score()}"
)
for class_name in metric.get_classes():
detailed_result += (
f"\n{class_name:<10} tp: {metric.get_tp(class_name)} - fp: {metric.get_fp(class_name)} - "
f"fn: {metric.get_fn(class_name)} - tn: {metric.get_tn(class_name)} - precision: "
f"{metric.precision(class_name):.4f} - recall: {metric.recall(class_name):.4f} - "
f"accuracy: {metric.accuracy(class_name):.4f} - f1-score: "
f"{metric.f_score(class_name):.4f}")
result = Result(
main_score=metric.micro_avg_f_score(),
log_line=
f"{metric.precision()}\t{metric.recall()}\t{metric.micro_avg_f_score()}",
log_header="PRECISION\tRECALL\tF1",
detailed_results=detailed_result,
)
if out_path is not None:
with open(out_path, "w", encoding="utf-8") as outfile:
outfile.write("".join(lines))
return result, eval_loss
def evaluate(self,
data_loader: DataLoader,
out_path: Path = None,
embeddings_storage_mode: str = 'cpu') -> (Result, float):
with torch.no_grad():
eval_loss = 0
metric = Metric('Evaluation')
lines = []
batch_count = 0
for batch in data_loader:
batch_count += 1
(labels, loss) = self.forward_labels_and_loss(batch)
eval_loss += loss
sentences_for_batch = [
sent.to_plain_string() for sent in batch
]
confidences_for_batch = [[
label.score for label in sent_labels
] for sent_labels in labels]
predictions_for_batch = [[
label.value for label in sent_labels
] for sent_labels in labels]
true_values_for_batch = [
sentence.get_label_names() for sentence in batch
]
available_labels = self.label_dictionary.get_items()
for (sentence, confidence, prediction, true_value) in zip(
sentences_for_batch, confidences_for_batch,
predictions_for_batch, true_values_for_batch):
eval_line = '{}\t{}\t{}\t{}\n'.format(
sentence, true_value, prediction, confidence)
lines.append(eval_line)
for (predictions_for_sentence,
true_values_for_sentence) in zip(predictions_for_batch,
true_values_for_batch):
for label in available_labels:
if ((label in predictions_for_sentence)
and (label in true_values_for_sentence)):
metric.add_tp(label)
elif ((label in predictions_for_sentence)
and (label not in true_values_for_sentence)):
metric.add_fp(label)
elif ((label not in predictions_for_sentence)
and (label in true_values_for_sentence)):
metric.add_fn(label)
elif ((label not in predictions_for_sentence)
and (label not in true_values_for_sentence)):
metric.add_tn(label)
store_embeddings(batch, embeddings_storage_mode)
eval_loss /= batch_count
detailed_result = ''.join([
'\nMICRO_AVG: acc ', '{}'.format(metric.micro_avg_accuracy()),
' - f1-score ', '{}'.format(metric.micro_avg_f_score()),
'\nMACRO_AVG: acc ', '{}'.format(metric.macro_avg_accuracy()),
' - f1-score ', '{}'.format(metric.macro_avg_f_score())
])
for class_name in metric.get_classes():
detailed_result += ''.join([
'\n', '{:<10}'.format(class_name), ' tp: ',
'{}'.format(metric.get_tp(class_name)), ' - fp: ',
'{}'.format(metric.get_fp(class_name)), ' - fn: ',
'{}'.format(metric.get_fn(class_name)), ' - tn: ',
'{}'.format(metric.get_tn(class_name)), ' - precision: ',
'{:.4f}'.format(metric.precision(class_name)),
' - recall: ', '{:.4f}'.format(metric.recall(class_name)),
' - accuracy: ', '{:.4f}'.format(
metric.accuracy(class_name)), ' - f1-score: ',
'{:.4f}'.format(metric.f_score(class_name))
])
result = Result(main_score=metric.micro_avg_f_score(),
log_line=''.join([
'{}'.format(metric.precision()), '\t',
'{}'.format(metric.recall()), '\t',
'{}'.format(metric.micro_avg_f_score())
]),
log_header='PRECISION\tRECALL\tF1',
detailed_results=detailed_result)
if (out_path is not None):
with open(out_path, 'w', encoding='utf-8') as outfile:
outfile.write(''.join(lines))
return (result, eval_loss)
def evaluate(
self,
sentences: Dataset,
eval_mini_batch_size: int = 32,
embeddings_in_memory: bool = True,
out_path: Path = None,
) -> (Result, float):
with torch.no_grad():
eval_loss = 0
batch_no: int = 0
batch_loader = torch.utils.data.DataLoader(
sentences,
batch_size=eval_mini_batch_size,
shuffle=False,
num_workers=4,
collate_fn=list,
)
metric = Metric("Evaluation")
lines: List[str] = []
for batch in batch_loader:
batch_no += 1
with torch.no_grad():
features = self.forward(batch)
loss = self._calculate_loss(features, batch)
tags, _ = self._obtain_labels(features, batch)
eval_loss += loss
for (sentence, sent_tags) in zip(batch, tags):
for (token, tag) in zip(sentence.tokens, sent_tags):
token: Token = token
token.add_tag_label("predicted", tag)
# append both to file for evaluation
eval_line = "{} {} {} {}\n".format(
token.text,
token.get_tag(self.tag_type).value,
tag.value,
tag.score,
)
lines.append(eval_line)
lines.append("\n")
for sentence in batch:
# make list of gold tags
gold_tags = [
(tag.tag, str(tag)) for tag in sentence.get_spans(self.tag_type)
]
# make list of predicted tags
predicted_tags = [
(tag.tag, str(tag)) for tag in sentence.get_spans("predicted")
]
# check for true positives, false positives and false negatives
for tag, prediction in predicted_tags:
if (tag, prediction) in gold_tags:
metric.add_tp(tag)
else:
metric.add_fp(tag)
for tag, gold in gold_tags:
if (tag, gold) not in predicted_tags:
metric.add_fn(tag)
else:
metric.add_tn(tag)
clear_embeddings(
batch, also_clear_word_embeddings=not embeddings_in_memory
)
eval_loss /= batch_no
if out_path is not None:
with open(out_path, "w", encoding="utf-8") as outfile:
outfile.write("".join(lines))
detailed_result = (
f"\nMICRO_AVG: acc {metric.micro_avg_accuracy()} - f1-score {metric.micro_avg_f_score()}"
f"\nMACRO_AVG: acc {metric.macro_avg_accuracy()} - f1-score {metric.macro_avg_f_score()}"
)
for class_name in metric.get_classes():
detailed_result += (
f"\n{class_name:<10} tp: {metric.get_tp(class_name)} - fp: {metric.get_fp(class_name)} - "
f"fn: {metric.get_fn(class_name)} - tn: {metric.get_tn(class_name)} - precision: "
f"{metric.precision(class_name):.4f} - recall: {metric.recall(class_name):.4f} - "
f"accuracy: {metric.accuracy(class_name):.4f} - f1-score: "
f"{metric.f_score(class_name):.4f}"
)
result = Result(
main_score=metric.micro_avg_f_score(),
log_line=f"{metric.precision()}\t{metric.recall()}\t{metric.micro_avg_f_score()}",
log_header="PRECISION\tRECALL\tF1",
detailed_results=detailed_result,
)
return result, eval_loss
def train(self,
base_path: str,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
max_epochs: int = 50,
anneal_factor: float = 0.5,
patience: int = 5,
save_model: bool = True,
embeddings_in_memory: bool = False,
train_with_dev: bool = False,
eval_on_train: bool = True):
"""
Trains the model using the training data of the corpus.
:param patience: number of 'bad' epochs before learning rate gets decreased
:param anneal_factor: learning rate will be decreased by this factor
:param base_path: the directory to which any results should be written to
:param learning_rate: the learning rate
:param mini_batch_size: the mini batch size
:param max_epochs: the maximum number of epochs to train
:param save_model: boolean value indicating, whether the model should be saved or not
:param embeddings_in_memory: boolean value indicating, if embeddings should be kept in memory or not
:param train_with_dev: boolean value indicating, if the dev data set should be used for training or not
:param eval_on_train: boolean value indicating, if evaluation metrics should be calculated on training data set
or not
"""
loss_txt = init_output_file(base_path, 'loss.tsv')
with open(loss_txt, 'a') as f:
f.write(
'EPOCH\tTIMESTAMP\tTRAIN_LOSS\t{}\tDEV_LOSS\t{}\tTEST_LOSS\t{}\n'
.format(Metric.tsv_header('TRAIN'), Metric.tsv_header('DEV'),
Metric.tsv_header('TEST')))
weight_extractor = WeightExtractor(base_path)
optimizer = torch.optim.SGD(self.model.parameters(), lr=learning_rate)
anneal_mode = 'min' if train_with_dev else 'max'
scheduler: ReduceLROnPlateau = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data.extend(self.corpus.dev)
# At any point you can hit Ctrl + C to break out of training early.
try:
# record overall best dev scores and best loss
best_score = 0
for epoch in range(max_epochs):
log.info('-' * 100)
if not self.test_mode:
random.shuffle(train_data)
self.model.train()
batches = [
self.corpus.train[x:x + mini_batch_size]
for x in range(0, len(self.corpus.train), mini_batch_size)
]
current_loss: float = 0
seen_sentences = 0
modulo = max(1, int(len(batches) / 10))
for group in optimizer.param_groups:
learning_rate = group['lr']
for batch_no, batch in enumerate(batches):
scores = self.model.forward(batch)
loss = self.model.calculate_loss(scores, batch)
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
seen_sentences += len(batch)
current_loss += loss.item()
clear_embeddings(
batch,
also_clear_word_embeddings=not embeddings_in_memory)
if batch_no % modulo == 0:
log.info(
"epoch {0} - iter {1}/{2} - loss {3:.8f}".format(
epoch + 1, batch_no, len(batches),
current_loss / seen_sentences))
iteration = epoch * len(batches) + batch_no
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
current_loss /= len(train_data)
self.model.eval()
log.info('-' * 100)
log.info("EPOCH {0}: lr {1:.4f} - bad epochs {2}".format(
epoch + 1, learning_rate, scheduler.num_bad_epochs))
dev_metric = train_metric = None
dev_loss = '_'
train_loss = current_loss
if eval_on_train:
train_metric, train_loss = self._calculate_evaluation_results_for(
'TRAIN', self.corpus.train, embeddings_in_memory,
mini_batch_size)
if not train_with_dev:
dev_metric, dev_loss = self._calculate_evaluation_results_for(
'DEV', self.corpus.dev, embeddings_in_memory,
mini_batch_size)
with open(loss_txt, 'a') as f:
train_metric_str = train_metric.to_tsv(
) if train_metric is not None else Metric.to_empty_tsv()
dev_metric_str = dev_metric.to_tsv(
) if dev_metric is not None else Metric.to_empty_tsv()
f.write('{}\t{:%H:%M:%S}\t{}\t{}\t{}\t{}\t{}\t{}\n'.format(
epoch, datetime.datetime.now(), train_loss,
train_metric_str, dev_loss, dev_metric_str, '_',
Metric.to_empty_tsv()))
# anneal against train loss if training with dev, otherwise anneal against dev score
scheduler.step(
current_loss) if train_with_dev else scheduler.step(
dev_metric.f_score())
is_best_model_so_far: bool = False
current_score = dev_metric.f_score(
) if not train_with_dev else train_metric.f_score()
if current_score >= best_score:
best_score = current_score
is_best_model_so_far = True
if is_best_model_so_far:
if save_model:
self.model.save(base_path + "/model.pt")
self.model.save(base_path + "/final-model.pt")
if save_model:
self.model = TextClassifier.load_from_file(base_path +
"/model.pt")
log.info('-' * 100)
log.info('Testing using best model ...')
self.model.eval()
test_metrics, test_loss = self.evaluate(
self.corpus.test,
mini_batch_size=mini_batch_size,
eval_class_metrics=True,
embeddings_in_memory=embeddings_in_memory)
for metric in test_metrics.values():
metric.print()
self.model.train()
log.info('-' * 100)
except KeyboardInterrupt:
log.info('-' * 100)
log.info('Exiting from training early.')
log.info('Saving model ...')
with open(base_path + "/final-model.pt", 'wb') as model_save_file:
torch.save(self.model, model_save_file, pickle_protocol=4)
model_save_file.close()
log.info('Done.')
def test_metric_with_classes():
metric = Metric('Test')
metric.add_tp('class-1')
metric.add_tn('class-1')
metric.add_tn('class-1')
metric.add_fp('class-1')
metric.add_tp('class-2')
metric.add_tn('class-2')
metric.add_tn('class-2')
metric.add_fp('class-2')
for i in range(0, 10):
metric.add_tp('class-3')
for i in range(0, 90):
metric.add_fp('class-3')
metric.add_tp('class-4')
metric.add_tn('class-4')
metric.add_tn('class-4')
metric.add_fp('class-4')
assert(metric.precision('class-1') == 0.5)
assert(metric.precision('class-2') == 0.5)
assert(metric.precision('class-3') == 0.1)
assert(metric.precision('class-4') == 0.5)
assert(metric.recall('class-1') == 1)
assert(metric.recall('class-2') == 1)
assert(metric.recall('class-3') == 1)
assert(metric.recall('class-4') == 1)
assert(metric.accuracy() == metric.micro_avg_accuracy())
assert(metric.f_score() == metric.micro_avg_f_score())
assert(metric.f_score('class-1') == 0.6667)
assert(metric.f_score('class-2') == 0.6667)
assert(metric.f_score('class-3') == 0.1818)
assert(metric.f_score('class-4') == 0.6667)
assert(metric.accuracy('class-1') == 0.75)
assert(metric.accuracy('class-2') == 0.75)
assert(metric.accuracy('class-3') == 0.1)
assert(metric.accuracy('class-4') == 0.75)
assert(metric.micro_avg_f_score() == 0.2184)
assert(metric.macro_avg_f_score() == 0.5714)
assert(metric.micro_avg_accuracy() == 0.1696)
assert(metric.macro_avg_accuracy() == 0.5875)
assert(metric.precision() == 0.1226)
assert(metric.recall() == 1)
def evaluate(self,
data_loader: DataLoader,
out_path: Path = None,
embeddings_storage_mode: str = 'cpu') -> (Result, float):
with torch.no_grad():
eval_loss = 0
batch_no = 0
metric = Metric('Evaluation')
lines = []
for batch in data_loader:
batch_no += 1
with torch.no_grad():
features = self.forward(batch)
loss = self._calculate_loss(features, batch)
(tags, _) = self._obtain_labels(features, batch)
eval_loss += loss
for (sentence, sent_tags) in zip(batch, tags):
for (token, tag) in zip(sentence.tokens, sent_tags):
token = token
token.add_tag_label('predicted', tag)
eval_line = '{} {} {} {}\n'.format(
token.text,
token.get_tag(self.tag_type).value, tag.value,
tag.score)
lines.append(eval_line)
lines.append('\n')
for sentence in batch:
gold_tags = [(tag.tag, str(tag))
for tag in sentence.get_spans(self.tag_type)]
predicted_tags = [
(tag.tag, str(tag))
for tag in sentence.get_spans('predicted')
]
for (tag, prediction) in predicted_tags:
if ((tag, prediction) in gold_tags):
metric.add_tp(tag)
else:
metric.add_fp(tag)
for (tag, gold) in gold_tags:
if ((tag, gold) not in predicted_tags):
metric.add_fn(tag)
else:
metric.add_tn(tag)
store_embeddings(batch, embeddings_storage_mode)
eval_loss /= batch_no
if (out_path is not None):
with open(out_path, 'w', encoding='utf-8') as outfile:
outfile.write(''.join(lines))
detailed_result = ''.join([
'\nMICRO_AVG: acc ', '{}'.format(metric.micro_avg_accuracy()),
' - f1-score ', '{}'.format(metric.micro_avg_f_score()),
'\nMACRO_AVG: acc ', '{}'.format(metric.macro_avg_accuracy()),
' - f1-score ', '{}'.format(metric.macro_avg_f_score())
])
for class_name in metric.get_classes():
detailed_result += ''.join([
'\n', '{:<10}'.format(class_name), ' tp: ',
'{}'.format(metric.get_tp(class_name)), ' - fp: ',
'{}'.format(metric.get_fp(class_name)), ' - fn: ',
'{}'.format(metric.get_fn(class_name)), ' - tn: ',
'{}'.format(metric.get_tn(class_name)), ' - precision: ',
'{:.4f}'.format(metric.precision(class_name)),
' - recall: ', '{:.4f}'.format(metric.recall(class_name)),
' - accuracy: ', '{:.4f}'.format(
metric.accuracy(class_name)), ' - f1-score: ',
'{:.4f}'.format(metric.f_score(class_name))
])
result = Result(main_score=metric.micro_avg_f_score(),
log_line=''.join([
'{}'.format(metric.precision()), '\t',
'{}'.format(metric.recall()), '\t',
'{}'.format(metric.micro_avg_f_score())
]),
log_header='PRECISION\tRECALL\tF1',
detailed_results=detailed_result)
return (result, eval_loss)
def evaluate(
self,
data_loader: DataLoader,
out_path: Path = None,
embedding_storage_mode: str = "none",
) -> (Result, float):
if type(out_path) == str:
out_path = Path(out_path)
metric = Metric("Evaluation", beta=self.beta)
parsing_metric = ParsingMetric()
lines: List[str] = []
eval_loss_arc = 0
eval_loss_rel = 0
for batch_idx, batch in enumerate(data_loader):
with torch.no_grad():
score_arc, score_rel = self.forward(batch)
loss_arc, loss_rel = self._calculate_loss(
score_arc, score_rel, batch)
arc_prediction, relation_prediction = self._obtain_labels_(
score_arc, score_rel)
parsing_metric(arc_prediction, relation_prediction, batch)
eval_loss_arc += loss_arc
eval_loss_rel += loss_rel
for (sentence, arcs, sent_tags) in zip(batch, arc_prediction,
relation_prediction):
for (token, arc, tag) in zip(sentence.tokens, arcs, sent_tags):
token: Token = token
token.add_tag_label("predicted", Label(tag))
token.add_tag_label("predicted_head_id", Label(str(arc)))
# append both to file for evaluation
eval_line = "{} {} {} {} {}\n".format(
token.text,
token.tags['dependency'].value,
str(token.head_id),
tag,
str(arc),
)
lines.append(eval_line)
lines.append("\n")
for sentence in batch:
# make list of gold tags
gold_tags = [
token.tags['dependency'].value for token in sentence.tokens
]
# make list of predicted tags
predicted_tags = [
tag.tag for tag in sentence.get_spans("predicted")
]
# check for true positives, false positives and false negatives
for tag_indx, predicted_tag in enumerate(predicted_tags):
if predicted_tag == gold_tags[tag_indx]:
metric.add_tp(tag)
else:
metric.add_fp(tag)
for tag_indx, label_tag in enumerate(gold_tags):
if label_tag != predicted_tags[tag_indx]:
metric.add_fn(tag)
else:
metric.add_tn(tag)
store_embeddings(batch, embedding_storage_mode)
eval_loss_arc /= len(data_loader)
eval_loss_rel /= len(data_loader)
if out_path is not None:
with open(out_path, "w", encoding="utf-8") as outfile:
outfile.write("".join(lines))
detailed_result = (
f"\nUAS : {parsing_metric.get_uas():.4f} - LAS : {parsing_metric.get_las():.4f}"
f"\neval loss rel : {eval_loss_rel:.4f} - eval loss arc : {eval_loss_arc:.4f}"
f"\nMICRO_AVG: acc {metric.micro_avg_accuracy()} - f1-score {metric.micro_avg_f_score()}"
f"\nMACRO_AVG: acc {metric.macro_avg_accuracy()} - f1-score {metric.macro_avg_f_score()}"
)
for class_name in metric.get_classes():
detailed_result += (
f"\n{class_name:<10} tp: {metric.get_tp(class_name)} - fp: {metric.get_fp(class_name)} - "
f"fn: {metric.get_fn(class_name)} - tn: {metric.get_tn(class_name)} - precision: "
f"{metric.precision(class_name):.4f} - recall: {metric.recall(class_name):.4f} - "
f"accuracy: {metric.accuracy(class_name):.4f} - f1-score: "
f"{metric.f_score(class_name):.4f}")
result = Result(
main_score=metric.micro_avg_f_score(),
log_line=
f"{metric.precision()}\t{metric.recall()}\t{metric.micro_avg_f_score()}",
log_header="PRECISION\tRECALL\tF1",
detailed_results=detailed_result,
)
return result, eval_loss_arc + eval_loss_rel
def eval_flair_spans(data, predicted_list, batch_size, out_path=None):
metric = Metric('Evaluation')
mini_batch_size = batch_size
batches = [
data[x:x + mini_batch_size]
for x in range(0, len(data), mini_batch_size)
]
lines: List[str] = []
word_counter = 0
for batch in batches:
for sentence in batch:
for token in sentence.tokens:
tag = Label(predicted_list[word_counter])
word_counter += 1
token.add_tag_label('predicted', tag)
# append both to file for evaluation
eval_line = '{} {} {} {}\n'.format(token.text,
token.get_tag('ner').value,
tag.value, tag.score)
lines.append(eval_line)
lines.append('\n')
for sentence in batch:
# make list of gold tags
gold_tags = [(tag.tag, str(tag))
for tag in sentence.get_spans('ner')]
# make list of predicted tags
predicted_tags = [(tag.tag, str(tag))
for tag in sentence.get_spans('predicted')]
# check for true positives, false positives and false negatives
for tag, prediction in predicted_tags:
if (tag, prediction) in gold_tags:
metric.add_tp(tag)
else:
metric.add_fp(tag)
for tag, gold in gold_tags:
if (tag, gold) not in predicted_tags:
metric.add_fn(tag)
else:
metric.add_tn(tag)
# add metrics scores at the beginning of the file
lines.insert(0, str(metric) + "\n\n")
if out_path is not None:
# create folder for json and corresponding output
if not os.path.exists(os.path.dirname(out_path)):
try:
os.makedirs(os.path.dirname(out_path))
except OSError as exc: # Guard against race condition
if exc.errno != errno.EEXIST:
raise
with open(out_path, "w", encoding='utf-8') as outfile:
outfile.write(''.join(lines))
#
# esnWrapper.model.output_activation = output_activation_training
return metric
def _evaluate_text_classifier(model,
sentences,
eval_mini_batch_size=32,
embeddings_in_memory=False,
out_path=None):
with torch.no_grad():
eval_loss = 0
batches = [
sentences[x:x + eval_mini_batch_size]
for x in range(0, len(sentences), eval_mini_batch_size)
]
metric = Metric('Evaluation')
lines = []
for batch in batches:
labels, loss = model.forward_labels_and_loss(batch)
clear_embeddings(
batch, also_clear_word_embeddings=not embeddings_in_memory)
eval_loss += loss
sentences_for_batch = [
sent.to_plain_string() for sent in batch
]
confidences_for_batch = [[
label.score for label in sent_labels
] for sent_labels in labels]
predictions_for_batch = [[
label.value for label in sent_labels
] for sent_labels in labels]
true_values_for_batch = [
sentence.get_label_names() for sentence in batch
]
available_labels = model.label_dictionary.get_items()
for sentence, confidence, prediction, true_value in zip(
sentences_for_batch, confidences_for_batch,
predictions_for_batch, true_values_for_batch):
eval_line = '{}\t{}\t{}\t{}\n'.format(
sentence, true_value, prediction, confidence)
lines.append(eval_line)
for predictions_for_sentence, true_values_for_sentence in zip(
predictions_for_batch, true_values_for_batch):
ModelTrainer._evaluate_sentence_for_text_classification(
metric, available_labels, predictions_for_sentence,
true_values_for_sentence)
eval_loss /= len(sentences)
if out_path is not None:
with open(out_path, "w", encoding='utf-8') as outfile:
outfile.write(''.join(lines))
return metric, eval_loss
def _evaluate_sequence_tagger(model,
sentences,
eval_mini_batch_size=32,
embeddings_in_memory=True,
out_path=None):
with torch.no_grad():
eval_loss = 0
batch_no = 0
batches = [
sentences[x:x + eval_mini_batch_size]
for x in range(0, len(sentences), eval_mini_batch_size)
]
metric = Metric('Evaluation')
lines = []
for batch in batches:
batch_no += 1
tags, loss = model.forward_labels_and_loss(batch)
eval_loss += loss
for (sentence, sent_tags) in zip(batch, tags):
for (token, tag) in zip(sentence.tokens, sent_tags):
token = token
token.add_tag_label('predicted', tag)
# append both to file for evaluation
eval_line = '{} {} {} {}\n'.format(
token.text,
token.get_tag(model.tag_type).value, tag.value,
tag.score)
lines.append(eval_line)
lines.append('\n')
for sentence in batch:
# make list of gold tags
gold_tags = [(tag.tag, str(tag))
for tag in sentence.get_spans(model.tag_type)]
# make list of predicted tags
predicted_tags = [
(tag.tag, str(tag))
for tag in sentence.get_spans('predicted')
]
# check for true positives, false positives and false negatives
for tag, prediction in predicted_tags:
if (tag, prediction) in gold_tags:
metric.add_tp(tag)
else:
metric.add_fp(tag)
for tag, gold in gold_tags:
if (tag, gold) not in predicted_tags:
metric.add_fn(tag)
else:
metric.add_tn(tag)
clear_embeddings(
batch, also_clear_word_embeddings=not embeddings_in_memory)
eval_loss /= len(sentences)
if out_path is not None:
with open(out_path, "w", encoding='utf-8') as outfile:
outfile.write(''.join(lines))
return metric, eval_loss
def train(self,
base_path,
evaluation_metric=EvaluationMetric.MICRO_F1_SCORE,
learning_rate=0.1,
mini_batch_size=32,
eval_mini_batch_size=None,
max_epochs=100,
anneal_factor=0.5,
patience=3,
anneal_against_train_loss=True,
train_with_dev=False,
monitor_train=False,
embeddings_in_memory=True,
checkpoint=False,
save_final_model=True,
anneal_with_restarts=False,
test_mode=False,
param_selection_mode=False,
**kwargs):
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 = 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
test_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 and self.corpus.test:
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 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)
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: str,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
max_epochs: int = 100,
anneal_factor: float = 0.5,
patience: int = 4,
train_with_dev: bool = False,
embeddings_in_memory: bool = True,
checkpoint: bool = False,
save_final_model: bool = True,
anneal_with_restarts: bool = False,
):
evaluation_method = 'F1'
if self.model.tag_type in ['pos', 'upos']:
evaluation_method = 'accuracy'
log.info('Evaluation method: {}'.format(evaluation_method))
loss_txt = init_output_file(base_path, 'loss.tsv')
with open(loss_txt, 'a') as f:
f.write(
'EPOCH\tTIMESTAMP\tTRAIN_LOSS\t{}\tDEV_LOSS\t{}\tTEST_LOSS\t{}\n'
.format(Metric.tsv_header('TRAIN'), Metric.tsv_header('DEV'),
Metric.tsv_header('TEST')))
weight_extractor = WeightExtractor(base_path)
optimizer = torch.optim.SGD(self.model.parameters(), lr=learning_rate)
# annealing scheduler
anneal_mode = 'min' if train_with_dev else 'max'
scheduler = ReduceLROnPlateau(optimizer,
factor=anneal_factor,
patience=patience,
mode=anneal_mode,
verbose=True)
train_data = self.corpus.train
# if training also uses dev data, include in training set
if train_with_dev:
train_data.extend(self.corpus.dev)
# At any point you can hit Ctrl + C to break out of training early.
try:
previous_learning_rate = learning_rate
for epoch in range(0, max_epochs):
bad_epochs = scheduler.num_bad_epochs
for group in optimizer.param_groups:
learning_rate = group['lr']
# reload last best model if annealing with restarts is enabled
if learning_rate != previous_learning_rate and anneal_with_restarts:
log.info('resetting to best model')
self.model.load_from_file(base_path + "/best-model.pt")
previous_learning_rate = learning_rate
# stop training if learning rate becomes too small
if learning_rate < 0.001:
log.info('learning rate too small - quitting training!')
break
if not self.test_mode: random.shuffle(train_data)
batches = [
train_data[x:x + mini_batch_size]
for x in range(0, len(train_data), mini_batch_size)
]
self.model.train()
current_loss: float = 0
seen_sentences = 0
modulo = max(1, int(len(batches) / 10))
for batch_no, batch in enumerate(batches):
batch: List[Sentence] = batch
optimizer.zero_grad()
# Step 4. Compute the loss, gradients, and update the parameters by calling optimizer.step()
loss = self.model.neg_log_likelihood(batch)
current_loss += loss.item()
seen_sentences += len(batch)
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(),
5.0)
optimizer.step()
if not embeddings_in_memory:
self.clear_embeddings_in_batch(batch)
if batch_no % modulo == 0:
log.info(
"epoch {0} - iter {1}/{2} - loss {3:.8f}".format(
epoch + 1, batch_no, len(batches),
current_loss / seen_sentences))
iteration = epoch * len(batches) + batch_no
weight_extractor.extract_weights(
self.model.state_dict(), iteration)
current_loss /= len(train_data)
# switch to eval mode
self.model.eval()
# if checkpointing is enable, save model at each epoch
if checkpoint:
self.model.save(base_path + "/checkpoint.pt")
log.info('-' * 100)
dev_score = dev_metric = None
if not train_with_dev:
dev_score, dev_metric = self.evaluate(
self.corpus.dev,
base_path,
evaluation_method=evaluation_method,
embeddings_in_memory=embeddings_in_memory)
test_score, test_metric = self.evaluate(
self.corpus.test,
base_path,
evaluation_method=evaluation_method,
embeddings_in_memory=embeddings_in_memory)
# anneal against train loss if training with dev, otherwise anneal against dev score
scheduler.step(
current_loss) if train_with_dev else scheduler.step(
dev_score)
# logging info
log.info("EPOCH {0}: lr {1:.4f} - bad epochs {2}".format(
epoch + 1, learning_rate, bad_epochs))
if not train_with_dev:
log.info(
"{0:<4}: f-score {1:.4f} - acc {2:.4f} - tp {3} - fp {4} - fn {5} - tn {6}"
.format('DEV', dev_metric.f_score(),
dev_metric.accuracy(), dev_metric._tp,
dev_metric._fp, dev_metric._fn,
dev_metric._tn))
log.info(
"{0:<4}: f-score {1:.4f} - acc {2:.4f} - tp {3} - fp {4} - fn {5} - tn {6}"
.format('TEST', test_metric.f_score(),
test_metric.accuracy(), test_metric._tp,
test_metric._fp, test_metric._fn, test_metric._tn))
with open(loss_txt, 'a') as f:
dev_metric_str = dev_metric.to_tsv(
) if dev_metric is not None else Metric.to_empty_tsv()
f.write('{}\t{:%H:%M:%S}\t{}\t{}\t{}\t{}\t{}\t{}\n'.format(
epoch, datetime.datetime.now(), '_',
Metric.to_empty_tsv(), '_', dev_metric_str, '_',
test_metric.to_tsv()))
# if we use dev data, remember best model based on dev evaluation score
if not train_with_dev and dev_score == scheduler.best:
self.model.save(base_path + "/best-model.pt")
# if we do not use dev data for model selection, save final model
if save_final_model:
if train_with_dev:
self.model.save(base_path + "/final-model.pt")
except KeyboardInterrupt:
log.info('-' * 100)
log.info('Exiting from training early.')
log.info('Saving model ...')
self.model.save(base_path + "/final-model.pt")
log.info('Done.')
def test_metric_get_classes():
metric = Metric('Test')
metric.add_fn('class-1')
metric.add_fn('class-3')
metric.add_tn('class-1')
metric.add_tp('class-2')
assert(3 == len(metric.get_classes()))
assert('class-1' in metric.get_classes())
assert('class-2' in metric.get_classes())
assert('class-3' in metric.get_classes())
def evaluate(
self,
data_loader: DataLoader,
out_path: Path = None,
embedding_storage_mode: str = "none",
) -> (Result, float):
if type(out_path) == str:
out_path = Path(out_path)
with torch.no_grad():
eval_loss = 0
batch_no: int = 0
metric = Metric("Evaluation", beta=self.beta)
lines: List[str] = []
if self.use_crf:
transitions = self.transitions.detach().cpu().numpy()
else:
transitions = None
for batch in data_loader:
batch_no += 1
with torch.no_grad():
features = self.forward(batch)
loss = self._calculate_loss(features, batch)
tags, _ = self._obtain_labels(
feature=features,
batch_sentences=batch,
transitions=transitions,
get_all_tags=False,
)
eval_loss += loss
for (sentence, sent_tags) in zip(batch, tags):
for (token, tag) in zip(sentence.tokens, sent_tags):
token: Token = token
token.add_tag("predicted", tag.value, tag.score)
# append both to file for evaluation
eval_line = "{} {} {} {}\n".format(
token.text,
token.get_tag(self.tag_type).value,
tag.value,
tag.score,
)
lines.append(eval_line)
lines.append("\n")
for sentence in batch:
# make list of gold tags
gold_tags = [(tag.tag, str(tag))
for tag in sentence.get_spans(self.tag_type)]
# make list of predicted tags
predicted_tags = [
(tag.tag, str(tag))
for tag in sentence.get_spans("predicted")
]
# check for true positives, false positives and false negatives
for tag, prediction in predicted_tags:
if (tag, prediction) in gold_tags:
metric.add_tp(tag)
else:
metric.add_fp(tag)
for tag, gold in gold_tags:
if (tag, gold) not in predicted_tags:
metric.add_fn(tag)
else:
metric.add_tn(tag)
store_embeddings(batch, embedding_storage_mode)
eval_loss /= batch_no
if out_path is not None:
with open(out_path, "w", encoding="utf-8") as outfile:
outfile.write("".join(lines))
detailed_result = (
f"\nMICRO_AVG: acc {metric.micro_avg_accuracy()} - f1-score {metric.micro_avg_f_score()}"
f"\nMACRO_AVG: acc {metric.macro_avg_accuracy()} - f1-score {metric.macro_avg_f_score()}"
)
for class_name in metric.get_classes():
detailed_result += (
f"\n{class_name:<10} tp: {metric.get_tp(class_name)} - fp: {metric.get_fp(class_name)} - "
f"fn: {metric.get_fn(class_name)} - tn: {metric.get_tn(class_name)} - precision: "
f"{metric.precision(class_name):.4f} - recall: {metric.recall(class_name):.4f} - "
f"accuracy: {metric.accuracy(class_name):.4f} - f1-score: "
f"{metric.f_score(class_name):.4f}")
result = Result(
main_score=metric.micro_avg_f_score(),
log_line=
f"{metric.precision()}\t{metric.recall()}\t{metric.micro_avg_f_score()}",
log_header="PRECISION\tRECALL\tF1",
detailed_results=detailed_result,
)
return result, eval_loss
def evaluate(
self,
sentences: List[Sentence],
eval_mini_batch_size: int = 32,
embeddings_in_memory: bool = False,
out_path: Path = None,
) -> (Result, float):
with torch.no_grad():
eval_loss = 0
batches = [
sentences[x:x + eval_mini_batch_size]
for x in range(0, len(sentences), eval_mini_batch_size)
]
metric = Metric("Evaluation")
lines: List[str] = []
for batch in batches:
labels, loss = self.forward_labels_and_loss(batch)
clear_embeddings(
batch, also_clear_word_embeddings=not embeddings_in_memory)
eval_loss += loss
sentences_for_batch = [
sent.to_plain_string() for sent in batch
]
confidences_for_batch = [[
label.score for label in sent_labels
] for sent_labels in labels]
predictions_for_batch = [[
label.value for label in sent_labels
] for sent_labels in labels]
true_values_for_batch = [
sentence.get_label_names() for sentence in batch
]
available_labels = self.label_dictionary.get_items()
for sentence, confidence, prediction, true_value in zip(
sentences_for_batch,
confidences_for_batch,
predictions_for_batch,
true_values_for_batch,
):
eval_line = "{}\t{}\t{}\t{}\n".format(
sentence, true_value, prediction, confidence)
lines.append(eval_line)
for predictions_for_sentence, true_values_for_sentence in zip(
predictions_for_batch, true_values_for_batch):
for label in available_labels:
if (label in predictions_for_sentence
and label in true_values_for_sentence):
metric.add_tp(label)
elif (label in predictions_for_sentence
and label not in true_values_for_sentence):
metric.add_fp(label)
elif (label not in predictions_for_sentence
and label in true_values_for_sentence):
metric.add_fn(label)
elif (label not in predictions_for_sentence
and label not in true_values_for_sentence):
metric.add_tn(label)
eval_loss /= len(sentences)
detailed_result = (
f"\nMICRO_AVG: acc {metric.micro_avg_accuracy()} - f1-score {metric.micro_avg_f_score()}"
f"\nMACRO_AVG: acc {metric.macro_avg_accuracy()} - f1-score {metric.macro_avg_f_score()}"
)
for class_name in metric.get_classes():
detailed_result += (
f"\n{class_name:<10} tp: {metric.get_tp(class_name)} - fp: {metric.get_fp(class_name)} - "
f"fn: {metric.get_fn(class_name)} - tn: {metric.get_tn(class_name)} - precision: "
f"{metric.precision(class_name):.4f} - recall: {metric.recall(class_name):.4f} - "
f"accuracy: {metric.accuracy(class_name):.4f} - f1-score: "
f"{metric.f_score(class_name):.4f}")
result = Result(
main_score=metric.micro_avg_f_score(),
log_line=
f"{metric.precision()}\t{metric.recall()}\t{metric.micro_avg_f_score()}",
log_header="PRECISION\tRECALL\tF1",
detailed_results=detailed_result,
)
if out_path is not None:
with open(out_path, "w", encoding="utf-8") as outfile:
outfile.write("".join(lines))
return result, eval_loss
