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")
print(metric)
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") == 0.6666666666666666
assert metric.f_score("class-2") == 0.6666666666666666
assert metric.f_score("class-3") == 0.18181818181818182
assert metric.f_score("class-4") == 0.6666666666666666
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() == 0.21848739495798317
assert metric.macro_avg_f_score() == 0.5454545454545454
assert metric.micro_avg_accuracy() == 0.16964285714285715
assert metric.macro_avg_accuracy() == 0.5875
assert metric.precision() == 0.12264150943396226
assert metric.recall() == 1
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") == 0.5
assert metric.accuracy("class-2") == 0.5
assert metric.accuracy("class-3") == 0.1
assert metric.accuracy("class-4") == 0.5
assert metric.micro_avg_f_score() == 0.2184
assert metric.macro_avg_f_score() == 0.5454749999999999
assert metric.micro_avg_accuracy() == 0.1226
assert metric.macro_avg_accuracy() == 0.4
assert metric.precision() == 0.1226
assert metric.recall() == 1
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() == 0.4)
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 = "none",
eval_mode: EvalMode = EvalMode.Standard,
misspell_mode: MisspellingMode = MisspellingMode.Random,
misspelling_rate: float = 0.0,
char_vocab: set = {},
lut: dict = {},
cmx: np.array = None,
typos: dict = {},
correction_mode: CorrectionMode = CorrectionMode.NotSpecified,
eval_dict_name=None,
evaluation_metric: EvaluationMetric = EvaluationMetric.MICRO_F1_SCORE,
) -> (Result, float):
if type(out_path) == str:
out_path = Path(out_path)
from robust_ner.spellcheck import load_correction_dict, get_lang_from_corpus_name
if correction_mode == CorrectionMode.NotSpecified:
eval_dict = None
else:
eval_dict = load_correction_dict(eval_dict_name, log)
# note: use 'save_correction_dict' to re-generate a dictionary
lang = get_lang_from_corpus_name(eval_dict_name)
eval_params = {}
eval_params["eval_mode"] = eval_mode
eval_params["misspelling_rate"] = misspelling_rate
eval_params["misspell_mode"] = misspell_mode
eval_params["char_vocab"] = char_vocab
eval_params["lut"] = lut
eval_params["cmx"] = cmx
eval_params["typos"] = typos
eval_params["correction_mode"] = correction_mode
eval_params["lang"] = lang
eval_params["dictionary"] = eval_dict
with torch.no_grad():
eval_loss = 0
batch_no: int = 0
metric = Metric("Evaluation")
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, eval_params)
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, tag.text)
for tag in sentence.get_spans(self.tag_type)]
# make list of predicted tags
predicted_tags = [
(tag.tag, tag.text)
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, 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 = (
f"\nMICRO_AVG: acc {metric.micro_avg_accuracy():.4f} - f1-score {metric.micro_avg_f_score():.4f}"
f"\nMACRO_AVG: acc {metric.macro_avg_accuracy():.4f} - f1-score {metric.macro_avg_f_score():.4f}"
)
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}")
if evaluation_metric == EvaluationMetric.MICRO_F1_SCORE:
main_score = metric.micro_avg_f_score()
elif evaluation_metric == EvaluationMetric.MACRO_F1_SCORE:
main_score = metric.macro_avg_f_score()
elif evaluation_metric == EvaluationMetric.MICRO_ACCURACY:
main_score = metric.micro_avg_accuracy()
elif evaluation_metric == EvaluationMetric.MACRO_ACCURACY:
main_score = metric.macro_avg_accuracy()
elif evaluation_metric == EvaluationMetric.MEAN_SQUARED_ERROR:
main_score = metric.mean_squared_error()
else:
log.error(f"unknown evaluation metric: {evaluation_metric}")
result = Result(
main_score=main_score,
log_line=
f"{metric.precision():.4f}\t{metric.recall():.4f}\t{main_score:.4f}",
log_header="PRECISION\tRECALL\tF1",
detailed_results=detailed_result,
)
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,
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)
