def test_train_test_tokenization_consistency(self):
filepath = os.path.abspath(
os.path.join(os.path.dirname(__file__), 'data', 'testdata.csv'))
df = pd.read_csv(filepath)
X = []
Y = []
for i, row in df.iterrows():
X.append(row["text"])
labels = json.loads(row["question_843"])
for label in labels:
label['start'] = label['startOffset']
label['end'] = label['endOffset']
label['text'] = row["text"][label['start']:label['end']]
Y.append(labels)
for multilabel_setting in [True, False]:
for base_model in [GPT, GPT2, BERT]:
model = SequenceLabeler(
chunk_long_sequences=True,
base_model=base_model,
multi_label_sequences=multilabel_setting)
train_encoded = [
x for x in model.input_pipeline._text_to_ids(
X, Y=Y, pad_token=model.config.pad_token)
]
test_encoded = [
x for x in model.input_pipeline._text_to_ids(X)
]
for chunk_id in range(len(train_encoded)):
for train_token_ids, test_token_ids in zip(
train_encoded[chunk_id].token_ids,
test_encoded[chunk_id].token_ids):
self.assertEqual(train_token_ids[0], test_token_ids[0])
def test_fit_predict_multi_model(self):
"""
Ensure model training does not error out
Ensure model returns predictions
"""
self.model = SequenceLabeler(batch_size=2,
max_length=256,
lm_loss_coef=0.0,
multi_label_sequences=True)
raw_docs = ["".join(text) for text in self.texts]
texts, annotations = finetune_to_indico_sequence(
raw_docs,
self.texts,
self.labels,
none_value=self.model.config.pad_token)
train_texts, test_texts, train_annotations, _ = train_test_split(
texts, annotations, test_size=0.1)
self.model.fit(train_texts, train_annotations)
self.model.predict(test_texts)
probas = self.model.predict_proba(test_texts)
self.assertIsInstance(probas, list)
self.assertIsInstance(probas[0], list)
self.assertIsInstance(probas[0][0], dict)
self.assertIsInstance(probas[0][0]['confidence'], dict)
self.model.save(self.save_file)
model = SequenceLabeler.load(self.save_file)
model.predict(test_texts)
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.model_config.update(
dict(
# general params that differ from finetune
base_model=RoBERTa,
batch_size=4,
predict_batch_size=10,
val_size=0.0,
crf_sequence_labeling=False,
low_memory_mode=True,
class_weights="log",
# auxiliary-specific params
use_auxiliary_info=True,
context_dim=4,
default_context={
'left': 0,
'right': 0,
'top': 0,
'bottom': 0,
},
n_context_embed_per_channel=48,
context_in_base_model=True,
n_layers_with_aux=-1)
)
self.model_config.update(kwargs)
self.model = SequenceLabeler(**self.model_config)
def setUp(self):
self.save_file = 'tests/saved-models/test-save-load'
random.seed(42)
np.random.seed(42)
with open(self.processed_path, 'rt') as fp:
self.texts, self.labels = json.load(fp)
self.model = SequenceLabeler(**self.default_config())
def setUp(self):
self.save_file = 'tests/saved-models/test-save-load'
with open(self.processed_path, 'rt') as fp:
self.texts, self.labels = json.load(fp)
tf.reset_default_graph()
self.model = SequenceLabeler(batch_size=2, max_length=256, verbose=False)
def test_sequence_labeler_no_auxiliary(self):
"""
Ensure model training does not error out
Ensure model returns reasonable predictions
"""
model = SequenceLabeler(**self.default_config(
use_auxiliary_info=False, val_set=(self.trainX, self.trainY)))
model.fit(self.trainX, self.trainY_seq)
preds = model.predict(self.trainX)
self._evaluate_sequence_preds(preds, includes_context=False)
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.model_config = dict(
use_auxiliary_info = False,
n_layers_with_aux = 0,
context_in_base_model = False,
context_dim = 0
)
self.model_config.update(kwargs)
self.model = SequenceLabeler(**self.model_config)
def test_sequence_labeler_auxiliary(self):
"""
Ensure model training does not error out
Ensure model returns reasonable predictions
"""
# here we want to make sure we're actually using context
model = SequenceLabeler(**self.default_config(n_epochs=1500))
model.fit(self.trainX, self.trainY_seq, context=self.train_context)
preds = model.predict(self.trainX, context=self.train_context)
self._evaluate_sequence_preds(preds, includes_context=True)
def setUpClass(cls):
cls._download_data()
#dataset preparation
cls.classifier_dataset = pd.read_csv(cls.classifier_dataset_path, nrows=cls.n_sample * 10)
path = os.path.join(os.path.dirname(__file__), "data", "testdata.json")
with open(path, 'rt') as fp:
cls.texts, cls.labels = json.load(fp)
cls.animals = ["dog", "cat", "horse", "cow", "pig", "sheep", "goat", "chicken", "guinea pig", "donkey", "turkey", "duck", "camel", "goose", "llama", "rabbit", "fox"]
cls.numbers = ["one", "two", "three", "four", "five", "six", "seven", "eight", "nine", "ten", "eleven", "twelve", "thirteen", "fourteen", "fifteen", "sixteen"]
#train and save sequence labeler for later use
try:
cls.s = SequenceLabeler.load(cls.sequence_labeler_path, **cls.default_seq_config(cls))
except FileNotFoundError:
cls.s = SequenceLabeler(**cls.default_seq_config(cls))
cls.s.fit(cls.texts * 10, cls.labels * 10)
cls.s.save(cls.sequence_labeler_path)
#train and save classifier for later use
train_sample = cls.classifier_dataset.sample(n=cls.n_sample*10)
try:
cls.cl = Classifier.load(cls.classifier_path)
except FileNotFoundError:
cls.cl = Classifier(**cls.default_config(cls))
cls.cl.fit(train_sample.Text, train_sample.Target)
cls.cl.save(cls.classifier_path)
if cls.do_comparison:
#train and save comparison regressor for use
cls.cr = ComparisonRegressor()
n_per = 150
similar = []
different = []
for dataset in [cls.animals, cls.numbers]:
for i in range(n_per // 2):
similar.append([random.choice(dataset), random.choice(dataset)])
for i in range(n_per):
different.append([random.choice(cls.animals), random.choice(cls.numbers)])
targets = np.asarray([1] * len(similar) + [0] * len(different))
data = similar + different
cls.x_tr, cls.x_te, cls.t_tr, cls.t_te = train_test_split(data, targets, test_size=0.3, random_state=42)
try:
cls.cr = ComparisonRegressor.load(cls.comparison_regressor_path, **cls.default_config(cls))
except FileNotFoundError:
cls.cr = ComparisonRegressor(**cls.default_config(cls))
cls.cr.fit(cls.x_tr, cls.t_tr)
cls.cr.save(cls.comparison_regressor_path)
def test_fit_predict(self):
"""
Ensure model training does not error out
Ensure model returns predictions
Ensure class reweighting behaves as intended
"""
raw_docs = ["".join(text) for text in self.texts]
texts, annotations = finetune_to_indico_sequence(
raw_docs, self.texts, self.labels, none_value=self.model.config.pad_token
)
train_texts, test_texts, train_annotations, test_annotations = train_test_split(
texts, annotations, test_size=0.1
)
reweighted_model = SequenceLabeler(
**self.default_config(class_weights={"Named Entity": 100.0})
)
reweighted_model.fit(train_texts, train_annotations)
reweighted_predictions = reweighted_model.predict(test_texts)
reweighted_token_recall = sequence_labeling_token_recall(
test_annotations, reweighted_predictions
)
self.model.fit(train_texts, train_annotations)
predictions = self.model.predict(test_texts)
probas = self.model.predict_proba(test_texts)
self.assertIsInstance(probas, list)
self.assertIsInstance(probas[0], list)
self.assertIsInstance(probas[0][0], dict)
self.assertIsInstance(probas[0][0]["confidence"], dict)
token_precision = sequence_labeling_token_precision(
test_annotations, predictions
)
token_recall = sequence_labeling_token_recall(test_annotations, predictions)
overlap_precision = sequence_labeling_overlap_precision(
test_annotations, predictions
)
overlap_recall = sequence_labeling_overlap_recall(test_annotations, predictions)
self.assertIn("Named Entity", token_precision)
self.assertIn("Named Entity", token_recall)
self.assertIn("Named Entity", overlap_precision)
self.assertIn("Named Entity", overlap_recall)
self.model.save(self.save_file)
self.assertGreater(
reweighted_token_recall["Named Entity"], token_recall["Named Entity"]
)
def test_auxiliary_sequence_labeler(self):
"""
Ensure model training does not error out
Ensure model returns reasonable predictions
"""
(trainX, testX, trainY, testY) = self.dataset
model = SequenceLabeler(**self.default_config())
model.fit(trainX, trainY)
preds = model.predict(testX)
token_precision = sequence_labeling_token_precision(preds, testY)
token_recall = sequence_labeling_token_recall(preds, testY)
self.assertIn("Named Entity", token_precision)
self.assertIn("Named Entity", token_recall)
token_precision = np.mean(list(token_precision.values()))
token_recall = np.mean(list(token_recall.values()))
self.assertGreater(token_precision, 0.6)
self.assertGreater(token_recall, 0.6)
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.model_config.update(dict(
pos_injection=True,
n_layers_with_aux = 0,
context_in_base_model = False
))
self.model_config.update(kwargs)
self.model = SequenceLabeler(**self.model_config)
def __init__(self, *args, **kwargs):
"""Initialize internal classifier."""
super().__init__(auto_resample=False, *args, **kwargs)
self.model = SequenceLabeler(val_size=0)
texts.append(c.text)
labels.append(label)
docs.append(texts)
docs_labels.append(labels)
fd.close()
os.remove(XML_PATH)
raw_texts = ["".join(doc) for doc in docs]
texts, annotations = finetune_to_indico_sequence(raw_texts, docs, docs_labels, none_value="<PAD>",
subtoken_predictions=True)
df = pd.DataFrame({'texts': texts, 'annotations': [json.dumps(annotation) for annotation in annotations]})
df.to_csv(DATA_PATH)
if __name__ == "__main__":
dataset = Reuters().dataframe
dataset['annotations'] = [json.loads(annotation) for annotation in dataset['annotations']]
trainX, testX, trainY, testY = train_test_split(
dataset.texts.values,
dataset.annotations.values,
test_size=0.3,
random_state=42
)
model = SequenceLabeler(base_model=GPT2, batch_size=2, val_size=0., chunk_long_sequences=True, subtoken_predictions=True)
model.fit(trainX, trainY)
predictions = model.predict(testX)
print(predictions)
print(annotation_report(testY, predictions))
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.model = SequenceLabeler(val_size=0)
fd.close()
os.remove(XML_PATH)
raw_texts = ["".join(doc) for doc in docs]
texts, annotations = finetune_to_indico_sequence(
raw_texts, docs, docs_labels)
df = pd.DataFrame({
'texts':
texts,
'annotations':
[json.dumps(annotation) for annotation in annotations]
})
df.to_csv(DATA_PATH)
if __name__ == "__main__":
dataset = Reuters().dataframe
dataset['annotations'] = [
json.loads(annotation) for annotation in dataset['annotations']
]
trainX, testX, trainY, testY = train_test_split(dataset.texts.values,
dataset.annotations.values,
test_size=0.3,
random_state=42)
model = SequenceLabeler(batch_size=2,
val_size=0.,
chunk_long_sequences=True)
model.fit(trainX, trainY)
predictions = model.predict(testX)
print(annotation_report(testY, predictions))
none_value="<PAD>",
subtoken_predictions=True)
df = pd.DataFrame({
'texts':
texts,
'annotations':
[json.dumps(annotation) for annotation in annotations]
})
df.to_csv(DATA_PATH)
if __name__ == "__main__":
dataset = Reuters().dataframe
dataset['annotations'] = [
json.loads(annotation) for annotation in dataset['annotations']
]
trainX, testX, trainY, testY = train_test_split(dataset.texts.values,
dataset.annotations.values,
test_size=0.7,
random_state=42)
model = SequenceLabeler(base_model=RoBERTa,
batch_size=1,
val_size=0.,
max_length=16,
chunk_long_sequences=True,
subtoken_predictions=True)
model.fit(trainX, trainY)
predictions = model.predict(testX)
print(predictions)
print(annotation_report(testY, predictions))
fd.close()
os.remove(XML_PATH)
raw_texts = ["".join(doc) for doc in docs]
texts, annotations = finetune_to_indico_sequence(
raw_texts, docs, docs_labels)
df = pd.DataFrame({
'texts':
texts,
'annotations':
[json.dumps(annotation) for annotation in annotations]
})
df.to_csv(DATA_PATH)
if __name__ == "__main__":
dataset = Reuters(nrows=1000).dataframe
dataset['annotations'] = [
json.loads(annotation) for annotation in dataset['annotations']
]
trainX, testX, trainY, testY = train_test_split(dataset.texts,
dataset.annotations,
test_size=0.3,
random_state=42)
model = SequenceLabeler(verbose=False)
model.fit(trainX, trainY)
predictions = model.predict(testX)
n_sample = 10
for i in range(n_sample):
print(testX.values[i], predictions[i])
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.model_config = dict(val_size=0)
self.model_config.update(kwargs)
self.model = SequenceLabeler(**self.model_config)
subtoken_predictions=True)
df = pd.DataFrame({
'texts':
texts,
'annotations':
[json.dumps(annotation) for annotation in annotations]
})
df.to_csv(DATA_PATH)
if __name__ == "__main__":
dataset = Reuters().dataframe
dataset['annotations'] = [
json.loads(annotation) for annotation in dataset['annotations']
]
trainX, testX, trainY, testY = train_test_split(dataset.texts.values,
dataset.annotations.values,
test_size=0.3,
random_state=42)
model = SequenceLabeler(base_model=GPT2,
batch_size=2,
val_size=0.,
max_length=16,
chunk_long_sequences=True,
subtoken_predictions=True,
filter_empty_examples=True)
model.fit(trainX, trainY)
predictions = model.predict(testX)
print(predictions)
print(annotation_report(testY, predictions))
df = pd.DataFrame({
'texts':
texts,
'annotations':
[json.dumps(annotation) for annotation in annotations]
})
df.to_csv(DATA_PATH)
if __name__ == "__main__":
dataset = Reuters().dataframe
dataset['annotations'] = [
json.loads(annotation) for annotation in dataset['annotations']
]
trainX, testX, trainY, testY = train_test_split(dataset.texts.values,
dataset.annotations.values,
test_size=0.7,
random_state=42)
model = SequenceLabeler(base_model=RoBERTa,
batch_size=1,
n_epochs=3,
val_size=0.0,
max_length=16,
chunk_long_sequences=True,
subtoken_predictions=True,
crf_sequence_labeling=True)
model.fit(trainX, trainY)
predictions = model.predict(testX)
print(predictions)
print(annotation_report(testY, predictions))
os.remove(XML_PATH)
raw_texts = ["".join(doc) for doc in docs]
texts, annotations = finetune_to_indico_sequence(
raw_texts, docs, docs_labels)
df = pd.DataFrame({
'texts':
texts,
'annotations':
[json.dumps(annotation) for annotation in annotations]
})
df.to_csv(DATA_PATH)
if __name__ == "__main__":
dataset = Reuters(nrows=1000).dataframe
dataset['annotations'] = [
json.loads(annotation) for annotation in dataset['annotations']
]
trainX, testX, trainY, testY = train_test_split(dataset.texts.values,
dataset.annotations.values,
test_size=0.3,
random_state=42)
model = SequenceLabeler(verbose=False,
max_length=64,
chunk_long_sequences=True)
model.fit(trainX, trainY)
predictions = model.predict(testX)
n_sample = 10
for i in range(n_sample):
print(testX[i], predictions[i])
