def convert_offsets_to_spacy_docs(doc_annotated: list) -> list:
"""
Convert a list of tuple of string with their offset to Spacy doc with entities ready
:param doc_annotated: list of tuple (string, array of offsets)
:return: list of spacy doc
"""
model = get_empty_model(load_labels_for_training=False)
docs = list()
for (index, (case_id, text, tags)) in enumerate(doc_annotated):
doc: Doc = model.make_doc(text)
ents = list()
for (start_offset, end_offset, type_name) in tags:
span_doc = doc.char_span(start_offset, end_offset, label=type_name)
if span_doc is not None:
ents.append(span_doc)
else:
print("Issue in offset",
"Index: " + str(index),
"case: " + case_id,
text[start_offset:end_offset],
text,
sep="|")
doc.ents = ents
docs.append(doc)
return docs
def annotate(model_dir_path: str, files_dir_path: List[str],
out_dir_path: str) -> None:
"""
Annotate a sample of the given XML files and save them into the given directory.
:param model_dir_path: the directory of the Spacy model
:param files_dir_path: the directory containing the XML files
:param out_dir_path: the directory where to write the annotations
"""
logging.info("Loading NER model…")
nlp = get_empty_model(load_labels_for_training=False)
nlp = nlp.from_disk(model_dir_path)
# TODO remove when we have retrained
infixes = nlp.Defaults.infixes + [r':', r"(?<=[\W\d_])-|-(?=[\W\d_])"]
infixes_regex = spacy.util.compile_infix_regex(infixes)
nlp.tokenizer.infix_finditer = infixes_regex.finditer
# end of deletion above
entity_typename_builder = EntityTypename()
logging.info("Loading cases…")
cases: List[Case] = list()
for path in files_dir_path:
if path.endswith(".xml"):
case: Case = get_paragraph_from_file(
path=path, keep_paragraph_without_annotation=True)
cases.append(case)
elif path.endswith(".txt"):
with open(path) as f:
lines = f.readlines()
case: Case = list()
for line in lines:
clean_text = line.strip()
if len(clean_text) > 1:
basename = os.path.basename(path)
basename = basename.split(".")[0]
case.append(
Paragraph(basename, clean_text, list(), list()))
cases.append(case)
def train_model(data: list, folder_to_save_model: str, n_iter: int,
batch_size: int, dropout_rate: float):
"""
Train a NER model using Spacy
:param data: list of tuples [(text, offset)]
:param folder_to_save_model: Where to save the learned model. None to skip. Will be overiden with new model
:param n_iter: number iterations of the CNN
:param batch_size: more = less precise / less time to learn
:param dropout_rate: more : learn less / better generalization
"""
nlp = get_empty_model(load_labels_for_training=True)
nlp.vocab.vectors.name = 'spacy_pretrained_vectors'
optimizer = nlp.begin_training()
with tqdm(total=n_iter * ceil(len(data) / batch_size),
unit=" paragraphs",
desc="Learn NER model") as pbar:
for itn in range(n_iter):
pbar.set_description(f"Learn NER model - iteration {itn + 1}")
losses = {}
random.shuffle(data)
batches = util.minibatch(data, batch_size)
for current_batch_item in batches:
case_id, texts, annotations = zip(*current_batch_item)
docs = [nlp.make_doc(text) for text in texts]
gold_with_unknown_bilou = convert_unknown_bilou_bulk(
docs=docs, offsets=annotations)
nlp.update(
docs, # batch of texts
gold_with_unknown_bilou, # batch of annotations
drop=
dropout_rate, # dropout - make it harder to memorise rules
sgd=optimizer, # callable to update weights
losses=losses)
pbar.postfix = "loss: " + str(losses['ner'])
pbar.update()
# save model to output directory
if folder_to_save_model is not None:
folder_to_save_model = Path(folder_to_save_model)
nlp.to_disk(folder_to_save_model)
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
from spacy.tokens.doc import Doc
from match_text_unsafe.build_entity_dictionary import EntityTypename
from misc.convert_to_bilou import convert_unknown_bilou, convert_unknown_bilou_bulk, no_action_bilou
from ner.model_factory import get_empty_model
import pytest
pytest.nlp = get_empty_model(load_labels_for_training=True)
def test_bilou_conv():
doc: Doc = pytest.nlp.make_doc("Ceci est un test.")
offset1 = [(5, 8, "UNKNOWN")]
assert convert_unknown_bilou(
doc, offsets=offset1).ner == ['O', '-', 'O', 'O', 'O']
assert convert_unknown_bilou_bulk(
[doc], [offset1])[0].ner == ['O', '-', 'O', 'O', 'O']
offset2 = [(5, 8, "PERS")]
assert convert_unknown_bilou(
doc, offsets=offset2).ner == ['O', 'U-PERS', 'O', 'O', 'O']
offset3 = [(0, 4, "UNKNOWN")]
assert convert_unknown_bilou(
doc, offsets=offset3).ner == ['-', 'O', 'O', 'O', 'O']
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied. See the License for the
# specific language governing permissions and limitations
# under the License.
from xml_extractions.extract_node_values import get_paragraph_from_file
from ner.model_factory import get_empty_model
from resources.config_provider import get_config_default
config_training = get_config_default()
model_dir_path = config_training["model_dir_path"]
xml_dev_path = config_training["xml_dev_path"]
nlp = get_empty_model(load_labels_for_training=False)
nlp = nlp.from_disk(model_dir_path)
DEV_DATA = get_paragraph_from_file(xml_dev_path,
keep_paragraph_without_annotation=True)
for case_id, texts, xml_extracted_text, annotations in DEV_DATA:
doc = nlp(texts)
spacy_extracted_text_ad_pp = [
ent.text for ent in doc.ents if ent.label_ in ["ADDRESS", "PERS"]
]
spacy_extracted_text = [ent.text for ent in doc.ents]
str_rep_spacy = ' '.join(spacy_extracted_text)
match = [span_xml in str_rep_spacy for span_xml in xml_extracted_text]
def main(data_folder: str, model_path: Optional[str],
output_model: Optional[str], dev_size: float, nb_epochs: int,
print_diff: bool) -> None:
nlp = get_empty_model(load_labels_for_training=True)
if model_path is not None:
nlp = nlp.from_disk(path=model_path)
nlp.tokenizer = get_tokenizer(nlp) # replace tokenizer
nlp.begin_training()
# ner = nlp.get_pipe("ner")
# ner.model.learn_rate = 0.0001
else:
nlp.tokenizer = get_tokenizer(nlp) # replace tokenizer
nlp.begin_training()
all_annotated_files: List[str] = [
os.path.join(data_folder, filename)
for filename in os.listdir(data_folder) if filename.endswith(".txt")
]
random.shuffle(all_annotated_files)
nb_doc_dev_set: int = int(len(all_annotated_files) * dev_size)
dev_file_names = all_annotated_files[0:nb_doc_dev_set]
train_file_names = [
file for file in all_annotated_files if file not in dev_file_names
]
# train_file_names = ["./resources/training_data/generated_annotations.txt"] + train_file_names
content_to_rate = load_content(txt_paths=train_file_names)
content_to_rate_test = load_content(txt_paths=dev_file_names)
print(
f"nb PERS entities {sum([1 for _, offsets in content_to_rate for o in offsets if o.type == 'PERS'])}"
)
if model_path is not None:
print("evaluation without fine tuning")
spacy_evaluate(nlp, content_to_rate_test, print_diff)
optimizer: Optimizer = nlp.resume_training()
for epoch in range(nb_epochs):
print(f"------- {epoch} -------")
random.shuffle(content_to_rate)
losses = dict()
batches = minibatch(content_to_rate, size=compounding(4., 16., 1.001))
for batch_id, batch in enumerate(
tqdm(iterable=batches, unit=" batches", desc="Training")):
try:
batch_gold = convert_batch_to_gold_dataset(model=nlp,
batch=batch)
texts, manual_annotations = zip(
*batch_gold) # type: List[str], List[GoldParse]
nlp.update(texts,
manual_annotations,
drop=0.5,
losses=losses,
sgd=optimizer)
if batch_id % 10000 == 0:
spacy_evaluate(model=nlp,
dev=content_to_rate_test,
print_diff=print_diff)
except Exception as e:
print(f"got exception [{e}] on batch id {batch_id}")
print(f"Epoch {epoch + 1}\nLoss: {losses}\n")
spacy_evaluate(model=nlp,
dev=content_to_rate_test,
print_diff=print_diff)
if output_model is not None:
nlp.to_disk(output_model)
