Rooba
/
spaCy
mirror of https://github.com/explosion/spaCy.git
1
1
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix concat bp and more efficient batch calls

This commit is contained in:
svlandeg 2019-05-21 13:43:59 +02:00
parent 0a15ee4541
commit 2fa3fac851
2 changed files with 76 additions and 89 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

163
examples/pipeline/wiki_entity_linking/train_el.py
View File

@ -52,27 +52,25 @@ class EL_Model:
# raise errors instead of runtime warnings in case of int/float overflow
np.seterr(all='raise')
Doc.set_extension("entity_id", default=None)
train_inst, train_pos, train_neg, train_texts = self._get_training_data(training_dir,
entity_descr_output,
False,
trainlimit,
to_print=False)
train_inst, train_pos, train_neg, train_doc = self._get_training_data(training_dir,
entity_descr_output,
False,
trainlimit,
to_print=False)
dev_inst, dev_pos, dev_neg, dev_doc = self._get_training_data(training_dir,
entity_descr_output,
True,
devlimit,
to_print=False)
dev_inst, dev_pos, dev_neg, dev_texts = self._get_training_data(training_dir,
entity_descr_output,
True,
devlimit,
to_print=False)
self._begin_training()
print()
self._test_dev(train_inst, train_pos, train_neg, train_doc, print_string="train_random", calc_random=True)
self._test_dev(dev_inst, dev_pos, dev_neg, dev_doc, print_string="dev_random", calc_random=True)
self._test_dev(train_inst, train_pos, train_neg, train_texts, print_string="train_random", calc_random=True)
self._test_dev(dev_inst, dev_pos, dev_neg, dev_texts, print_string="dev_random", calc_random=True)
print()
self._test_dev(train_inst, train_pos, train_neg, train_doc, print_string="train_pre", avg=False)
self._test_dev(dev_inst, dev_pos, dev_neg, dev_doc, print_string="dev_pre", avg=False)
self._test_dev(train_inst, train_pos, train_neg, train_texts, print_string="train_pre", avg=False)
self._test_dev(dev_inst, dev_pos, dev_neg, dev_texts, print_string="dev_pre", avg=False)
instance_pos_count = 0
instance_neg_count = 0
@ -99,26 +97,22 @@ class EL_Model:
# print()
# print(article_count, "Training on article", article_id)
article_count += 1
article_docs = list()
article_text = train_texts[article_id]
entities = list()
golds = list()
for inst_cluster in inst_cluster_set:
article_docs.append(train_doc[article_id])
entities.append(train_pos.get(inst_cluster))
golds.append(float(1.0))
instance_pos_count += 1
for neg_entity in train_neg.get(inst_cluster, []):
article_docs.append(train_doc[article_id])
entities.append(neg_entity)
golds.append(float(0.0))
instance_neg_count += 1
self.update(article_docs=article_docs, entities=entities, golds=golds)
self.update(article_text=article_text, entities=entities, golds=golds)
# dev eval
# self._test_dev(dev_inst, dev_pos, dev_neg, dev_doc, print_string="dev_inter", avg=False)
self._test_dev(dev_inst, dev_pos, dev_neg, dev_doc, print_string="dev_inter_avg", avg=True)
print()
self._test_dev(dev_inst, dev_pos, dev_neg, dev_texts, print_string="dev_inter_avg", avg=True)
except ValueError as e:
print("Error in article id", article_id)
@ -127,13 +121,9 @@ class EL_Model:
print("Trained on", instance_pos_count, "/", instance_neg_count, "instances pos/neg")
if self.PRINT_TRAIN:
# print()
# self._test_dev(train_inst, train_pos, train_neg, train_doc, print_string="train_post", avg=False)
self._test_dev(train_inst, train_pos, train_neg, train_doc, print_string="train_post_avg", avg=True)
# self._test_dev(dev_inst, dev_pos, dev_neg, dev_doc, print_string="dev_post", avg=False)
# self._test_dev(dev_inst, dev_pos, dev_neg, dev_doc, print_string="dev_post_avg", avg=True)
self._test_dev(train_inst, train_pos, train_neg, train_texts, print_string="train_post_avg", avg=True)
def _test_dev(self, instances, pos, neg, doc, print_string, avg=False, calc_random=False):
def _test_dev(self, instances, pos, neg, texts_by_id, print_string, avg=False, calc_random=False):
predictions = list()
golds = list()
@ -144,22 +134,18 @@ class EL_Model:
article = inst_cluster.split(sep="_")[0]
entity_id = inst_cluster.split(sep="_")[1]
article_doc = doc[article]
article_doc = self.nlp(texts_by_id[article])
entities = [self.nlp(pos_ex)]
golds.append(float(1.0))
for neg_ex in neg_exs:
entities.append(self.nlp(neg_ex))
golds.append(float(0.0))
if calc_random:
prediction = self._predict_random(entity=pos_ex)
preds = self._predict_random(entities=entities)
else:
prediction = self._predict(article_doc=article_doc, entity=pos_ex, avg=avg)
predictions.append(prediction)
golds.append(float(1.0))
for neg_ex in neg_exs:
if calc_random:
prediction = self._predict_random(entity=neg_ex)
else:
prediction = self._predict(article_doc=article_doc, entity=neg_ex, avg=avg)
predictions.append(prediction)
golds.append(float(0.0))
preds = self._predict(article_doc=article_doc, entities=entities, avg=avg)
predictions.extend(preds)
# TODO: combine with prior probability
p, r, f = run_el.evaluate(predictions, golds, to_print=False)
@ -172,39 +158,38 @@ class EL_Model:
return loss, p, r, f
def _predict(self, article_doc, entity, avg=False, apply_threshold=True):
def _predict(self, article_doc, entities, avg=False, apply_threshold=True):
if avg:
with self.article_encoder.use_params(self.sgd_article.averages) \
and self.entity_encoder.use_params(self.sgd_entity.averages):
doc_encoding = self.article_encoder([article_doc])[0]
entity_encoding = self.entity_encoder([entity])[0]
entity_encodings = self.entity_encoder(entities)
else:
doc_encoding = self.article_encoder([article_doc])[0]
entity_encoding = self.entity_encoder([entity])[0]
entity_encodings = self.entity_encoder(entities)
concat_encoding = list(entity_encoding) + list(doc_encoding)
np_array = np.asarray([concat_encoding])
concat_encodings = [list(entity_encodings[i]) + list(doc_encoding) for i in range(len(entities))]
np_array_list = np.asarray(concat_encodings)
if avg:
with self.model.use_params(self.sgd.averages):
prediction = self.model(np_array)
with self.model.use_params(self.sgd.averages):
predictions = self.model(np_array_list)
else:
prediction = self.model(np_array)
predictions = self.model(np_array_list)
if not apply_threshold:
return float(prediction)
if prediction > self.CUTOFF:
return float(1.0)
return float(0.0)
predictions = self.model.ops.flatten(predictions)
predictions = [float(p) for p in predictions]
if apply_threshold:
predictions = [float(1.0) if p > self.CUTOFF else float(0.0) for p in predictions]
def _predict_random(self, entity, apply_threshold=True):
r = random.uniform(0, 1)
return predictions
def _predict_random(self, entities, apply_threshold=True):
if not apply_threshold:
return r
if r > self.CUTOFF:
return float(1.0)
return float(0.0)
return [float(random.uniform(0,1)) for e in entities]
else:
return [float(1.0) if random.uniform(0,1) > self.CUTOFF else float(0.0) for e in entities]
def _build_cnn(self, hidden_entity_width, hidden_article_width):
with Model.define_operators({">>": chain, "|": concatenate, "**": clone}):
@ -252,20 +237,27 @@ class EL_Model:
loss = (d_scores ** 2).sum()
return loss, d_scores
def update(self, article_docs, entities, golds, apply_threshold=True):
doc_encodings, bp_doc = self.article_encoder.begin_update(article_docs, drop=self.DROP)
# print("doc_encodings", len(doc_encodings), doc_encodings)
# TODO: multiple docs/articles
def update(self, article_text, entities, golds, apply_threshold=True):
article_doc = self.nlp(article_text)
doc_encodings, bp_doc = self.article_encoder.begin_update([article_doc], drop=self.DROP)
doc_encoding = doc_encodings[0]
entity_encodings, bp_entity = self.entity_encoder.begin_update(entities, drop=self.DROP)
entity_docs = list(self.nlp.pipe(entities))
# print("entity_docs", type(entity_docs))
entity_encodings, bp_entity = self.entity_encoder.begin_update(entity_docs, drop=self.DROP)
# print("entity_encodings", len(entity_encodings), entity_encodings)
concat_encodings = [list(entity_encodings[i]) + list(doc_encodings[i]) for i in range(len(entities))]
concat_encodings = [list(entity_encodings[i]) + list(doc_encoding) for i in range(len(entities))]
# print("concat_encodings", len(concat_encodings), concat_encodings)
predictions, bp_model = self.model.begin_update(np.asarray(concat_encodings), drop=self.DROP)
predictions = self.model.ops.flatten(predictions)
# print("predictions", predictions)
golds = self.model.ops.asarray(golds)
# print("golds", golds)
loss, d_scores = self.get_loss(predictions, golds)
@ -275,7 +267,7 @@ class EL_Model:
if self.PRINT_F and self.PRINT_TRAIN:
predictions_f = [x for x in predictions]
if apply_threshold:
predictions_f = [1.0 if x > self.CUTOFF else 0.0 for x in predictions_f]
predictions_f = [float(1.0) if x > self.CUTOFF else float(0.0) for x in predictions_f]
p, r, f = run_el.evaluate(predictions_f, golds, to_print=False)
print("p/r/F train", round(p, 1), round(r, 1), round(f, 1))
@ -286,17 +278,17 @@ class EL_Model:
model_gradient = bp_model(d_scores, sgd=self.sgd)
# print("model_gradient", model_gradient)
doc_gradient = list()
entity_gradient = list()
# concat = entity + doc, but doc is the same within this function (TODO: multiple docs/articles)
doc_gradient = model_gradient[0][self.ENTITY_WIDTH:]
entity_gradients = list()
for x in model_gradient:
doc_gradient.append(list(x[0:self.ARTICLE_WIDTH]))
entity_gradient.append(list(x[self.ARTICLE_WIDTH:]))
entity_gradients.append(list(x[0:self.ENTITY_WIDTH]))
# print("doc_gradient", doc_gradient)
# print("entity_gradient", entity_gradient)
# print("entity_gradients", entity_gradients)
bp_doc(doc_gradient, sgd=self.sgd_article)
bp_entity(entity_gradient, sgd=self.sgd_entity)
bp_doc([doc_gradient], sgd=self.sgd_article)
bp_entity(entity_gradients, sgd=self.sgd_entity)
def _get_training_data(self, training_dir, entity_descr_output, dev, limit, to_print):
id_to_descr = kb_creator._get_id_to_description(entity_descr_output)
@ -305,9 +297,9 @@ class EL_Model:
collect_correct=True,
collect_incorrect=True)
instance_by_doc = dict()
instance_by_article = dict()
local_vectors = list() # TODO: local vectors
doc_by_article = dict()
text_by_article = dict()
pos_entities = dict()
neg_entities = dict()
@ -319,33 +311,28 @@ class EL_Model:
if cnt % 500 == 0 and to_print:
print(datetime.datetime.now(), "processed", cnt, "files in the training dataset")
cnt += 1
if article_id not in doc_by_article:
if article_id not in text_by_article:
with open(os.path.join(training_dir, f), mode="r", encoding='utf8') as file:
text = file.read()
doc = self.nlp(text)
doc_by_article[article_id] = doc
instance_by_doc[article_id] = set()
text_by_article[article_id] = text
instance_by_article[article_id] = set()
for mention, entity_pos in correct_entries[article_id].items():
descr = id_to_descr.get(entity_pos)
if descr:
instance_by_doc[article_id].add(article_id + "_" + mention)
doc_descr = self.nlp(descr)
doc_descr._.entity_id = entity_pos
pos_entities[article_id + "_" + mention] = doc_descr
instance_by_article[article_id].add(article_id + "_" + mention)
pos_entities[article_id + "_" + mention] = descr
for mention, entity_negs in incorrect_entries[article_id].items():
for entity_neg in entity_negs:
descr = id_to_descr.get(entity_neg)
if descr:
doc_descr = self.nlp(descr)
doc_descr._.entity_id = entity_neg
descr_list = neg_entities.get(article_id + "_" + mention, [])
descr_list.append(doc_descr)
descr_list.append(descr)
neg_entities[article_id + "_" + mention] = descr_list
if to_print:
print()
print("Processed", cnt, "training articles, dev=" + str(dev))
print()
return instance_by_doc, pos_entities, neg_entities, doc_by_article
return instance_by_article, pos_entities, neg_entities, text_by_article

2
examples/pipeline/wiki_entity_linking/wiki_nel_pipeline.py
View File

@ -111,7 +111,7 @@ if __name__ == "__main__":
print("STEP 6: training", datetime.datetime.now())
my_nlp = spacy.load('en_core_web_md')
trainer = EL_Model(kb=my_kb, nlp=my_nlp)
trainer.train_model(training_dir=TRAINING_DIR, entity_descr_output=ENTITY_DESCR, trainlimit=1000, devlimit=200)
trainer.train_model(training_dir=TRAINING_DIR, entity_descr_output=ENTITY_DESCR, trainlimit=10, devlimit=10)
print()
# STEP 7: apply the EL algorithm on the dev dataset