def train_crf(train_path, model_path, _log, _run, dev_path=None):
train_reader = read_corpus(train_path)
_log.info('Extracting features from train corpus')
train_itemseq = ItemSequence([
fs for sent in train_reader.sents()
for fs in extract_crf_features(sent)
])
train_labels = [tag for _, tag in train_reader.tagged_words()]
trainer = make_crf_trainer()
trainer.append(train_itemseq, train_labels, group=0)
if dev_path is not None:
dev_reader = read_corpus(dev_path, name='dev')
_log.info('Extracting features from dev corpus')
dev_itemseq = ItemSequence([
fs for sent in dev_reader.sents()
for fs in extract_crf_features(sent)
])
dev_labels = [tag for _, tag in dev_reader.tagged_words()]
trainer.append(dev_itemseq, dev_labels, group=1)
_log.info('Begin training; saving model to %s', model_path)
holdout = -1 if dev_path is None else 1
trainer.train(model_path, holdout=holdout)
if SACRED_OBSERVE_FILES:
_run.add_artifact(model_path)
def train(model_path, _log, _run, window=2):
"""Train a CRF model."""
train_reader = read_train_corpus()
sents, tags = separate_tagged_sents(train_reader.tagged_sents())
sents = preprocess(sents)
_log.info('Extracting features from train corpus')
train_itemseq = ItemSequence(
[fs for sent in sents for fs in extract_crf_features(sent)])
train_labels = [tag for tags_ in tags for tag in tags_]
trainer = make_crf_trainer()
trainer.append(train_itemseq, train_labels, group=0)
dev_reader = read_dev_corpus()
if dev_reader is not None:
_log.info('Extracting features from dev corpus')
sents, tags = separate_tagged_sents(dev_reader.tagged_sents())
sents = preprocess(sents)
dev_itemseq = ItemSequence(
[fs for sent in sents for fs in extract_crf_features(sent)])
dev_labels = [tag for tags_ in tags for tag in tags_]
trainer.append(dev_itemseq, dev_labels, group=1)
_log.info('Begin training; saving model to %s', model_path)
holdout = -1 if dev_reader is None else 1
trainer.train(model_path, holdout=holdout)
if SACRED_OBSERVE_FILES:
_run.add_artifact(model_path)
def build_model_features(sample, window_size, label_data):
"""
:param sample: 一个DataSample对象
:param window_size:
:param label_data:
:return: sample对应的ItemSequence
"""
features = []
chars = list(sample.sentence)
for index, word in enumerate(chars):
'''
build features
'''
windows_words = [i if i > 0 else -1 for i in range(index - window_size // 2, index)]
windows_words.extend([i if i < len(chars) else -1 for i in range(index, index + window_size // 2 + 1)])
feature = word_feature(sample, windows_words, label_data)
features.append(feature.calWord_features(index))
# print(features)
item_sequence = ItemSequence(features)
return item_sequence
def train(self,
training_data,
classifier_path="classifier/cache/label_crf_classifier",
c1=0,
c2=10,
period=300,
minfreq=5):
self.preprocess(training_data)
train = Trainer()
for i1, i in enumerate(self.x):
train.append(ItemSequence(i), self.y[i1])
params = {
"c1": c1,
"c2": c2,
"period": period,
"feature.minfreq": minfreq,
"max_iterations": 1000
# "calibration.eta": 0.05,
# "calibration_samples": 400,
}
# train.select(algorithm = "l2sgd")
train.set_params(params)
train.train(classifier_path)
self.tagger = Tagger()
self.tagger.open(classifier_path)
def make_preds(tagger, sents, _log):
sents = preprocess(sents)
_log.info('Extracting features')
itemseq = ItemSequence(
[fs for sent in sents for fs in extract_crf_features(sent)])
_log.info('Making predictions with the model')
return tagger.tag(itemseq)
def train(features: pd.Series, labels: pd.Series) -> None:
trainer = Trainer(verbose=False)
features = features.tolist()
labels = labels.tolist()
for idx in range(len(features)):
trainer.append(ItemSequence(features[idx]), literal_eval(labels[idx]))
trainer.train('crf.model')
def train(features, labels):
print("Training..")
trainer = Trainer(verbose=False)
features = features.tolist()
labels = labels.tolist()
for idx in range(0, len(features)):
trainer.append(ItemSequence(features[idx]), literal_eval(labels[idx]))
trainer.train('crf.model')
def predict_crf(reader, model_path, _log, _run):
_log.info('Loading model from %s', model_path)
if SACRED_OBSERVE_FILES:
_run.add_resource(model_path)
tagger = Tagger()
tagger.open(model_path)
_log.info('Extracting features from test corpus')
itemseq = ItemSequence(
[fs for sent in reader.sents() for fs in extract_crf_features(sent)])
_log.info('Making predictions with the model')
return tagger.tag(itemseq)
def predict(self, test_data):
"""Input: x should be a list of strings"""
result = self.tagger.tag(ItemSequence(test_data["crf_feature"]))
test_data["section_label"] = result
return test_data
# def score(self, training_data, classifier_path="classifier/cache/label_crf_classifier", portion=0.8, c1=0, c2=10, period=300, minfreq=10):
# # split resume_id
# resume_ids = np.unique([resume['resume_id'] for resume in training_data])
# length = len(resume_ids)
# shuffle(resume_ids)
# train_ids = resume_ids[:int(length*portion)]
# test_ids = resume_ids[int(length*portion):]
# train_df = [resume for resume in training_data if resume['resume_id'] in train_ids]
# test_df = [resume for resume in training_data if resume['resume_id'] in test_ids]
# # train model on train_ids
# self.train(train_df, classifier_path=classifier_path, c1=c1, c2=c2, period=period, minfreq=minfreq)
# test_pred = self.predict_all(test_df)
# train_pred = self.predict_all(train_df)
# # print out result
# return train_pred, test_pred
# if __name__ == "__main__":
# data = MongoRetriveData()
# resumes = data.get_data_mongo()
# # pickle.dump(resumes, open('./resume_data.pkl', 'wb'))
# # resumes = pickle.load(open('./resume_data.pkl', 'rb'))
# stopword_path = './stopword.txt'
# model_path = './model.txt'
# resume_data = resumes
# clf = Crf(stopword_path, model_path, resume_data)
# clf.CleanData()
# clf.Fit()
# clf.Score()
# # result = clf.Predict(clf.data)
# # print result
def train(self, feat_seqs, label_seqs):
print >>sys.stderr, "Training on %d sequences"%len(feat_seqs)
if self.algorithm == "crf":
for feat_seq, label_seq in zip(feat_seqs, label_seqs):
self.trainer.append(ItemSequence(feat_seq), label_seq)
self.trainer.train(self.trained_model_name)
else:
for fs in feat_seqs:
for feat_dict in fs:
for f in feat_dict:
if f not in self.feat_index:
self.feat_index[f] = len(self.feat_index)
Xs = []
for fs in feat_seqs:
X = []
for feat_dict in fs:
x = [0] * len(self.feat_index)
for f in feat_dict:
x[self.feat_index[f]] = feat_dict[f]
X.append(x)
Xs.append(numpy.asarray(X))
for ls in label_seqs:
for label in ls:
if label not in self.label_index:
self.label_index[label] = len(self.label_index)
Ys = []
for ls in label_seqs:
Y = []
for label in ls:
Y.append(self.label_index[label])
Ys.append(numpy.asarray(Y))
self.trainer.fit(Xs, Ys)
pickle.dump(self.trainer, open(self.trained_model_name, "wb"))
pickle.dump(self.feat_index, open("ssvm_feat_index.pkl", "wb"))
pickle.dump(self.label_index, open("ssvm_label_index.pkl", "wb"))
def predict(self, feat_seqs):
print >>sys.stderr, "Tagging %d sequences"%len(feat_seqs)
if self.algorithm == "crf":
self.tagger.open(self.trained_model_name)
preds = [self.tagger.tag(ItemSequence(feat_seq)) for feat_seq in feat_seqs]
else:
Xs = []
for fs in feat_seqs:
X = []
for feat_dict in fs:
x = [0] * len(self.feat_index)
for f in feat_dict:
if f in self.feat_index:
x[self.feat_index[f]] = feat_dict[f]
X.append(x)
Xs.append(numpy.asarray(X))
pred_ind_seqs = self.tagger.predict(Xs)
preds = []
for ps in pred_ind_seqs:
pred = []
for pred_ind in ps:
pred.append(self.rev_label_index[pred_ind])
preds.append(pred)
return preds
def one_hot_encoding(self,features):
encoding = ItemSequence(features[0])
self.encoding = encoding.items()
return self.encoding
def _get_xseq(model, matrix):
xseq = [{
'feat{}'.format(i): float(w)
for i, w in enumerate(list(features))
} for features in model.predict(matrix)]
return ItemSequence(xseq)
def test_tag_item_sequence(model_filename, xseq, yseq):
with Tagger().open(model_filename) as tagger:
assert tagger.tag(ItemSequence(xseq)) == yseq
def feed_trainer(self, features_list, labels):
features = ItemSequence(features_list)
self.trainer.append(features, labels)
def tag_tweets(self, tweet_features_list):
features = ItemSequence(tweet_features_list)
labels = self.tagger.tag(features)
return labels
def sent2features(sent):
return ItemSequence([word2features(sent, i) for i in range(len(sent))])
