transformers_sklearn
A Fitting, Scoring and Predicting Toolkit for Transformer Series Models.
transformers_sklearn aims at bringing Transformer series models to non-specialists.
Main requirements
- pandas https://pandas.pydata.org/
- pytorch https://pytorch.org/
- transformers https://github.com/huggingface/transformers
- scikit-learn https://scikit-learn.org/stable/
Architecture
transformers_sklearn wrap the powerful APIs of transformers into three
class. They are BERTologyClassifier for classification task, BERTologyNERClassifier
for name entity recognition(NER) task and BERTologyRegressor for regression task.
Within each class, there are three methods, which are fit for fine-tuning
the pre-trained models downloaded from community, score for scoring the performance of
fine-tuned model, and predict for predicting the labels of given datasets.
fit and score methods accept two parameters, X and y. The type of X and y
could be list, ndarray and DataFrame. predict only need X.
Install
git clone https://github.com/trueto/transformers_sklearn
cd transformers_sklearn
pip install .
Tutorials
Classification tasks
English
Fine-tuning and scoring BERT model on MRPC corpus The whole code is as following:
import pandas as pd
from transformers_sklearn import BERTologyClassifier
if __name__ == '__main__':
## 1. preparing X,y
train_df = pd.read_csv('datasets/mrpc/train.txt',sep='\t',names=['label','id1','id2','s1','s2'])
X_train = pd.concat([train_df['s1'],train_df['s2']],axis=1)
y_train = train_df['label']
test_df = pd.read_csv('datasets/mrpc/test.txt', sep='\t', names=['label', 'id1', 'id2', 's1', 's2'])
X_test = pd.concat([test_df['s1'], test_df['s2']], axis=1)
y_test = test_df['label']
## 2. customize the model
cls = BERTologyClassifier(
model_type='bert',
model_name_or_path='bert-base-cased',
data_dir='ts_data/mrpc',
output_dir='results/mrpc',
num_train_epochs=3,
learning_rate=5e-5
)
## 3. fit
cls.fit(X_train,y_train)
## 4. score
report = cls.score(X_test,y_test)
with open('mrpc.txt','w',encoding='utf8') as f:
f.write(report)Running code above, the following result will be returned:
***** Eval results 50 *****acc = 0.7360406091370558
acc_and_f1 = 0.7810637828293975
f1 = 0.8260869565217391
***** Eval results 100 *****acc = 0.6802030456852792
acc_and_f1 = 0.707748581666169
f1 = 0.7352941176470589
***** Eval results 150 *****acc = 0.8121827411167513
acc_and_f1 = 0.8418552594472646
f1 = 0.8715277777777778
***** Eval results 200 *****acc = 0.817258883248731
acc_and_f1 = 0.8452491599342247
f1 = 0.8732394366197184
***** Eval results 250 *****acc = 0.8096446700507615
acc_and_f1 = 0.8367642588003353
f1 = 0.8638838475499092
***** Eval results 300 *****acc = 0.8121827411167513
acc_and_f1 = 0.840488533678943
f1 = 0.8687943262411348
Chinese
Fine-tuning and scoring BERT model on LCQMC corpus The whole code is as following:
import pandas as pd
from transformers_sklearn import BERTologyClassifier
from sklearn.metrics import classification_report
if __name__ == '__main__':
## 1. preparing X,y
train_df = pd.read_csv('datasets/lcqmc/train.tsv',sep='\t',names=['s1','s2','label'])
X_train = pd.concat([train_df['s1'],train_df['s2']],axis=1)
y_train = train_df['label']
dev_df = pd.read_csv('datasets/lcqmc/dev.tsv', sep='\t', names=['s1', 's2','label'])
X_dev = pd.concat([dev_df['s1'], dev_df['s2']], axis=1)
y_dev = dev_df['label']
test_df = pd.read_csv('datasets/lcqmc/test.tsv', sep='\t', names=['s1', 's2', 'label'])
X_test = pd.concat([test_df['s1'], test_df['s2']], axis=1)
y_test = test_df['label']
## 2. customize the model
cls = BERTologyClassifier(
model_type='bert',
model_name_or_path='bert-base-chinese',
data_dir='ts_data/lcqmc',
output_dir='results/lcqmc',
num_train_epochs=3,
learning_rate=5e-5
)
#
## 3. fit
cls.fit(X_train,y_train)
#
## 4. score
report = cls.score(X_dev,y_dev)
with open('lcqmc.txt','w',encoding='utf8') as f:
f.write(report)
## 5. predict
y_pred = cls.predict(X_test)
test_report = classification_report(y_test,y_pred,digits=4)
with open('lcqmc_test.txt','w',encoding='utf8') as f:
f.write(test_report)Running code above, the following result will be returned:
precision recall f1-score support
0 0.8919 0.8777 0.8848 4400
1 0.8797 0.8937 0.8866 4402
accuracy 0.8857 8802
macro avg 0.8858 0.8857 0.8857 8802
weighted avg 0.8858 0.8857 0.8857 8802
NER tasks
English
Fine-tuning and scoring BERT model on GMBNER corpus The whole code is as following:
import pandas as pd
from sklearn.model_selection import train_test_split
from transformers_sklearn import BERTologyNERClassifer
if __name__ == '__main__':
data_df = pd.read_csv('datasets/gmbner/ner_dataset.csv',encoding="utf8")
data_df.fillna(method="ffill",inplace=True)
value_counts = data_df['Tag'].value_counts()
label_list = list(value_counts.to_dict().keys())
# ## 1. preparing datasets
X = []
y = []
for label, batch_df in data_df.groupby(by='Sentence #',sort=False):
words = batch_df['Word'].tolist()
labels = batch_df['Tag'].tolist()
assert len(words) == len(labels)
X.append(words)
y.append(labels)
X_train,X_test,y_train,y_test = train_test_split(X,y,test_size=0.1,random_state=520)
## 2. customize model
ner = BERTologyNERClassifer(
labels=label_list,
model_type='bert',
model_name_or_path='bert-base-cased',
data_dir='ts_data/gmbner',
output_dir='results/gmbner',
num_train_epochs=3,
learning_rate=5e-5,
logging_steps=50,
save_steps=50,
overwrite_output_dir=True
)
#
## 3. fit
ner.fit(X_train, y_train)
# # # #
## 4. score
report = ner.score(X_test, y_test)
with open('gmbner.txt', 'w', encoding='utf8') as f:
f.write(report)Chinese
Fine-tuning and scoring BERT model on MSRANER corpus The whole code is as following:
import pandas as pd
from transformers_sklearn import BERTologyNERClassifer
tag_dict = {
'o': 'O',
'nr': 'Per',
'ns': 'Adr',
'nt': 'Org'
}
def get_X_y(file_path):
words_list = []
labels_list = []
with open(file_path,'r', encoding='utf8') as f:
for line in f.readlines():
line = line.strip()
words = []
labels = []
for sequence in line.split(' '):
tokens = sequence.split('/')[0]
tag = sequence.split('/')[-1]
tlen = len(tokens)
if tag == 'o':
labels = labels + ['O'] * tlen
else:
labels = labels + ['B-' + tag_dict[tag]] + (tlen - 1) * ['I-' + tag_dict[tag]]
assert len(words) == len(labels)
words_list.append(words)
labels_list.append(labels)
return words_list,labels_list
if __name__ == '__main__':
## 1. preparing X,y
X_train, y_train = get_X_y('datasets/msraner/train.txt')
X_test, y_test = get_X_y('datasets/msraner/test.txt')
## 2. customize model
ner = BERTologyNERClassifer(
model_type='bert',
model_name_or_path='bert-base-chinese',
data_dir='ts_data/msraner',
output_dir='results/msraner',
num_train_epochs=3,
learning_rate=5e-5
)
## 3. fit
ner.fit(X_train, y_train)
## 4. score
report = ner.score(X_test,y_test)
with open('msraner.txt','w',encoding='utf8') as f:
f.write(report)Regression task
Fine-tuning and scoring BERT model on STS-B corpus The whole code is as following:
import pandas as pd
from transformers_sklearn import BERTologyRegressor
if __name__ == '__main__':
## 1. preparing X,y
train_df = pd.read_csv('datasets/sts-b/train.tsv',sep='\t')
X_train = pd.concat([train_df['sentence1'],train_df['sentence2']],axis=1)
y_train = train_df['score']
dev_df = pd.read_csv('datasets/sts-b/dev.tsv', sep='\t')
X_dev = pd.concat([dev_df['sentence1'], dev_df['sentence2']], axis=1)
y_dev = dev_df['score']
test_df = pd.read_csv('datasets/sts-b/test.tsv', sep='\t')
X_test = pd.concat([test_df['sentence1'], test_df['sentence2']], axis=1)
## 2. customize the model
cls = BERTologyRegressor(
model_type='bert',
model_name_or_path='bert-base-cased',
data_dir='ts_data/sts-b',
output_dir='results/sts-b',
num_train_epochs=3,
learning_rate=5e-5
)
#
## 3. fit
cls.fit(X_train,y_train)
#
## 4. score
report = cls.score(X_dev,y_dev)
with open('sts-b.txt','w',encoding='utf8') as f:
f.write(report)
## 5. predict
y_pred = cls.predict(X_test)
temp_df = X_test.copy()
temp_df['score'] = y_pred
temp_df.to_csv('sts-b.tsv',index=False)