mochodek/comment-bert-subject

The model is a port of our CommentBERT model from the paper:

@inproceedings{ochodek2022automated,
  title={Automated code review comment classification to improve modern code reviews},
  author={Ochodek, Miroslaw and Staron, Miroslaw and Meding, Wilhelm and S{\"o}der, Ola},
  booktitle={International Conference on Software Quality},
  pages={23--40},
  year={2022},
  organization={Springer}
}

The original model was implemented in Keras with two outputs - comment-purpose and subject-purpose. Here, we divided it into two separate model with one output each.

---

license: apache-2.0

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import numpy as np

def sigmoid(x):
    return 1/(1 + np.exp(-x))

classes = [
 'code_design',
 'code_style',
 'code_naming',
 'code_logic',
 'code_io',
 'code_data',
 'code_doc',
 'code_api',
 'compatibility',
 'rule_def',
 'config_commit_patch_review',
 'config_building_installing',
]
class2id = {class_:id for id, class_ in enumerate(classes)}
id2class = {id:class_ for class_, id in class2id.items()}

checkpoint = 'mochodek/bert4comment-subject'
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint)

text = "What do you think about making this constant?"
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)

logits = output.logits.detach().numpy()

scores = sigmoid(logits)
scores = (scores > 0.5).astype(int).reshape(-1)
scores_labels = [class_name for class_name in classes if scores[class2id[class_name]] == 1 ]