CodeT5-base for Code Summarization
CodeT5-base model fine-tuned on CodeSearchNet data in a multi-lingual training setting ( Ruby/JavaScript/Go/Python/Java/PHP) for code summarization. It was introduced in this EMNLP 2021 paper CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Understanding and Generation by Yue Wang, Weishi Wang, Shafiq Joty, Steven C.H. Hoi. Please check out more at this repository.
How to use
Here is how to use this model:
from transformers import RobertaTokenizer, T5ForConditionalGeneration
if __name__ == '__main__':
tokenizer = RobertaTokenizer.from_pretrained('Salesforce/codet5-base-multi-sum')
model = T5ForConditionalGeneration.from_pretrained('Salesforce/codet5-base-multi-sum')
text = """def svg_to_image(string, size=None):
if isinstance(string, unicode):
string = string.encode('utf-8')
renderer = QtSvg.QSvgRenderer(QtCore.QByteArray(string))
if not renderer.isValid():
raise ValueError('Invalid SVG data.')
if size is None:
size = renderer.defaultSize()
image = QtGui.QImage(size, QtGui.QImage.Format_ARGB32)
painter = QtGui.QPainter(image)
renderer.render(painter)
return image"""
input_ids = tokenizer(text, return_tensors="pt").input_ids
generated_ids = model.generate(input_ids, max_length=20)
print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))
# this prints: "Convert a SVG string to a QImage."
Fine-tuning data
We employ the filtered version of CodeSearchNet data [Husain et al., 2019] from CodeXGLUE benchmark for fine-tuning on code summarization. The data is tokenized with our pre-trained code-specific BPE (Byte-Pair Encoding) tokenizer. One can prepare text (or code) for the model using RobertaTokenizer with the vocab files from codet5-base.
Data statistic
Programming Language | Training | Dev | Test |
---|---|---|---|
Python | 251,820 | 13,914 | 14,918 |
PHP | 241,241 | 12,982 | 14,014 |
Go | 167,288 | 7,325 | 8,122 |
Java | 164,923 | 5,183 | 10,955 |
JavaScript | 58,025 | 3,885 | 3,291 |
Ruby | 24,927 | 1,400 | 1,261 |
Training procedure
We fine-tune codet5-base on these six programming languages (Ruby/JavaScript/Go/Python/Java/PHP) in the multi-task learning setting. We employ the balanced sampling to avoid biasing towards high-resource tasks. Please refer to the paper for more details.
Evaluation results
Unlike the paper allowing to select different best checkpoints for different programming languages (PLs), here we employ one checkpoint for all PLs. Besides, we remove the task control prefix to specify the PL in training and inference. The results on the test set are shown as below:
Model | Ruby | Javascript | Go | Python | Java | PHP | Overall |
---|---|---|---|---|---|---|---|
Seq2Seq | 9.64 | 10.21 | 13.98 | 15.93 | 15.09 | 21.08 | 14.32 |
Transformer | 11.18 | 11.59 | 16.38 | 15.81 | 16.26 | 22.12 | 15.56 |
RoBERTa | 11.17 | 11.90 | 17.72 | 18.14 | 16.47 | 24.02 | 16.57 |
CodeBERT | 12.16 | 14.90 | 18.07 | 19.06 | 17.65 | 25.16 | 17.83 |
PLBART | 14.11 | 15.56 | 18.91 | 19.30 | 18.45 | 23.58 | 18.32 |
CodeT5-small | 14.87 | 15.32 | 19.25 | 20.04 | 19.92 | 25.46 | 19.14 |
CodeT5-base | 15.24 | 16.16 | 19.56 | 20.01 | 20.31 | 26.03 | 19.55 |
CodeT5-base-multi-sum | 15.24 | 16.18 | 19.95 | 20.42 | 20.26 | 26.10 | 19.69 |
Citation
@inproceedings{
wang2021codet5,
title={CodeT5: Identifier-aware Unified Pre-trained Encoder-Decoder Models for Code Understanding and Generation},
author={Yue Wang, Weishi Wang, Shafiq Joty, Steven C.H. Hoi},
booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021},
year={2021},
}
- Downloads last month
- 30,820