Official code release of our NAACL 2021 work, Unified Pre-training for Program Understanding and Generation .
***** PLBART's performances on the downstream tasks are recorded in this spreadsheet. *****
News • Setup • Pre-training • Fine-tuning • FAQs • Acknowledgement • License • Citation
- PLBART is a sequence-to-sequence model pre-trained on a large collection Java and Python functions and natural language descriptions collected from Github and StackOverflow, respectively.
- PLBART is pre-trained via denoising autoencoding (DAE) and uses three noising strategies: token masking, token deletion, and token infilling (shown below in the three examples).
| Noisy Input | Original Sequence |
|---|---|
| Is 0 the [MASK] Fibonacci [MASK] ? <En> | <En> Is 0 the first Fibonacci number ? |
| public static main ( String args [ ] ) { date = Date ( ) ; System . out . ( String . format ( " Current Date : % tc " , ) ) ; } <java> | <java> public static void main ( String args [ ] ) { Date date = new Date ( ) ; System . out . printf ( String . format ( " Current Date : % tc " , date ) ) ; } |
| def addThreeNumbers ( x , y , z ) : NEW_LINE INDENT return [MASK] <python> | <python> def addThreeNumbers ( x , y , z ) : NEW_LINE INDENT return x + y + z |
- November 2021: Released PLBART-large (12L-1024H-16A) checkpoint
- October 2021: Released PLBART checkpoints pre-trained on CodeSearchNet
- August 2021: Multilingual multi-task learning using PLBART
- July 2021: Released PLBART checkpoints fine-tuned on downstream tasks
- June 2021: Official code release
- March 2021: Pre-release of source code
We can setup a conda environment in order to run PLBART experiments, the first step is to download the dependencies. We assume anaconda is installed. The additional requirements (noted in requirements.txt) can be installed by running the following script:
bash install_env.sh
Go to data/github directory and follow instructions.
Go to data/stackoverflow directory and follow instructions.
cd pretrain
bash binarize.sh
bash pretrain.sh GPU_IDS[Note] We pre-trained PLBART on 8 GeForce RTX 2080 (11gb) GPUs (took ~11.5 days). If you want to pre-train PLBART
using more GPUs or GPUs with more memory, adjust MAX_SENTENCES, MAX_TOKENS, UPDATE_FREQ accordingly to maintain an
effective batch size of 2048. According to fairseq, effective batch size is equal
to:
PER_GPU_TRAIN_BATCH_SIZE * NUM_GPU * UPDATE_FREQ
Note that, MAX_TOKENS refers to the size of each mini-batch, in terms of the number of tokens. During our experiments,
we noticed that in an 11gb GPU, maximum 2048 tokens can be accommodated which is equivalent to 4-5 examples. Therefore,
we set UPDATE_FREQ to 60, so that we can achieve an effective batch size of ~2048.
We fine-tune and evaluate PLBART on three types of downstream tasks.
| Type | Task | Language(s) | Data | Scripts | Checkpoints |
|---|---|---|---|---|---|
| Code to Text | Code summarization | Python, Java, Ruby, PHP, Javascript, Go |
[LINK] | [LINK] | [LINK] |
| Text to Code | Code generation | Java | [LINK] | [LINK] | [LINK] |
| Code to Code | Code translation | Java, C# | [LINK] | [LINK] | [LINK] |
| Code refinement | Java | [LINK] | [LINK] | ||
| Clone detection | Java | [LINK] | [LINK] | ||
| Defect detection | C/C++ | [LINK] | [LINK] |
cd pretrain
bash download.sh
cd ..cd data/codeXglue
bash download.sh
cd ../..cd evaluation/CodeBLEU/parser
bash build.sh
cd ../../..For example, we want to fine-tune PLBART on Text-to-Code task. Then,
cd scripts/text_to_code
bash prepare.sh
bash run.sh GPU_IDS
cd ../..Note. We fine-tuned PLBART on 1 GeForce RTX 2080 (11gb) GPU.
[NOTE] We present the file structure of this repository here .
How to download Github data from Google BigQuery?
We provided a detailed guide here .
Mismatch in performance reported in the paper and achieved using the released checkpoints.
There is a difference between PLBART's performances mentioned in the paper and the performance achieved with the released checkpoints. We noted them here. Note that, there is no change in the hyper-parameter setting. We provided the exact same value we used in the bash scripts. The performance difference we observed is perhaps due to running experiments at different point of time. Although we didn't but we recommend to fine-tune PLBART with multiple different seeds and report the average scores.
mbart_base task is not present in fairseq==0.9.0 official release.
Although we used fairseq==0.9.0 but we used a different commit which consists of mbart_base task. You may do the
following which should work.
git clone https://github.com/pytorch/fairseq
cd fairseq
git checkout 698e3b91ffa832c286c48035bdff78238b0de8ae
pip install .
Otherwise, you may consider installing fairseq==0.10.0. Please refer to
this issue
to make other adjustments.
What can be the maximum input and output lengths for PLBART?
The maximum length is 512.
PLBART uses Fairseq, codeXglue, and TransCoder and thanks the authors of these works for their contribution.
Contents of this repository is under the MIT license. The license applies to the pre-trained and fine-tuned models as well.
@inproceedings{ahmad-etal-2021-unified,
title = "Unified Pre-training for Program Understanding and Generation",
author = "Ahmad, Wasi and
Chakraborty, Saikat and
Ray, Baishakhi and
Chang, Kai-Wei",
booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
month = jun,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2021.naacl-main.211",
pages = "2655--2668"
}