This is a Tensorflow Implementation for "InferCode: Self-Supervised Learning of Code Representations by Predicting Subtrees" (ICSE'21). InferCode works based on the key idea of using an encoder to predict subtrees as a pretext task. Then the weights learned from the encoder can be used to transfer for other downstream tasks. This is to alleviate the need for the huge amount of labeled data to build decent code learning models in Software Engineering. With this concept, representation learning models for source code can now learn from unlabeled data.
We trained our model on a dataset comprised of 19 languages, which are: java, c, c++, c#, golang, javascript, lua, php, python, ruby, rust, scala, kotlin, solidity, haskell, r, html, css, bash. We use tree-sitter as the backbone to parse these languages to AST. This is a bit different from the implementation we reported in our paper, which used srcml as the AST parser. The reasons are that we found that tree-sitter supports more language than srcml, and tree-sitter also provides a python binding interface, which makes it easy to parse any code snippet into AST by using python code. A details of our old implementation using srcml can be found in old_version.
Install the Pypi package (current version is 0.0.28):
pip3
install
infercode
Infercode can be tested/used as a command
infercode <file1>.<ext1> [<file2>.<ext2>...]
where <file>
is a file name, and <ext>
is the file extension. The file extension
will be used to select the programming language for infercode to choose the corresponding
parser. It will generate a numpy vector for each file in the argument.
You can also use infercode
as a python library for more advanced uses:
from infercode.client.infercode_client import InferCodeClient
import os
import logging
logging.basicConfig(level=logging.INFO)
# Change from -1 to 0 to enable GPU
os.environ['CUDA_VISIBLE_DEVICES'] = "-1"
infercode = InferCodeClient(language="c")
infercode.init_from_config()
vectors = infercode.encode(["for (i = 0; i < n; i++)", "struct book{ int num; char s[27]; }shu[1000];"])
print(vectors)
Then we have the output embeddings:
[[ 0.00455336 0.00277071 0.00299444 -0.00264732 0.00424443 0.02380365
0.00802475 0.01927063 0.00889819 0.01684897 0.03249155 0.01853252
0.00930241 0.02532686 0.00152953 0.0027509 0.00200306 -0.00042401
0.00093602 0.044968 -0.0041187 0.00760367 0.01713051 0.0051542
-0.00033204 0.01757674 -0.00852873 0.00510181 0.02680481 0.00579945
0.00298177 0.00650377 0.01903037 0.00188015 0.00644581 0.02502727
-0.00599149 0.00339381 0.01834774 -0.0012807 -0.00413265 0.01172356
0.01524384 0.00769007 0.01364587 -0.00340345 0.02757765 0.03651286
0.01334631 0.01464784]
[-0.00017088 0.01376707 0.01347563 0.00545072 0.01674811 0.01347677
0.01061796 0.02521674 0.01205592 0.03466582 0.01449588 0.02479498
-0.00011303 0.01174722 0.00444653 0.01382409 -0.00396148 -0.00195686
0.00527923 0.03169966 -0.00935379 0.01904526 0.02334653 -0.00742705
0.00405659 0.0158342 -0.00599484 0.01687686 0.03012032 0.01365279
0.01936428 0.00576922 0.01786506 0.00244599 0.00816536 0.03116215
-0.00721357 0.01265837 0.029279 0.00394636 0.00475944 0.0057507
0.02005564 0.00345545 0.01078242 0.00763404 0.01771503 0.02223164
0.01541999 0.03995579]]
Note that on the initial step, the script will build tree-sitter parsers from sources into ~/.tree-sitter/bin
,
download our pretrained model, and store it into ~/.infercode_data/model_checkpoint
.
-
There are a few other techniques for code representation learning, but none of them are designed with the intention to have a pretrained model to convert code to vector. For example, Code2vec (Alon et al.), despite the attractive name, Code2vec is not suitable to convert code to vector since they trained the model to predict the method name. If one wants to reuse the Code2vec model to convert code to vector, their implementation is not ready for this purpose.
-
There are also other pretrained models for code, such as CodeBert, GraphCodeBert, CuBert, etc, but they did not wrap their code into usable inferfaces.
-
None of the above work supports such many languages like InferCode.
If you find this work useful for your research, please consider citing our paper:
@inproceedings{bui2021infercode,
title={InferCode: Self-Supervised Learning of Code Representations by Predicting Subtrees},
author={Bui, Nghi DQ and Yu, Yijun and Jiang, Lingxiao},
booktitle={2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE)},
pages={1186--1197},
year={2021},
organization={IEEE}
}