A neural network for learning distributed representations of code. This is an official implementation of the model described in:
Uri Alon, Meital Zilberstein, Omer Levy and Eran Yahav, "code2vec: Learning Distributed Representations of Code", POPL'2019 [PDF]
October 2018 - The paper was accepted to POPL'2019!
April 2019 - The talk video is available here.
July 2019 - Add tf.keras model implementation (see here).
An online demo is available at https://code2vec.org/.
- code2seq (ICLR'2019) is our newer model. It uses LSTMs to encode paths node-by-node (rather than monolithic path embeddings as in code2vec), and an LSTM to decode a target sequence (rather than predicting a single label at a time as in code2vec). See PDF, demo at http://www.code2seq.org and code.
- Structural Language Models of Code is a new paper that learns to generate the missing code within a larger code snippet. This is similar to code completion, but is able to predict complex expressions rather than a single token at a time. See PDF, demo at http://AnyCodeGen.org.
- Adversarial Examples for Models of Code is a new paper that shows how to slightly mutate the input code snippet of code2vec and GNNs models (thus, introducing adversarial examples), such that the model (code2vec or GNNs) will output a prediction of our choice. See PDF (code: soon).
- Neural Reverse Engineering of Stripped Binaries is a new paper that learns to predict procedure names in stripped binaries, thus use neural networks for reverse engineering. See PDF (code: soon).
This is a TensorFlow implementation, designed to be easy and useful in research, and for experimenting with new ideas in machine learning for code tasks. By default, it learns Java source code and predicts Java method names, but it can be easily extended to other languages, since the TensorFlow network is agnostic to the input programming language (see Extending to other languages. Contributions are welcome. This repo actually contains two model implementations. The 1st uses pure TensorFlow and the 2nd uses TensorFlow's Keras (more details).
- Requirements
- Quickstart
- Configuration
- Features
- Extending to other languages
- Additional datasets
- Citation
On Ubuntu:
- Python3 (>=3.6). To check the version:
python3 --version
- TensorFlow - version 2.0.0 (install). To check TensorFlow version:
python3 -c 'import tensorflow as tf; print(tf.__version__)'
- If you are using a GPU, you will need CUDA 10.0 (download) as this is the version that is currently supported by TensorFlow. To check CUDA version:
nvcc --version
- For GPU: cuDNN (>=7.5) (download) To check cuDNN version:
cat /usr/include/cudnn.h | grep CUDNN_MAJOR -A 2
- For creating a new dataset or manually examining a trained model (any operation that requires parsing of a new code example) - Java JDK
git clone https://github.com/tech-srl/code2vec
cd code2vec
In order to have a preprocessed dataset to train a network on, you can either download our preprocessed dataset, or create a new dataset of your own.
wget https://s3.amazonaws.com/code2vec/data/java14m_data.tar.gz
tar -xvzf java14m_data.tar.gz
This will create a data/java14m/ sub-directory, containing the files that hold that training, test and validation sets, and a vocabulary file for various dataset properties.
In order to create and preprocess a new dataset (for example, to compare code2vec to another model on another dataset):
- Edit the file preprocess.sh using the instructions there, pointing it to the correct training, validation and test directories.
- Run the preprocess.sh file:
source preprocess.sh
You can either download an already-trained model, or train a new model using a preprocessed dataset.
We already trained a model for 8 epochs on the data that was preprocessed in the previous step. The number of epochs was chosen using early stopping, as the version that maximized the F1 score on the validation set. This model can be downloaded here or using:
wget https://s3.amazonaws.com/code2vec/model/java14m_model.tar.gz
tar -xvzf java14m_model.tar.gz
This trained model is in a "released" state, which means that we stripped it from its training parameters and can thus be used for inference, but cannot be further trained. If you use this trained model in the next steps, use 'saved_model_iter8.release' instead of 'saved_model_iter8' in every command line example that loads the model such as: '--load models/java14_model/saved_model_iter8'. To read how to release a model, see Releasing the model.
A non-stripped trained model can be obtained here or using:
wget https://s3.amazonaws.com/code2vec/model/java14m_model_trainable.tar.gz
tar -xvzf java14m_model_trainable.tar
This model weights more than twice than the stripped version, and it is recommended only if you wish to continue training a model which is already trained. To continue training this trained model, use the --load flag to load the trained model; the --data flag to point to the new dataset to train on; and the --save flag to provide a new save path.
We provide an additional code2vec model that was trained on the "Java-large" dataset (this dataset was introduced in the code2seq paper). See Java-large
To train a model from scratch:
- Edit the file train.sh to point it to the right preprocessed data. By default, it points to our "java14m" dataset that was preprocessed in the previous step.
- Before training, you can edit the configuration hyper-parameters in the file config.py, as explained in Configuration.
- Run the train.sh script:
source train.sh
- By default, the network is evaluated on the validation set after every training epoch.
- The newest 10 versions are kept (older are deleted automatically). This can be changed, but will be more space consuming.
- By default, the network is training for 20 epochs. These settings can be changed by simply editing the file config.py. Training on a Tesla v100 GPU takes about 50 minutes per epoch. Training on Tesla K80 takes about 4 hours per epoch.
Once the score on the validation set stops improving over time, you can stop the training process (by killing it) and pick the iteration that performed the best on the validation set. Suppose that iteration #8 is our chosen model, run:
python3 code2vec.py --load models/java14_model/saved_model_iter8.release --test data/java14m/java14m.test.c2v
While evaluating, a file named "log.txt" is written with each test example name and the model's prediction.
To manually examine a trained model, run:
python3 code2vec.py --load models/java14_model/saved_model_iter8.release --predict
After the model loads, follow the instructions and edit the file Input.java and enter a Java method or code snippet, and examine the model's predictions and attention scores.
Changing hyper-parameters is possible by editing the file config.py.
Here are some of the parameters and their description:
The max number of epochs to train the model. Stopping earlier must be done manually (kill).
After how many training iterations a model should be saved.
Batch size in training.
Batch size in evaluating. Affects only the evaluation speed and memory consumption, does not affect the results.
Number of words with highest scores in $ y_hat $ to consider during prediction and evaluation.
Number of batches (during training / evaluating) to complete between two progress-logging records.
Number of training batches to complete between model evaluations on the test set.
The number of threads enqueuing examples to the reader queue.
Size of buffer in reader to shuffle example within during training. Bigger buffer allows better randomness, but requires more amount of memory and may harm training throughput.
The buffer size (in bytes) of the CSV dataset reader.
The number of contexts to use in each example.
The max size of the token vocabulary.
The max size of the target words vocabulary.
The max size of the path vocabulary.
Default embedding size to be used for token and path if not specified otherwise.
Embedding size for tokens.
Embedding size for paths.
config.CODE_VECTOR_SIZE = config.PATH_EMBEDDINGS_SIZE + 2 * config.TOKEN_EMBEDDINGS_SIZE
Size of code vectors.
Embedding size for target words.
Keep this number of newest trained versions during training.
Dropout rate used during training.
Whether to treat <OOV> and <PAD> as two different special tokens whenever possible.
Code2vec supports the following features:
This repo comes with two model implementations:
(i) uses pure TensorFlow (written in tensorflow_model.py);
(ii) uses TensorFlow's Keras (written in keras_model.py).
The default implementation used by code2vec.py is the pure TensorFlow.
To explicitly choose the desired implementation to use, specify --framework tensorflow or --framework keras
as an additional argument when executing the script code2vec.py.
Particularly, this argument can be added to each one of the usage examples (of code2vec.py) detailed in this file.
Note that in order to load a trained model (from file), one should use the same implementation used during its training.
If you wish to keep a trained model for inference only (without the ability to continue training it) you can release the model using:
python3 code2vec.py --load models/java14_model/saved_model_iter8 --release
This will save a copy of the trained model with the '.release' suffix. A "released" model usually takes 3x less disk space.
Token and target embeddings are available to download:
[Token vectors] [Method name vectors]
These saved embeddings are saved without subtoken-delimiters ("toLower" is saved as "tolower").
In order to export embeddings from a trained model, use the "--save_w2v" and "--save_t2v" flags:
Exporting the trained token embeddings:
python3 code2vec.py --load models/java14_model/saved_model_iter8.release --save_w2v models/java14_model/tokens.txt
Exporting the trained target (method name) embeddings:
python3 code2vec.py --load models/java14_model/saved_model_iter8.release --save_t2v models/java14_model/targets.txt
This saves the tokens/targets embedding matrices in word2vec format to the specified text file, in which: the first line is: <vocab_size> <dimension> and each of the following lines contains: <word> <float_1> <float_2> ... <float_dimension>
These word2vec files can be manually parsed or easily loaded and inspected using the gensim python package:
python3
>>> from gensim.models import KeyedVectors as word2vec
>>> vectors_text_path = 'models/java14_model/targets.txt' # or: `models/java14_model/tokens.txt'
>>> model = word2vec.load_word2vec_format(vectors_text_path, binary=False)
>>> model.most_similar(positive=['equals', 'to|lower']) # or: 'tolower', if using the downloaded embeddings
>>> model.most_similar(positive=['download', 'send'], negative=['receive'])The above python commands will result in the closest name to both "equals" and "to|lower", which is "equals|ignore|case". Note: In embeddings that were exported manually using the "--save_w2v" or "--save_t2v" flags, the input token and target words are saved using the symbol "|" as a subtokens delimiter ("toLower" is saved as: "to|lower"). In the embeddings that are available to download (which are the same as in the paper), the "|" symbol is not used, thus "toLower" is saved as "tolower".
The flag --export_code_vectors allows to export the code vectors for the given examples.
If used with the --test <TEST_FILE> flag,
a file named <TEST_FILE>.vectors will be saved in the same directory as <TEST_FILE>.
Each row in the saved file is the code vector of the code snipped in the corresponding row in <TEST_FILE>.
If used with the --predict flag, the code vector will be printed to console.
This project currently supports Java and C# as the input languages.
April 2020 - an extension for code2vec that addresses obfuscated Java code was developed by @basedrhys, and is available here: https://github.com/basedrhys/obfuscated-code2vec.
January 2020 - an extractor for predicting TypeScript type annotations for JavaScript input using code2vec was developed by @izosak and Noa Cohen, and is available here: https://github.com/tech-srl/id2vec.
June 2019 - an extractor for C that is compatible with our model was developed by CMU SEI team. - removed by CMU SEI team.
June 2019 - an extractor for Python, Java, C, C++ by JetBrains Research is available here: PathMiner.
In order to extend code2vec to work with other languages, a new extractor (similar to the JavaExtractor) should be implemented, and be called by preprocess.sh. Basically, an extractor should be able to output for each directory containing source files:
- A single text file, where each row is an example.
- Each example is a space-delimited list of fields, where:
- The first "word" is the target label, internally delimited by the "|" character.
- Each of the following words are contexts, where each context has three components separated by commas (","). Each of these components cannot include spaces nor commas. We refer to these three components as a token, a path, and another token, but in general other types of ternary contexts can be considered.
For example, a possible novel Java context extraction for the following code example:
void fooBar() {
System.out.println("Hello World");
}Might be (in a new context extraction algorithm, which is different than ours since it doesn't use paths in the AST):
foo|Bar System,FIELD_ACCESS,out System.out,FIELD_ACCESS,println THE_METHOD,returns,void THE_METHOD,prints,"hello_world"
Consider the first example context "System,FIELD_ACCESS,out". In the current implementation, the 1st ("System") and 3rd ("out") components of a context are taken from the same "tokens" vocabulary, and the 2nd component ("FIELD_ACCESS") is taken from a separate "paths" vocabulary.
We preprocessed additional three datasets used by the code2seq paper, using the code2vec preprocessing. These datasets are available in raw format (i.e., .java files) at https://github.com/tech-srl/code2seq/blob/master/README.md#datasets, and are also available to download in a preprocessed format (i.e., ready to train a code2vec model on) here:
wget https://s3.amazonaws.com/code2vec/data/java-small_data.tar.gz
This dataset is based on the dataset of Allamanis et al. (ICML'2016), with the difference that training/validation/test are split by-project rather than by-file. This dataset contains 9 Java projects for training, 1 for validation and 1 testing. Overall, it contains about 700K examples.
wget https://s3.amazonaws.com/code2vec/data/java-med_data.tar.gz
A dataset of the 1000 top-starred Java projects from GitHub. It contains 800 projects for training, 100 for validation and 100 for testing. Overall, it contains about 4M examples.
wget https://s3.amazonaws.com/code2vec/data/java-large_data.tar.gz
A dataset of the 9500 top-starred Java projects from GitHub that were created since January 2007. It contains 9000 projects for training, 200 for validation and 300 for testing. Overall, it contains about 16M examples.
Additionally, we provide a trained code2vec model that was trained on the Java-large dataset (this model was not part of the original code2vec paper, but was later used as a baseline in the code2seq paper which introduced this dataset). Trainable model (3.5 GB):
wget https://code2vec.s3.amazonaws.com/model/java-large-model.tar.gz
"Released model" (1.4 GB, cannot be further trained).
wget https://code2vec.s3.amazonaws.com/model/java-large-released-model.tar.gz
code2vec: Learning Distributed Representations of Code
@article{alon2019code2vec,
author = {Alon, Uri and Zilberstein, Meital and Levy, Omer and Yahav, Eran},
title = {Code2Vec: Learning Distributed Representations of Code},
journal = {Proc. ACM Program. Lang.},
issue_date = {January 2019},
volume = {3},
number = {POPL},
month = jan,
year = {2019},
issn = {2475-1421},
pages = {40:1--40:29},
articleno = {40},
numpages = {29},
url = {http://doi.acm.org/10.1145/3290353},
doi = {10.1145/3290353},
acmid = {3290353},
publisher = {ACM},
address = {New York, NY, USA},
keywords = {Big Code, Distributed Representations, Machine Learning},
}