This is an official implementation of the model described in:
Uri Alon, Shaked Brody, Omer Levy and Eran Yahav, "code2seq: Generating Sequences from Structured Representations of Code" [PDF]
to appear in ICLR'2019
An online demo is available at https://code2seq.org.
This is a TensorFlow implementation of the network, with Java and C# extractors for preprocessing the input code. 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.
- Requirements
- Quickstart
- Configuration
- Releasing a trained mode
- Extending to other languages
- Datasets
- Citation
python3 -c 'import tensorflow as tf; print(tf.__version__)'
- For creating a new Java dataset or manually examining a trained model (any operation that requires parsing of a new code example): JDK
- For creating a C# dataset: dotnet-core version 2.2 or newer.
git clone https://github.com/tech-srl/code2seq
cd code2seq
To obtain a preprocessed dataset to train a network on, you can either download our preprocessed dataset, or create a new dataset from Java source files.
Download our preprocessed dataset Java-large dataset (~16M examples, compressed: 11G, extracted 125GB)
mkdir data
cd data
wget https://s3.amazonaws.com/code2seq/datasets/java-large-preprocessed.tar.gz
tar -xvzf java-large-preprocessed.tar.gz
This will create a data/java-large/ sub-directory, containing the files that hold training, test and validation sets,
and a dict file for various dataset properties.
To create and preprocess a new dataset (for example, to compare code2seq 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:
bash 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 52 epochs on the data that was preprocessed in the previous step. This model is the same model that was used in the paper and the same model that serves the demo at code2seq.org.
wget https://s3.amazonaws.com/code2seq/model/java-large/java-large-model.tar.gz
tar -xvzf java-large-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.
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 "java-large" 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:
bash train.sh
After config.PATIENCE iterations of no improvement on the validation set, training stops by itself.
Suppose that iteration #52 is our chosen model, run:
python3 code2seq.py --load models/java-large-model/model_iter52.release --test data/java-large/java-large.test.c2s
While evaluating, a file named "log.txt" is written to the same dir as the saved models, with each test example name and the model's prediction.
To manually examine a trained model, run:
python3 code2seq.py --load models/java-large-model/model_iter52.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.
Due to TensorFlow's limitations, if using beam search (config.BEAM_WIDTH > 0), then BEAM_WIDTH hypotheses will be printed, but
without attention weights. If not using beam search (config.BEAM_WIDTH == 0), then a single hypothesis will be printed with
the attention weights in every decoding timestep.
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.
The frequency, in epochs, of saving a model and evaluating on the validation set during training.
Controlling early stopping: how many epochs of no improvement should training continue before stopping.
Batch size during training.
Batch size during evaluation. Affects only the evaluation speed and memory consumption, does not affect the results.
The buffer size that the reader uses for shuffling the training data. Controls the randomness of the data. Increasing this value might hurt training throughput.
The buffer size (in bytes) of the CSV dataset reader.
The number of contexts to sample in each example during training (resampling a different subset of this size every training iteration).
The max size of the subtoken vocabulary.
The max size of the target words vocabulary.
Embedding size for subtokens, AST nodes and target symbols.
The total size of the two LSTMs that are used to embed the paths if config.BIRNN is True, or the size of the single LSTM if config.BIRNN is False.
Size of each LSTM layer in the decoder.
Number of decoder LSTM layers. Can be increased to support long target sequences.
The max number of nodes in a path
The max number of subtokens in an input token. If the token is longer, only the first subtokens will be read.
The max number of symbols in the target sequence. Set to 6 by default for method names, but can be increased for learning datasets with longer sequences.
If True, use a bidirectional LSTM to encode each path. If False, use a unidirectional LSTM only.
When True, sample MAX_CONTEXT from every example every training iteration.
When False, take the first MAX_CONTEXTS only.
Beam width in beam search. Inactive when 0.
If True, use Momentum optimizer with nesterov. If False, use Adam
(Adam converges in fewer epochs; Momentum leads to slightly better results).
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 code2seq.py --load models/java-large-model/model_iter52 --release
This will save a copy of the trained model with the '.release' suffix. A "released" model usually takes ~3x less disk space.
To extend code2seq to other languages other than Java and C#, 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 field is the target label, internally delimited by the "|" character (for example:
compare|ignore|case - Each of the following field are contexts, where each context has three components separated by commas (","). None of these components can 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.
Each "token" component is a token in the code, split to subtokens using the "|" character.
Each path is a path between two tokens, split to path nodes (or other kinds of building blocks) using the "|" character. Example for a context:
my|key,StringExression|MethodCall|Name,get|value
Here my|key and get|value are tokens, and StringExression|MethodCall|Name is the syntactic path that connects them.
To download the Java-small, Java-med and Java-large datasets used in the Code Summarization task as raw *.java files, use:
The C# dataset used in the Code Captioning task can be downloaded from the CodeNN repository.
code2seq: Generating Sequences from Structured Representations of Code
@inproceedings{
alon2018codeseq,
title={code2seq: Generating Sequences from Structured Representations of Code},
author={Uri Alon and Shaked Brody and Omer Levy and Eran Yahav},
booktitle={International Conference on Learning Representations},
year={2019},
url={https://openreview.net/forum?id=H1gKYo09tX},
}