Source code for "Exploring Dynamic Selection of Branch Expansion Orders for Code Generation".
The source code is developed upon TranX, for technical details please refer to ACL '18 paper and EMNLP '18 demo paper.
File Structure is mainly composed of two components:
-
asdl
defines a general-purpose transition system based on the ASDL formalism, and its instantiations in different programming languages and datasets. The transition system defines how an AST is constructed using a sequence of actions. This package can be used as a standalone library independent of tranX. -
model
contains the neural network implementation of the transition system defined inasdl
, which computes action probabilities using neural networks.
Here is a detailed map of the file strcuture:
├── asdl (grammar-based transition system)
├── datasets (dataset specific code like data preprocessing/evaluation/etc.)
├── model (PyTorch implementation of neural nets)
├── components (helper functions and classes like vocabulary)
cd tranX
bash ./pull_data.sh # get datasets from the TranX
sh scripts/django/train-pre.sh # pre-train model on the django dataset
sh scripts/django/train.sh # train model on the django dataset
At last, it will show the accuracy on test dataset.
We provide the trained model in checkpoint
for convenience.
Here is a list of performance results on four datasets using trained models in checkpoint
Dataset | Results | Metric |
---|---|---|
Django | 78.22 | Accuracy |
CoNaLa | 26.42 | Corpus BLEU |
ATIS | 89.95 | Accuracy |
GEO | 90.35 | Accuracy |
You can run sh scripts/<lang>/test.sh
to get these results.
Please note that
Django dataset only support Python 2.7.
The main example conda environment (config/env/tranx.yml
) supports Python 3, and
environment (config/env/tranx-py2.yml
) supports Python 2.
You can export the enviroments using the following command:
conda env create -f config/env/(tranx.yml,tranx-py2.yml)
Most of this code and data are borrowed from:
https://github.com/pcyin/tranX