xiye17 / OpSynth

Optimal Neural Program Synthesis from Multimodal Specifications

OpSynth

Code for the paper Optimal Neural Program Synthesis from Multimodal Specifications (Findings of EMNLP, 2021).

@inproceedings{ye2021optimal,
  title={Optimal Neural Program Synthesis from Multimodal Specifications},
  author = {Xi Ye, Qiaochu Chen, Isil Dillig, and Greg Durrett},
  booktitle = {Findings of EMNLP},
  year={2021}
}

Requirements

• python==3.8
• pytorch==1.6.0
• JAVA 1.8.0

Code

We've already attached trained checkpoint at checkpoints/streg/streg.enc.src100.field100.bin.

Preprocess data

python -c 'from datasets.streg.make_dataset import make_dataset;make_dataset()'

Run Optimal Synthesis

# <split>: the split (dev,testi, or teste) to evaluate on.
sh scripts/streg/synth.sh checkpoints/streg/streg.src100.field100.bin <split>

Train a Model

If you'd like to train a new ASN model, run the following command. The checkpoints will be stored at checkpoints/streg/

sh scripts/streg/train.sh

Run RobustFill

python -c 'from datasets.streg.make_deepcoder_data import make_exs_vocab;make_exs_vocab()'

sh scripts/streg/test_fill.sh checkpoints/streg/streg.robustfill.ioenc100.src100.field100.bin teste

Train RobustFill

sh scripts/streg/train_fill.sh

Credit

Part of the codes and system design are modified from TranX.