Code Generation as a Dual Task of Code Summarization

Ad-hoc implementation of the CS/CG model proposed by Wei et al.

Getting started

• Each dataset must be defined as a sub-class of torch.utils.data.Dataset, with methods for
  • preprocessing and vocab builder (text -> vocab look-up indices)
  • __getitem__ which must return a training example
  • __len__
  • generating train/test/valid splits
  • computing language model probabilites (i.e. P(x), where x: anno/code tensor)

Computing LM probabilities

• Get train/test/valid splits for a dataset.
• Construct a configuration for the LM.
• For each kind (anno/code), train a LM and dump the model as lm-{dataset_name}-{kind}.pt (e.g. lm-django-anno.pt).
• Finally, using these models, compute P(x) for each x (anno/code tensor).

Reference

@article{wei2019code,
  title={Code Generation as a Dual Task of Code Summarization},
  author={Wei, Bolin and Li, Ge and Xia, Xin and Fu, Zhiyi and Jin, Zhi},
  journal={arXiv preprint arXiv:1910.05923},
  year={2019}
}