vegarab / code2seq-reproducibility-challenge Goto Github PK

Following the model descriptions in the original paper and the implementation in the code2seq repository, implement an equivalent model in pytorch for testing reproducibility.

Look into different wrappers for pytorch, such as torchbearer. The final trainable model will require; AST embeddings, an encoder, an MLP, and a decoder with an attention head.

For the experiment setup, we will require functionality for saving and loading the model to file. This needs to support the given frameworks for implementing the model (likely pytorch with torchbearer).

Necessarily functions for saving weights to file, loading weights from file, global configuration file with model-size and other parameters. The original code2seq repository should be used as a basis. Different ways of evaluating a trained model need to be investigated. These should be equivalent to those used in the original paper (F1, BLEU-score).

Using the code2seq repository as a basis, implement/port an AST pre-processing pipeline. This will probably entail the forking of multiple files and functions from the original code2seq project.

All outputted data needs to be compatible with this project's model, which is likely to be implemented using pytorch. As the original code2seq implementation utilizes numpy arrays, we will probably need a translation step to cast the structures into pytorch Tensors.