Skip to content

Latest commit

 

History

History
104 lines (83 loc) · 3.44 KB

README.md

File metadata and controls

104 lines (83 loc) · 3.44 KB
Raw

MuCoS

WIP

This repo provides the code for reproducing the experiments in MuCoS: Is a Single Model Enough? MuCoS: A Multi-Model EnsembleLearning for Semantic Code Search. MuCoS is a multi-model ensemble learning architecture for semantic code search.

Dependency

  • pip install torch==1.4.0
  • pip install transformers==2.5.0
  • pip install javalang nltk numpy h5py # data process
  • pip install filelock more_itertools tensorboardX scikit-learn matplotlib # run MuCoS

Code Search

Data Generation

An example CodeSearchNet data for the process is csn.pkl(which is available in data/). For the full data, you should parse the data in CodeSearchNet into similar format. Then we generate adversarial data using generate_adversarial_datasets.py, this file call a jar package in https://github.com/mdrafiqulrabin/tnpa-framework#1-variable-renaming to process the origin data and do 6 types of program transformation, then save the data to data/output/<transform_type>/<filename>. We use save_dataset.py to read all augmented data and add to origin CodeSearchNet data, then parse the api, tokens and method name, and save all these data to valid.adv_data.pkl or train.adv_data.pkl.

Step 1: generate data

cd src

python generate_adversarial_datasets.py

The output will be the same as in output/ of this repo.

Step 2: parse and save augmented and origin data

all augmented data into origin data, and parse origin and augmented data, save to a new pickle file

python save_dataset.py --data_split train
python save_dataset.py --data_split val
python save_dataset.py --data_split test

Step 3: run scripts to build different dataset for each baseline

python transform2codebert.py --data_split train
python transform2codebert.py --data_split val
python transform2codebert.py --data_split test

Train Individual Model

Fine-Tune

lang=java #fine-tuning a language-specific model for each programming language
pretrained_model=microsoft/codebert-base  #Roberta: roberta-base

python ensemble_train.py \
--model_type roberta \
--model_name_or_path microsoft/codebert-base \
--task_name codesearch \
--do_predict \
--output_dir ../data/ \
--train_file train_origin.txt \
--dev_file valid_origin.txt \
--data_dir ../data/ \
--max_seq_length 200 \
--pred_modelA_dir ./models_api/java/checkpoint-best/ \
--pred_modelB_dir ./models_var/java/checkpoint-best/ \
--pred_modelC_dir ./models_struct/java/checkpoint-best/ \
--per_gpu_train_batch_size 32 \
--per_gpu_eval_batch_size 32 \
--learning_rate 1e-5 \
--num_train_epochs 8 \
--pred_model_dir ./models_ensemble/$lang/checkpoint-best/

Inference

lang=java #programming language
idx=0 #test batch idx

python ensemble_train.py \
--model_type roberta \
--model_name_or_path microsoft/codebert-base \
--task_name codesearch \
--do_predict \
--output_dir ../data/codesearch/test/$lang \
--data_dir ../data/codesearch/test/$lang \
--max_seq_length 200 \
--per_gpu_train_batch_size 32 \
--per_gpu_eval_batch_size 32 \
--learning_rate 1e-5 \
--pred_modelA_dir ./models_api/java/checkpoint-best/ \
--pred_modelB_dir ./models_var/java/checkpoint-best/ \
--pred_modelC_dir ./models_struct/java/checkpoint-best/ \
--num_train_epochs 8 \
--test_file batch_${idx}.txt \
--pred_model_dir ./models_ensemble/$lang/checkpoint-best/ \
--test_result_dir ./results_ensemble/$lang/${idx}_batch_result.txt

Evaluation

python mrr.py