ml4se2022 / group7

CodeXGLUE - Code Search (WebQueryTest)

The modified project for testing the relevance of the source code and natural language description from CodeXGLUE competition. The original code is available in a separate GitHub repository.

Prerequisites

Running with Docker

• Task (tested with v3.15.2)
• Docker 20.10 (tested with 20.10.19) with Compose and BuildKit installed
  • Compose and BuildKit are included by default in the latest versions of Docker

Running locally

• Linux or Windows machine with WSL
  • Not tested on macOS
• Python 3.7
  • pyenv can be used to install legacy versions of Python
• Task (tested with v3.15.2)
• (Optional) CUDA toolkit and Nvidia graphics card/accelerator
  • In case no CUDA acceleration is available, PyTorch will default to CPU

Evaluating a pre-trained model

• gdown can be used to download the best performing model of ours to perform evaluations on it.

pip install gdown
gdown https://drive.google.com/uc\?id\=1QSwfPHi83P0-fFaSdTsPhelp3l7V1f9S
unzip best_model.zip -d model_cosqa_continue_training

• If you do not want to use gdown, you can simply download the zip file via this link.
• IMPORTANT: predicting on this model will require a CUDA capable device, as it was trained on a CUDA-capable device.

Launching the learning pipeline

With Docker

Make sure that the Docker daemon is running and build the Docker image with:

task docker:build

This step has to be re-executed whenever changes are made to the source code.

Start the container with:

task docker:up

It will sleep in the background and wait for the commands to be executed with docker:exec, so that its local data are available for extraction with docker:extract. Wxecute arbitrary task by passing its name after --:

# Without access to GPU
task docker:exec -- task code:train
# With access to GPU
task docker:exec-gpu -- task code:train

To display the tasks relevant to the training, execute the following:

task docker:exec-gpu -- task -l

To execute the code:train production training task, execute the following:

task docker:exec -- code:train

Executing commands manually from within the container

In order to enter the container's shell and execute commands, run the following:

# Without access to GPU
task docker:run -- bash
# With access to GPU
task docker:run-gpu -- bash

Copying data from within the container

The data in the container are not persisted when the container is rebuilt. In order to extract it from the container, use the docker:extract task. For example, to extract the trained production models after running task docker:train

task docker:extract -- model_codesearchnet
task docker:extract -- model_cosqa_continue_training

Stopping and removing the container

To stop the container without removing data, execute the following:

task docker:stop

To remove the container and all its data (like the trained models), execute the following:

task docker:down

Without Docker

Ensure that you are running Python 3.7 (check with python -V).

To display the tasks relevant to the training, execute the following:

task -l

To display all the tasks (including intermediate ones, like data processing or virtual environment creation), execute the following:

task --list-all

For instance, to run the production training pipeline, call:

task code:train

The virtual environment will be created automatically. The dataset will also be downloaded and preprocessed.

Dataset

We are using the CoSQA dataset, available in the repository, as well as the CodeSearchNet dataset, which can be downloaded from either:

• executing task task data:download-model
• directly: https://s3.amazonaws.com/code-search-net/CodeSearchNet/v2/python.zip

Expected results

Model configuration	Precision	Recall	F1-score	Accuracy
Baseline model	51%	100%	0.68	51%
Model with deeper FC	55%	1%	0.02	48%
Model with new loss func	61%	2%	0.04	48%
Model with new AF	52%	100%	0.68	52%
Model with AST	56%	14%	0.22	49%
Model with everything combined	0%	0%	0.00	48%

It is important to mention that we obtained these results by fine-tuning the original codebert model on the CoSQA dataset. We achieved this by fine-tuning (training) on only one-tenth of the dataset, as training on more than that would have been unfeasible.

---

Original README

Update 2021.07.30: We add CoSQA into CodeXGLUE, and recommend you to use this dataset as the training and development set of CodeXGLUE -- Code Search (WebQueryTest) Challenge. The dataset can be found in ./cosqa. For more details about the dataset collection and usage, please refer to the ACL 2021 paper and the GitHub repo.

Task Description

Code Search is aimed to find a code snippet which best matches the demand of the query. This task can be formulated in two scenarios: retrieval scenario and text-code classification scenario. In WebQueryTest , we present the Code Search in text-code classification scenario.

In WebQueryTest, a trained model needs to judge whether a code snippet answers a given natural language query, which can be formulated into a binary classification problem.

Most existing code search datasets use code documentations or questions from online communities for software developers as queries, which is still different from real user search queries. Therefore we provide WebQueryTest testing set.

Dependency

• python 3.6 or 3.7
• torch==1.5.0
• transformers>=2.5.0

Data

Here we present WebQueryTest dataset, a testing set of Python code search of 1,046 query-code pairs with code search intent and their human annotations. The realworld user queries are collected from Bing query logs and the code for queries are from CodeSearchNet. You can find our testing set in ./data/test_webquery.json .

Since there's no direct training set for our WebQueryTest dataset, we suggest using two external training sets: 1. CodeSearchNet; 2. CoSQA.

1. To finetune the models on CodeSearchNet, we provide scripts to obtain the documentation-function pairs in the training set o CodeSearchNet AdvTest as positive instances. For each documentation, we also randomly sample 7 more functions to form negative instances. The following command is used to download and preprocess the data:

cd data
wget https://s3.amazonaws.com/code-search-net/CodeSearchNet/v2/python.zip
unzip python.zip
python preprocess.py
rm -r python
rm -r *.pkl
rm python.zip
cd ..

2. You can also continue finetuning the model on CoSQA dataset. CoSQA includes 20,604 labels for pairs of natural language queries and Python codes, with almost the same collections and data format as the 1,046 pairs in WebQueryTest. You can find the CoSQA training and dev set in ./data/cosqa_train.json and ./data/cosqa_dev.json . The detailed construction of CoSQA can be found in the paper CoSQA: 20,000+ Web Queries for Code Search and Question Answering (In Proceedings of ACL 2021).

Data statistics

Data statistics of WebQueryTest are shown in the table below:

	#Examples
WebQueryTest	1,046

Data statistics of CoSQA are shown in the table below:

	#Examples
CoSQA-training	20,000
CoSQA-dev	604

Fine-tuning

You can use the following command to finetune the model on CodeSearchNet:

python code/run_classifier.py \
			--model_type roberta \
			--do_train \
			--do_eval \
			--eval_all_checkpoints \
			--train_file train_codesearchnet_7.json \
			--dev_file dev_codesearchnet.json \
			--max_seq_length 200 \
			--per_gpu_train_batch_size 16 \
			--per_gpu_eval_batch_size 16 \
			--learning_rate 1e-5 \
			--num_train_epochs 3 \
			--gradient_accumulation_steps 1 \
			--warmup_steps 1000 \
			--evaluate_during_training \
			--data_dir ./data/ \
			--output_dir ./model_codesearchnet \
			--encoder_name_or_path microsoft/codebert-base 

You can then use the following command to continue finetuning the model on CoSQA:

python code/run_classifier.py \
			--model_type roberta \
			--do_train \
			--do_eval \
			--eval_all_checkpoints \
			--train_file cosqa_train.json \
			--dev_file cosqa_dev.json \
			--max_seq_length 200 \
			--per_gpu_train_batch_size 16 \
			--per_gpu_eval_batch_size 16 \
			--learning_rate 1e-5 \
			--num_train_epochs 3 \
			--gradient_accumulation_steps 1 \
			--warmup_steps 5000 \
			--evaluate_during_training \
			--data_dir ./data/ \
			--output_dir ./model_cosqa_continue_training \
			--encoder_name_or_path ./model_codesearchnet 

Evaluation

To test on the WebQueryTest, you run the following command. Also it will automatically generate predictions to --prediction_file.

python code/run_classifier.py \
			--model_type roberta \
			--do_predict \
			--test_file test_webquery.json \
			--max_seq_length 200 \
			--per_gpu_eval_batch_size 2 \
			--data_dir ./data \
			--output_dir ./model_cosqa_continue_training/checkpoint-best-aver/ \
			--encoder_name_or_path microsoft/codebert-base \
			--pred_model_dir ./model_cosqa_continue_training/checkpoint-last/ \
			--prediction_file ./evaluator/webquery_predictions.txt 
			

After generate predictions for WebQueryTest, you can use our provided script to evaluate:

python evaluator/evaluator.py \
		--answers_webquery ./evaluator/webquery_answers.txt \
		--predictions_webquery evaluator/webquery_predictions.txt

Results

The results on WebQueryTest are shown as below:

dataset	model	Accuracy
CodeSearchNet	RoBERTa	40.92
CodeSearchNet	CodeBERT	47.80
CodeSearchNet+ CoSQA	CodeBERT	52.87

Cite

If you use this code or our WebQueryTest dataset, please considering citing CodeXGLUE and CodeSearchNet:

@article{Lu2021CodeXGLUEAM,
  title={CodeXGLUE: A Machine Learning Benchmark Dataset for Code Understanding and Generation},
  author={Shuai Lu and Daya Guo and Shuo Ren and Junjie Huang and Alexey Svyatkovskiy and Ambrosio Blanco and Colin Clement and Dawn Drain and Daxin Jiang and Duyu Tang and Ge Li and Lidong Zhou and Linjun Shou and Long Zhou and Michele Tufano and Ming Gong and Ming Zhou and Nan Duan and Neel Sundaresan and Shao Kun Deng and Shengyu Fu and Shujie Liu},
  journal={ArXiv},
  year={2021},
  volume={abs/2102.04664}
}

@inproceedings{Huang2020CoSQA20,
  title={CoSQA: 20, 000+ Web Queries for Code Search and Question Answering},
  author={Junjie Huang and Duyu Tang and Linjun Shou and Ming Gong and Ke Xu and Daxin Jiang and Ming Zhou and Nan Duan},
  booktitle={ACL/IJCNLP},
  year={2020}
}

@article{husain2019codesearchnet,
  title={Codesearchnet challenge: Evaluating the state of semantic code search},
  author={Husain, Hamel and Wu, Ho-Hsiang and Gazit, Tiferet and Allamanis, Miltiadis and Brockschmidt, Marc},
  journal={arXiv preprint arXiv:1909.09436},
  year={2019}
}