🇨🇳中文 | 🌐English | 📖文档/Docs | 🤖模型/Models

---

CodeAssist: Advanced Code Completion Tool

Introduction

CodeAssist is an advanced code completion tool that intelligently provides high-quality code completions for Python, Java, and C++ and so on.

CodeAssist 是一个高级代码补全工具，高质量为 Python、Java 和 C++ 等编程语言补全代码

Features

• GPT based code completion
• Code completion for Python, Java, C++, javascript and so on
• Line and block code completion
• Train(Fine-tuning) and predict model with your own data

Release Models

Arch	BaseModel	Model	Model Size
GPT	gpt2	shibing624/code-autocomplete-gpt2-base	487MB
GPT	distilgpt2	shibing624/code-autocomplete-distilgpt2-python	319MB
GPT	bigcode/starcoder	WizardLM/WizardCoder-15B-V1.0	29GB

Demo

HuggingFace Demo: https://huggingface.co/spaces/shibing624/code-autocomplete

backend model: shibing624/code-autocomplete-gpt2-base

Install

pip install torch # conda install pytorch
pip install -U codeassist

or

git clone https://github.com/shibing624/codeassist.git
cd CodeAssist
python setup.py install

Usage

WizardCoder model

WizardCoder-15b is fine-tuned bigcode/starcoder with alpaca code data, you can use the following code to generate code:

example: examples/wizardcoder_demo.py

import sys

sys.path.append('..')
from codeassist import WizardCoder

m = WizardCoder("WizardLM/WizardCoder-15B-V1.0")
print(m.generate('def load_csv_file(file_path):')[0])

output:

import csv

def load_csv_file(file_path):
    """
    Load data from a CSV file and return a list of dictionaries.
    """
    # Open the file in read mode
    with open(file_path, 'r') as file:
        # Create a CSV reader object
        csv_reader = csv.DictReader(file)
        # Initialize an empty list to store the data
        data = []
        # Iterate over each row of data
        for row in csv_reader:
            # Append the row of data to the list
            data.append(row)
    # Return the list of data
    return data

model output is impressively effective, it currently supports English and Chinese input, you can enter instructions or code prefixes as required.

distilgpt2 model

distilgpt2 fine-tuned code autocomplete model, you can use the following code:

example: examples/distilgpt2_demo.py

import sys

sys.path.append('..')
from codeassist import GPT2Coder

m = GPT2Coder("shibing624/code-autocomplete-distilgpt2-python")
print(m.generate('import torch.nn as')[0])

output:

import torch.nn as nn
import torch.nn.functional as F

Use with huggingface/transformers：

example: examples/use_transformers_gpt2.py

Train Model

Train WizardCoder model

example: examples/training_wizardcoder_mydata.py

cd examples
CUDA_VISIBLE_DEVICES=0,1 python training_wizardcoder_mydata.py --do_train --do_predict --num_epochs 1 --output_dir outputs-wizard --model_name WizardLM/WizardCoder-15B-V1.0

• GPU memory: 31GB
• finetune need 2*V100(32GB)
• inference need 1*V100(32GB)

Train distilgpt2 model

example: examples/training_gpt2_mydata.py

cd examples
python training_gpt2_mydata.py --do_train --do_predict --num_epochs 15 --output_dir outputs-gpt2 --model_name gpt2

PS: fine-tuned result model is GPT2-python: shibing624/code-autocomplete-gpt2-base, I spent about 24 hours with V100 to fine-tune it.

Server

start FastAPI server:

example: examples/server.py

cd examples
python server.py

open url: http://0.0.0.0:8001/docs

Dataset

This allows to customize dataset building. Below is an example of the building process.

Let's use Python codes from Awesome-pytorch-list

1. We want the model to help auto-complete codes at a general level. The codes of The Algorithms suits the need.
2. This code from this project is well written (high-quality codes).

dataset tree:

examples/download/python
├── train.txt
└── valid.txt
└── test.txt

There are three ways to build dataset:

1. Use the huggingface/datasets library load the dataset huggingface datasets https://huggingface.co/datasets/shibing624/source_code

from datasets import load_dataset
dataset = load_dataset("shibing624/source_code", "python") # python or java or cpp
print(dataset)
print(dataset['test'][0:10])

output:

DatasetDict({
    train: Dataset({
        features: ['text'],
        num_rows: 5215412
    })
    validation: Dataset({
        features: ['text'],
        num_rows: 10000
    })
    test: Dataset({
        features: ['text'],
        num_rows: 10000
    })
})
{'text': [
"            {'max_epochs': [1, 2]},\n", 
'            refit=False,\n', '            cv=3,\n', 
"            scoring='roc_auc',\n", '        )\n', 
'        search.fit(*data)\n', 
'', 
'    def test_module_output_not_1d(self, net_cls, data):\n', 
'        from skorch.toy import make_classifier\n', 
'        module = make_classifier(\n'
]}

2. Download dataset from Cloud

Name	Source	Download	Size
Python+Java+CPP source code	Awesome-pytorch-list(5.22 Million lines)	github_source_code.zip	105M

download dataset and unzip it, put to examples/.

3. Get source code from scratch and build dataset

prepare_code_data.py

cd examples
python prepare_code_data.py --num_repos 260

Contact

• Issue(建议) ：
• 邮件我：xuming: [email protected]
• 微信我： 加我微信号：xuming624, 备注：个人名称-公司-NLP 进NLP交流群。

Citation

如果你在研究中使用了codeassist，请按如下格式引用：

APA:

Xu, M. codeassist: Code AutoComplete with GPT model (Version 1.0.0) [Computer software]. https://github.com/shibing624/codeassist

BibTeX:

@software{Xu_codeassist,
author = {Ming Xu},
title = {CodeAssist: Code AutoComplete with Generation model},
url = {https://github.com/shibing624/codeassist},
version = {1.0.0}
}

License

This repository is licensed under the The Apache License 2.0.

Please follow the Attribution-NonCommercial 4.0 International to use the WizardCoder model.

Contribute

项目代码还很粗糙，如果大家对代码有所改进，欢迎提交回本项目，在提交之前，注意以下两点：

• 在tests添加相应的单元测试
• 使用python setup.py test来运行所有单元测试，确保所有单测都是通过的

之后即可提交PR。

Reference

• gpt-2-simple
• galois-autocompleter
• WizardLM/WizardCoder-15B-V1.0