DySymNet

KO | EN

gitlite — search

#python #hacktoberfest #ai #llm #golang #machine-learning #deep-learning #macos #java #windows #awesome #kubernetes #openai #open-source

(★ 22)

[ICML 2024] Official Pytorch implementation of the paper "A Neural-Guided Dynamic Symbolic Network for Exploring Mathematical Expressions from Data"

DySymNet 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

## DySymNet

![overview](https://raw.githubusercontent.com/AILWQ/DySymNet/main/img/Overview.png)

This repository contains the official Pytorch implementation for the paper [***A Neural-Guided Dynamic Symbolic Network for Exploring Mathematical Expressions from Data***](https://openreview.net/forum?id=pTmrk4XPFx) accepted by ICML'24.

[![Paper](https://img.shields.io/badge/Paper-📖-blue?)](https://openreview.net/pdf?id=IejxxE9DO2)
[![Paper](https://img.shields.io/badge/arXiv-2309.13705-b31b1b)](https://arxiv.org/abs/2309.13705)
![visitor badge](https://visitor-badge.laobi.icu/badge?page_id=AILWQ.DySymNet&left_color=red&right_color=green&left_text=Hello%20Visitors)
![GitHub Repo stars](https://img.shields.io/github/stars/AILWQ/DySymNet?style=social)

## 🔥 News
[2024/10/12] ***Now DySymNet can be installed via 'pip install DySymNet'. You only need 1 command to start exploring expressions!***

## 🚀 Highlights

- ***DySymNet*** is a new search paradigm for symbolic regression (SR) that searches the symbolic network with various architectures instead of searching expressions in the large functional space.
- ***DySymNet*** possesses promising capabilities in solving high-dimensional problems and optimizing coefficients, which are lacking in current SR methods.
- ***DySymNet*** outperforms state-of-the-art baselines across various SR standard benchmark datasets and the well-known SRBench with more variables.

## 📦 Install

Install ***DySymNet*** using only one command:

```setup
pip install DySymNet
```

## 🤗 Quick start

You can create and run the following script in any directory:

```python
# Demo.py
import numpy as np
from DySymNet import SymbolicRegression
from DySymNet.scripts.params import Params
from DySymNet.scripts.functions import *

# You can customize some hyperparameters according to parameter configuration
config = Params()

# such as operators 
funcs = [Identity(), Sin(), Cos(), Square(), Plus(), Sub(), Product()]
config.funcs_avail = funcs

# Example 1: Input ground truth expression
SR = SymbolicRegression.SymboliRegression(config=config, func="x_1**3 + x_1**2 + x_1", func_name="Nguyen-1")
eq, R2, error, relative_error = SR.solve_environment()
print('Expression: ', eq)
print('R2: ', R2)
print('error: ', error)
print('relative_error: ', relative_error)
print('log(1 + MSE): ', np.log(1 + error))
```

Then you can get a folder named as "results" in the current directory, which contains subfolders named func_name that record the logs of the script running process.

## ⚙️ Parameter configuration

The main running script is `SymbolicRegression.py` and it relies on configuring runs via `params.py`. The `params.py` includes various hyperparameters of the controller RNN and the symbolic network. You can configure the following hyperparameters as required:

#### parameters for symbolic network structure

Note: You can add the additional operators in the `functions.py` by referring to existing operators and place them inside `funcs_avail`  if you want to use them.

#### parameters for controller RNN

#### parameters for symbolic network training

**Note:** `threshold` controls the complexity of the final expression, and is a trade-off between complexity and precision, which you can customise according to your actual requirements.

#### parameters for genearting input data

#### Additional parameters

`results_dir` configures the save path for all results

## 🤖 Symbolic Regression

We provide two ways to perform symbolic regression tasks.

#### Option1: Input ground truth expression

When you want to discover an expression for which the ground truth is known, for example to test a standard benchmark, you can run the script `Demo.py`  as follows:

```python
# Demo.py
import numpy as np
from DySymNet import SymbolicRegression
from DySymNet.scripts.params import Params
from DySymNet.scripts.functions import *

# You can customize some hyperparameters according to READEME
config = Params()

# config operators 
funcs = [Identity(), Sin(), Cos(), Square(), Plus(), Sub(), Product()]
config.funcs_avail = funcs
params = Params()  # configuration for a specific task

# input ground truth expression
ground_truth_eq = "x_1 + x_2"  # variable names should be written as x_i, where i>=1.
eq_name = "x_1+x_2"
SR = SymbolicRegression.SymbolicRegression(config=params, func=ground_truth_eq, fun_name=eq_name)  # A new folder named "func_name" will be created to store the result files.
eq, R2, error, relative_error = SR.solve_environment()  # return results
print('Expression: ', eq)
print('R2: ', R2)
print('error: ', error)
print('relative_error: ', relative_error)
print('log(1 + MSE): ', np.log(1 + error))
```

In this way, the function `generate_data` is used to automatically generate the corresponding data set $\mathcal{D}(X, y)$ for inference, instead of you generating the data yourself.

Then, you can run `Demo.py` directly, or you can run it in the terminal as follows:

```python
python Demo.py
```

After running this script, the results will be stored in path `./results/test/func_name`.

#### Option2: Load the data file

When you only have observed data and do not know the ground truth, you can perform symbolic regression by entering the path to the csv data file:

```python
# Demo.py
import numpy as np
from DySymNet import SymbolicRegression
from DySymNet.scripts.params import Params
from DySymNet.scripts.functions import *

# You can customize some hyperparameters according to READEME
config = Params()

# config operators 
funcs = [Identity(), Sin(), Cos(), Square(), Plus(), Sub(), Product()]
config.funcs_avail = funcs
params = Params()  # configuration for a specific task

# specify the data file path
data_path = './data/Nguyen-1.csv'  # data file should be in csv format
SR = SymbolicRegression.SymbolicRegression(config=params, func_name='Nguyen-1', data_path=data_path)  # you can rename the func_name as any other you want.
eq, R2, error, relative_error = SR.solve_environment()  # return results
print('Expression: ', eq)
print('R2: ', R2)
print('error: ', error)
print('relative_error: ', relative_error)
print('log(1 + MSE): ', np.log(1 + error))
```

**Note:** the data file should contains ($X\_{dim} + 1$) colums, which $X\_{dim}$ is the number of independent variables and the last colum is the corresponding $y$ values.

Then, you can run `Demo.py` directly, or you can run it in the terminal as follows:

```python
python Demo.py
```

After running this script, the results will be stored in path `./results/test/func_name`.

#### Output

Once the script stops early or finishes running, you will get the following output:

```
Expression: x_1 + x_2
R2: 1.0
error: 4.3591795754679974e-13
relative_error:  2.036015757767018e-06
log(1 + MSE):  4.3587355946774144e-13
```

## 🌟 Citing this work

If you find our work and this codebase helpful, please consider starring this repo and cite:

```bibtex
@inproceedings{li2024neural,
  title={A neural-guided dynamic symbolic network for exploring mathematical expressions from data},
  author={Li, Wenqiang and Li, Weijun and Yu, Lina and Wu, Min and Sun, Linjun and Liu, Jingyi and Li, Yanjie and Wei, Shu and Deng, Yusong and Hao, Meilan},
  booktitle={Proceedings of the 41st International Conference on Machine Learning},
  pages={28222--28242},
  year={2024}
}
```

DySymNet

DySymNet 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

DySymNet

DySymNet 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README