nnapprox (Neural Network Approximation)

Neural network function approximation using PyTorch (Future backends planned). Approximate arbitrary functions with automatic scaling, transformations, and serialization.

Full Documentation

---

Installation

Apple M1: Install pytorch with metal support via conda as described in the Apple Instructions

pip install nnapprox

Quick Start

import nnapprox as nna
import numpy as np

# Generate data
x = np.linspace(0, 10, 100)
y = np.sin(x) * np.exp(-x/10)

# Train
func = nna.create_approximator(
    input=['x'],
    output=['y'],
    backend='torch'
)
func.fit({'x': x, 'y': y}, epochs=5000)

# Predict
y_pred = func(x)

# Save/load
func.save('model.nna')
func2 = nna.load_approximator('model.nna', backend='torch')

Features

• Simple API: Train in 3 lines of code
• Flexible inputs: NumPy arrays, pandas DataFrames, scalars, or mixed types
• Transformations: Built-in (log, sqrt) or custom transforms
• GPU support: Automatic acceleration when available
• Serialization: Save/load complete model state

Multi-dimensional Example

import pandas as pd

x1, x2 = np.meshgrid(np.linspace(0,10,101), np.linspace(0,10,101))
x1 = x1.flatten()
x2 = x2.flatten()
def y_true(x1, x2):
    return 2*x1 + 3*x2**2 + 1
data = pd.DataFrame({'x1': x1, 'x2': x2, 'y': y_true(x1, x2)})

func = nna.create_approximator(
    input=['x1', 'x2'],
    output=['y'],
    backend='torch'
)
func.fit(data, epochs=3000)
predictions = func(data, return_dataframe=True)

Transformations

# Predefined transforms
func.set_transform('x1', predefined='log')
func.set_transform('y', predefined='exp')

# Custom transforms
def forward_fn(x):
    return x**3
def inverse_fn(y):
    return y**(1/3)
func.set_transform('x1', forward=forward_fn, inverse=inverse_fn)

Requirements

• Python >= 3.8
• PyTorch >= 1.9
• NumPy >= 1.20
• pandas >= 1.3
• scikit-learn >= 1.0

Optional: cloudpickle for lambda function serialization

Contributing

Contributions welcome! Please open an issue or submit a pull request.

License

MIT License - see LICENSE file.