pytorch-nam 0.1.0

Neural Additive Models (NAM) library for interpretable machine learning with shape function visualization

NAM - Neural Additive Models

A PyTorch library for building and visualizing Neural Additive Models (NAMs) - interpretable machine learning models that decompose predictions into individual feature contributions.

Features

• Interpretable ML: Build Neural Additive Models that decompose predictions into individual feature contributions
• Shape Function Visualization: Plot and analyze how each feature affects predictions
• Architecture Diagrams: Generate visual explanations of the NAM structure
• PyTorch-based: Efficient training with automatic differentiation
• Model Persistence: Save and load trained models
• Easy to Use: Simple API for training and inference

Installation

pip install nam

Quick Start

Training a NAM

import torch
from nam import NAM, train_nam

# Prepare your data
X = torch.randn(1000, 5)  # 1000 samples, 5 features
y = X[:, 0] * 2 + X[:, 1] ** 2 - X[:, 2]  # Some non-linear relationship

# Train the model
model = train_nam(X, y, num_features=5, hidden_dim=32, depth=5, epochs=1000)

# Make predictions
predictions = model(X)

Using the Model

from nam import NAM

# Create model
model = NAM(num_features=5, hidden_dim=32, depth=5)

# Forward pass
output = model(X)

# Save model
model.save_model('my_nam_model.pth')

# Load model
loaded_model = NAM.load_model('my_nam_model.pth')

Visualizing Shape Functions

Use the built-in visualization tools to understand feature contributions:

from nam import plot_shape_functions, get_shape_function_values
import matplotlib.pyplot as plt

# Plot all shape functions
figures = plot_shape_functions(model, X, feature_names=['f1', 'f2', 'f3', 'f4', 'f5'])

# Or get the raw values for custom plotting
values = get_shape_function_values(model, X, feature_names=['f1', 'f2', 'f3', 'f4', 'f5'])
for feature_name, (x_range, y_values) in values.items():
    plt.figure()
    plt.plot(x_range, y_values)
    plt.title(f'Shape Function: {feature_name}')
    plt.show()

Architecture Visualization

Generate publication-ready architecture diagrams:

from nam import make_nam_architecture_figure

fig = make_nam_architecture_figure(
    feature_names=['feature1', 'feature2', 'feature3'],
    example_inputs=[0.5, -1.2, 0.8],
    example_outputs=[0.3, -0.5, 0.2]
)
fig.savefig('nam_architecture.png', dpi=300, bbox_inches='tight')

What are Neural Additive Models?

Neural Additive Models (NAMs) are interpretable machine learning models that learn a separate neural network for each input feature. The final prediction is the sum of all feature contributions:

y = f₁(x₁) + f₂(x₂) + ... + fₙ(xₙ)

This additive structure makes it easy to understand how each feature affects predictions, while still capturing non-linear relationships.

Examples

See the experiments/ directory for complete examples including:

• Housing price prediction with interactive Gradio visualization
• Shape function plotting
• Model training and evaluation

API Reference

NAM

Main model class implementing a Neural Additive Model.

Parameters:

• num_features (int): Number of input features
• hidden_dim (int): Hidden dimension for shape function networks
• depth (int): Depth of each shape function network

Methods:

• forward(x): Forward pass returning predictions
• save_model(path): Save model to disk
• load_model(path): Load model from disk (classmethod)

train_nam

Train a NAM model on your data.

Parameters:

• X (torch.Tensor): Input features (n_samples, n_features)
• y (torch.Tensor): Target values (n_samples,)
• num_features (int): Number of features
• hidden_dim (int): Hidden dimension (default: 32)
• depth (int): Network depth (default: 5)
• epochs (int): Training epochs (default: 1000)
• lr (float): Learning rate (default: 0.01)
• verbose (bool): Print progress (default: True)

Returns:

• Trained NAM model

Contributing

Contributions are welcome! Please open issues or pull requests at the GitHub repository.

License

This project is licensed under the MIT License. See LICENSE.md for details.

References

Based on the Neural Additive Models paper:

• Agarwal, R., Melnick, L., Frosst, N., Zhang, X., Lengerich, B., Caruana, R., & Hinton, G. E. (2021). Neural additive models: Interpretable machine learning with neural nets. NeurIPS.