Warning

illia is under active development. The library is evolving rapidly to ensure stable support across all frameworks. Expect ongoing changes as we improve functionality and performance.

Introduction

illia is a library for Bayesian Neural Networks that brings uncertainty quantification to deep learning, a capability that is critical in sectors such as telecommunications, medicine, and beyond. Designed with flexibility in mind, it integrates seamlessly with multiple backends and popular frameworks, enabling a single codebase to support multiple backends with minimal modifications.

For full documentation, please visit the site: https://ericssonresearch.github.io/illia/

Why Choose illia?

• Multi-Backend Support: Works with PyTorch, TensorFlow, and JAX.
• Graph Neural Networks: Currently integrated with PyTorch Geometric, with planned support for DGL and/or Spektral in future releases.
• Developer Friendly: Intuitive API design and comprehensive documentation.

Quick Start

To show how easy it is to use illia, here’s a quick example to get started. In this case, we explicitly choose the backend PyTorch, the underlying framework, and define a convolutional layer:

import os
import torch

# Configure backend (PyTorch is default)
os.environ["ILLIA_BACKEND"] = "torch"

import illia
from illia.nn import Conv2d

# Create a Bayesian convolutional layer
conv_layer = Conv2d(
    input_channels=1,
    output_channels=1,
    kernel_size=3,
)

# Define input tensor
input_tensor = torch.rand(1, 1, 4, 4)

# Define the number of iterations to apply the forward pass
num_passes = 10
outputs = [conv_layer(input_tensor) for _ in range(num_passes)]

# Stack outputs into a single tensor
outputs = torch.stack(outputs)

print(f"Output shape: {outputs.shape}")
print(f"Output std: {outputs.std()}")
print(f"Output var: {outputs.var()}")

Contributing

We welcome contributions from the community! Whether you're fixing bugs, adding features, or improving documentation:

1. Read our contributing guide for development setup.
2. Check open issues for ways to help.
3. Submit bug reports using our issue templates.

License

illia is released under the MIT License. We hope you find it useful and inspiring for your projects!