INR_BEP: Modular Framework for Training and Benchmarking Implicit Neural Representations

This repository offers a flexible and extensible framework for research and education on Implicit Neural Representations (INRs). It supports a range of architectures, layer types, activation functions, positional encodings, and data modalities.

---

Repository Structure

• inr_utils/
  Utilities for training single INRs (coordinate MLPs/neural fields): includes training loops, loss functions, metrics, image utilities, callbacks, and more.

• hypernetwork_utils/
  Tools for training hypernetworks, with a similar setup to inr_utils.

• model_components/
  Core modules for INRs and hypernetworks:

  • inr_layers.py: Various INR layer types & positional encoding layers
  • activation_functions.py: Supported activation functions
  • inr_modules.py: High-level model modules (MLPs, NeRF, etc.)
  • hypernetwork_components.py: Hypernetwork-specific logic
  • Support modules: auxiliary.py, initialization_schemes.py, nerf_components.py

• Example Notebooks
  Demonstrations for images, audio, and more.

---

INR Layer Types

A wide variety of INR layers are implemented, including:

• Sinusoidal Representation Networks (SIREN)
• Cardinal Sine (sinc)
• Higher-order sine/cosine (Hosc, AdaHosc)
• Wavelet-inspired (Real/Complex WIRE)
• Linear (no activation)
• Gaussian-featured, Quadratic, Multi-Quadratic, Laplacian, Super-Gaussian, Exponential-Sine, and FINER layers

All layers are compatible with hypernetwork-based weight generation.

---

Activation Functions

Choose from a modular set of activations, such as:

• Sinusoidal (SIREN)
• Sinc / Cardinal Sine
• HOSC and Adaptive HOSC
• Quadratic & Multi-Quadratic
• Laplacian & Super-Gaussian
• Exponential Sine
• FINER (variable-periodic sine)
• Unscaled Gaussian Bump, Real/Complex Gabor Wavelets
• Linear (identity)

Activation functions can be selected or customized in your configuration.

---

Positional Encodings

Available as plug-and-play layers:

• Classical (NeRF-style Fourier features)
• Trident (from the TRIDENT paper)
• Integer Lattice (learnable, stateful, with dynamic pruning)

---

Supported Modalities

• Images: Continuous representations, super-resolution, inpainting, and more
• Audio: 1D signal modeling (see inr_audio_ex.ipynb)
• 3D/Volumes: Neural/radiance fields (NeRF)
• General: Any coordinate-to-value function

---

Example: Model Configuration

Define architectures in a Python config dictionary:

config = dict(
    architecture="my_model_code.py",
    model_type="AutoEncoder",
    encoder="ConvEncoder",
    encoder_config=dict(hidden_channels=64, kernel_size=3, depth=6),
    auxiliary_regressor="Regressor",
    auxiliary_regressor_config=dict(pred_size=10, depth=4),
    decoder="INRDecoder",
    decoder_config=dict(
        mlp_depth=3,
        mlp_width=1024,
        inr_depth=6,
        inr_layer_type="inr_layers.SirenLayer",
        inr_layer_kwargs={"w0": 30.}
    ),
    latent_size=64,
    data_channels=3,
    ndim=2,
)

---

Neural Tangent Kernels (NTKs) & Spectral Bias

This repository includes tools for analyzing spectral bias in INRs using NTKs. Spectral bias describes a neural network’s preference for learning low-frequency components before high-frequency ones. By computing the NTK spectrum, you can quantitatively compare architectures, layer types, and activation functions.

• Modular NTK computation: Analyze any INR or hypernetwork model.
• Frequency response: Probe how model design affects learning across frequencies.
• Example notebooks: Demonstrations for reproducible NTK analysis.

---

Extensibility

• Add new layers, activations, or encodings by extending model_components.
• Swap encoders, decoders, or positional encodings by editing configuration.
• Designed for clarity, collaboration, and robust experimentation.

---

For further details, consult the module docstrings and example notebooks.

Questions and contributions are welcome!

---