BGU-CS-VIL/WTConv

Wavelet Convolutions for Large Receptive Fields. ECCV 2024.

What it solves

It addresses the challenge of achieving large receptive fields in convolutional neural networks (CNNs) without the typical computational overhead associated with very large kernels.

How it works

WTConv implements Wavelet Convolutions, which use wavelet transforms to expand the network's field of view. It provides both 2D and 1D implementations (WTConv2d and WTConv1d) that can be integrated into existing CNN architectures. To improve performance, the project includes a high-performance implementation called fast_wtconv optimized for CUDA, Metal (MPS), and Triton backends.

Who it’s for

Researchers and developers working with computer vision and signal processing who want to improve the receptive field of their models while maintaining efficiency.

Highlights

  • Large Receptive Fields: Specifically designed to increase the area of the image a convolution operation can "see."
  • Optimized Backends: Includes fast_wtconv for high-performance execution on CUDA, Metal, and Triton.
  • Pre-trained Models: Provides WTConvNeXt models (Tiny, Small, and Base) trained on ImageNet-1K.
  • Flexible Integration: Compatible with the timm model registry for easy deployment.

Related

  • Project
    v0lta/PyTorch-Wavelet-Toolbox

    A PyTorch-native toolbox for discrete and continuous wavelet transforms that provides GPU acceleration and gradient support for signal and image processing.

  • Project
    rwth-i6/returnn

    An extensible training framework for universal recurrent neural networks optimized for fast multi-GPU training and decoding.

  • Project
    philipperemy/keras-tcn

    A Keras implementation of Temporal Convolutional Networks (TCN) that provides a more stable and parallelizable alternative to LSTMs and GRUs for long sequence modeling.

  • Project
    PyWavelets/pywt

    A Python library for wavelet transforms that enables time-frequency analysis of signals and data across multiple dimensions.

  • Project
    CVCUDA/CV-CUDA

    CV‑CUDA is NVIDIA’s open‑source GPU‑accelerated computer‑vision library. It provides fast, scalable image/video operators (resize, filter, color conversion, etc.) that run entirely on the GPU and integrate with C/C++ and Python AI frameworks. Install via pre‑built wheels (`cvcuda-cu12` or `cvcuda-cu13`) for Linux, decode images directly to GPU with `nvimgcodec`, and chain CV‑CUDA operators before feeding data to deep‑learning models. The library targets modern NVIDIA GPUs (Turing and newer) and is used in production systems such as Microsoft Bing visual search and Tencent Cloud video processing.