OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Kolmogorov–Arnold Transformer

Kolmogorov–Arnold Transformer is an MIT-licensed Python and PyTorch research implementation of a transformer architecture that replaces standard MLP layers with Kolmogorov–Arnold network layers. It includes computer-vision classification examples, ImageNet training and validation code, and model checkpoints.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Kolmogorov–Arnold Transformer — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use KAT when you want research code and model artifacts for experimenting with KAN-based transformer image classification, rather than a hosted model endpoint.

Who it is for

  • Researchers comparing transformer architectures that use KAN layers.
  • Computer-vision practitioners exploring image-classification models and ImageNet training.

Documented uses

  • Explore a transformer design that substitutes KAN layers for standard MLP layers.
  • Use the documented timm example to classify an image.
  • Train or validate KAT models using the provided ImageNet workflow and checkpoints.
Before you choose

Requirements & limits

The repository is public and MIT-licensed. Its documented workflows use Python and PyTorch; the README indicates PyTorch 1.x and 2.x.

  • The implementation replaces transformer MLP layers with KAN layers and describes grouped weights and rational base functions.
  • The README points to a separate repository for CUDA and Triton implementation options.
  • ImageNet training requires the dataset arranged in train and validation folders.
  • The README provides image-classification guidance; detection, segmentation and NLP are marked as unfinished.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install the KAT CUDA implementation and prepare ImageNet folders

Set up the linked KAT CUDA implementation and arrange ImageNet data in the folder structure shown by the project.

Before you start

  • PyTorch and other dependencies; the README lists timm==1.0.3 and wandb among the installation commands.
  • ImageNet data organized into train and val folders, with class subfolders.

Steps

  1. Install timm with the command below.

    Documented command
    pip install timm==1.0.3

  2. Install wandb with the command below.

    Documented command
    pip install wandb

  3. Clone the linked CUDA implementation, enter its directory, and install it in editable mode: the command below, cd rational_kat_cu, then the command below.

    Documented command
    git clone https://github.com/Adamdad/rational_kat_cu.git
    Documented command
    pip install -e .

  4. Arrange ImageNet under an imagenet/ directory with train/ and val/ subdirectories containing class-named folders and images.

Expected outcome

The linked CUDA implementation is installed in editable mode, and ImageNet is arranged in the documented directory layout.

Guide limitations

  • These installation instructions target the separately linked rational_kat_cu repository, not the kat repository itself.
  • This setup guide does not cover running training or evaluation.
Documentation 1 ↗
Source-based practical guide

Train KAT-Tiny using the provided ImageNet training script

Start KAT-Tiny training with the repository's provided training script after setup and data preparation.

Before you start

  • The training scripts are under scripts/.
  • ImageNet is prepared in the documented folder layout.
  • A configured training environment and dataset path.

Steps

  1. Set DATA_PATH in the training configuration to the location of the ImageNet dataset.

  2. From the repository, run the command below.

    Documented command
    bash scripts/train_kat_tiny_8x128.sh

  3. To change hyperparameters, edit the documented training configuration, including its DATA_PATH and training arguments.

Expected outcome

The provided KAT-Tiny training script is invoked with the configured dataset path.

Guide limitations

  • The README does not provide an execution result or guarantee successful training.
  • The sample configuration uses a local dataset path that must be adapted to the user's environment.
  • The README identifies the training scripts but does not document resource requirements.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
Adamdad/kat · Official homepage title · retained in its source language
How we review sourcesSuggest a correction