harDyT: A Post-Training Hardware-Aware LayerNorm Replacement

Code Repository: https://gitlab.laas.fr/lgauthier/harDyT

Associated Paper: Accepted in IEEE ISCAS 2026, To be published

Model Description

• License: MIT
• Model Type: Image Classification
• Architecture: ConvNeXt, DeiT, Swin, ViT, derivated from timm's implementations
• Dataset: ImageNet-1k

To evaluate the models:

• Clone the repository and setup the environment:

git clone https://gitlab.laas.fr/lgauthier/harDyT.git
cd harDyT
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

• Run the evaluation script:

python evaluate.py --model <model_name> --data-path <path_to_imagenet_root>

This code will download the pre-trained model weights and evaluate the model on the ImageNet-1k validation set. Typically, run python evaluate.py --model vit_base_patch16_224 --data-path imagenet_train_subsets_20k/subset_1 to evaluate the ViT-B model on a subset of ImageNet-1k that includes the validation set and 20k training samples.

Multiple models are available:

• ConvNeXt-B: convnext_base
• ConvNeXt-L: convnext_large
• DeiT-S: deit_small_patch16_224
• DeiT-B: deit_base_patch16_224
• Swin-S: swin_small_patch4_window7_224
• Swin-B: swin_base_patch4_window7_224
• ViT-S: vit_small_patch16_224
• ViT-B: vit_base_patch16_224
• ViT-L: vit_large_patch16_224

TODO: add model accuracies