Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations

Official PyTorch implementation

---

Requirements

We provide installation instructions for ImageNet classification experiments here, based on ConvNeXt instructions.

Dependency Setup

Create an new conda virtual environment

conda create -n convnext python=3.8 -y
conda activate convnext

Install Pytorch>=1.8.0, torchvision>=0.9.0 following official instructions. For example:

pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install timm==0.3.2 tensorboardX six

The results in the paper are produced with torch==1.8.0+cu111 torchvision==0.9.0+cu111 timm==0.3.2.

Dataset Preparation

Download the ImageNet-1K classification dataset and structure the data as follows:

/path/to/imagenet-1k/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class2/
      img4.jpeg

---

Train models

Original ConvNeXt-Tiny

python -m torch.distributed.launch --nproc_per_node=8 --master_port=1440 main.py \
--model convnext_tiny \
--drop_path 0.1 \
--batch_size 32  --update_freq 16 \
--lr 4e-3 \
--model_ema true --model_ema_eval true \
--data_set IMNET \
--data_path </path/to/imagenet> \
--output_dir </path/to/output/dir> \
 --epochs 300 --warmup_epochs 20 \

ConvNeXt-Tiny Baseline (circular convolutions)

python -m torch.distributed.launch --nproc_per_node=8 --master_port=1440 main.py \
--model convnext_afc_tiny \
--blurpool_kwargs "{\"filt_size\": 1, \"scale_l2\":false}" \
--activation gelu \
--normalization_type C \
--drop_path 0.1 \
--batch_size 32  --update_freq 16 \
--lr 4e-3 \
--model_ema true --model_ema_eval true \
--data_set IMNET \
--data_path </path/to/imagenet> \
--output_dir </path/to/output/dir> \
 --epochs 300 --warmup_epochs 20 \

ConvNeXt-Tiny-AFC

python -m torch.distributed.launch --nproc_per_node=8 --master_port=1440 main.py \
--model convnext_afc_tiny \
--drop_path 0.1 \
--blurpool_kwargs "{\"filter_type\": \"ideal\", \"scale_l2\":false}" \
--activation up_poly_per_channel \
--activation_kwargs "{\"in_scale\":7, \"out_scale\":7, \"train_scale\":true}" \
--model_kwargs "{\"stem_mode\":\"activation_residual\", \"stem_activation\": \"lpf_poly_per_channel\"}" \
--stem_activation_kwargs "{\"in_scale\":7, \"out_scale\":7, \"train_scale\":true, \"cutoff\":0.75}" \
--normalization_type CHW2 \
--batch_size 32  --update_freq 16 \
--lr 4e-3 \
--model_ema true --model_ema_eval true \
--data_set IMNET \
--data_path </path/to/imagenet> \
--output_dir </path/to/output/dir> \
 --epochs 300 --warmup_epochs 20 \

ConvNeXt-Tiny-APS

python -m torch.distributed.launch --nproc_per_node=8 --master_port=1440 main.py \
--model convnext_aps_tiny \
--drop_path 0.1 \
 --blurpool_kwargs "{\"filt_size\": 1}" \
--activation gelu \
--normalization_type C \
--batch_size 32  --update_freq 16 \
--lr 4e-3 \
--model_ema true --model_ema_eval true \
--data_set IMNET \
--data_path </path/to/imagenet> \
--output_dir </path/to/output/dir> \
 --epochs 300 --warmup_epochs 20 \

---

Checkpoints

Trained models can be Found in: https://huggingface.co/hmichaeli/convnext-afc

---

Acknowledgement

This repository is built using Truly shift invariant CNNs and ConvNeXt repositories.

• Truly shift invariant CNNs:

  • https://arxiv.org/abs/2011.14214

• ConvNeXt

  • https://arxiv.org/abs/2201.03545
  • LICENSE

---

Citation

If you find this repository helpful, please consider citing:

@InProceedings{Michaeli_2023_CVPR,
    author    = {Michaeli, Hagay and Michaeli, Tomer and Soudry, Daniel},
    title     = {Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {16333-16342}
}

---