visresearch / aps Goto Github PK

PyTorch implementation and pre-trained models for paper APS: Asymmetric Patch Sampling for Contrastive Learning.

APS is a novel asymmetric patch sampling strategy for contrastive learning, to further boost the appearance asymmetry for better representations. APS significantly outperforms the existing self-supervised methods on both ImageNet-1K and CIFAR dataset, e.g., 2.5% finetune accuracy improvement on CIFAR100. Additionally, compared to other self-supervised methods, APS is more efficient on both memory and computation during training.

conda create -n asp python=3.9
pip install -r requirements.txt

Torchvision provides CIFAR10, CIFAR100 datasets. The root paths of data are respectively set to ./dataset/cifar10 and ./dataset/cifar100. ImageNet-1K dataset is placed at ./dataset/ILSVRC.

To start the APS pre-training, simply run the following commands.

• arch is the architecture of the pre-trained models，you can choose vit-tiny, vit-small and vit-base.
• dataset is the pre-trained dataset.
• data-root is the path of the dataset.
• nepoch is the pre-trained epochs.

Run APS with ViT-Small/2 network on a single node on CIFAR100 for 1600 epochs with the following command.

python main_pretrain.py --arch='vit-small' --dataset='cifar100' --data-root='./dataset/cifar100' --nepoch=1600

To finetune ViT-Small/2 on CIFAR100 with the following command.

python main_finetune.py --arch='vit-small' --dataset='cifar100' --data-root='./dataset/cifar100'  \
                   --pretrained-weights='./weight/pretrain/cifar100/small_1600ep_5e-4_100.pth'

• CIFAR10 and CIFAR100

• ImageNet-1K

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.