PiLaMIM: Toward Richer Visual Representation by Integrating Pixel and Latent Masked Image Modeling

Setting

conda create -n name python=3.8

pip install -r requirements.txt

Pretrain

python pretrain.py

Linear probing

python linear_probing.py --dataset data_name --pretrained_model_path model_path

All datasets were utilized with torchvision

data_name : imagenet1k, cifar100, cifar10, Places365, INat2021, CLEVR, CLEVR_Dist

Imagenet-1k linear probing checkpoint

https://drive.google.com/drive/folders/1KCPFP8tLvz7vf_2qRlPcCamRIYqOtziI?usp=sharing