/ccMIM

ICLR 2023 paper

Primary LanguagePython

Contextual Image Masking Modeling via Synergized Contrasting without View Augmentation for Faster and Better Visual Pretraining (ICLR 2023)

Code of ICLR 23 paper "Contextual Image Masking Modeling via Synergized Contrasting without View Augmentation for Faster and Better Visual Pretraining"

ccMIM

This paper presents to attentively masked semantic-richer patches by importance sampling strategy.

To pre-train the encoder on ImageNet-1K, run:

spring.submit arun -n 32 --ntasks-per-node=8 --gres=gpu:8 --cpus-per-task=5 --job-name ccMIM_pretrain\
 "python $ccMIM/submit_pretrain.py\
  --batch_size 32\
  --epochs 800\
  --model ccmim_vit_base_patch16\
  --mask_ratio 0.75\
  --world_size 32\
  --warmup_epochs 40 \
  --norm_pix_loss \
  --blr 1.5e-4 --weight_decay 0.05 \
  --output_dir $ccMIM/output/\
  --log_dir $ccMIM/output/\
  --mae false\
  --accum_iter 4\
  --contrastive \
  --resume" &

The evaluation protocol follows MAE, which can be get in MAE.

If you use ccMIM as baseline or find this repository useful, please consider citing our paper:

@inproceedings{
zhang2023contextual,
title={Contextual Image Masking Modeling via Synergized Contrasting without View Augmentation for Faster and Better Visual Pretraining},
author={Shaofeng Zhang and Feng Zhu and Rui Zhao and Junchi Yan},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=A3sgyt4HWp}
}