Awesome

SDMP

This is the offical implementation of "A Simple Data Mixing Prior for Improving Self-Supervised Learning" by Sucheng Ren, Huiyu Wang, Zhengqi Gao, Shengfeng He, Alan Yuille, Yuyin Zhou, Cihang Xie

MoCo with SDMP

Installing packages as Moco v3.

Training ViT small on 4 nodes.

cd moco
# ImageNet  
# On the first node
python main_moco.py \
  -a vit_small \
  --optimizer=adamw --lr=1.5e-4 --weight-decay=.1 \
  --epochs=300 --warmup-epochs=40 \
  --stop-grad-conv1 --moco-m-cos --moco-t=.2 \
  --dist-url 'tcp://hostnode:port' \
  --multiprocessing-distributed --world-size 4 --rank 0 \
  /path/to/imagenet
# On the rest node --rank 1, 2, 3

# On single node
python main_moco.py \
  -a vit_small -b 1024 \
  --optimizer=adamw --lr=1.5e-4 --weight-decay=.1 \
  --epochs=300 --warmup-epochs=40 \
  --stop-grad-conv1 --moco-m-cos --moco-t=.2 \
  --dist-url 'tcp://localhost:10001' \
  --multiprocessing-distributed --world-size 1 --rank 0 \
  /path/to/imagenet

Testing

python main_lincls.py \
  -a vit_small --lr=3 \
  --dist-url 'tcp://localhost:10001' \
  --multiprocessing-distributed --world-size 1 --rank 0 \
  --pretrained /path/to/checkpoint.pth.tar \
  /path/to/imagenet

Checkpoints

The checkpoint of SDMP with 100 epochs pretraining (--epochs=100) is at Google Drive. We also release the checkpoint of moco with 100 epochs pretraining for comparison. We make 1.2% improvements. The checkpoint of 300 epochs is coming soon.

If you have any question, feel free to email Sucheng Ren

Citing SDMP

If you think think this paper and responsity is useful, please cite

@inproceedings{ren2022sdmp,
  title     = {A Simple Data Mixing Prior for Improving Self-Supervised Learning},
  author    = {Ren, Sucheng and Wang, Huiyu and Gao, Zhengqi and He, Shengfeng and Yuille, Alan and Zhou, Yuyin and Xie, Cihang},
  booktitle = {CVPR},
  year      = {2022}
}