MaskAlign (CVPR 2023)

This is the official PyTorch repository for CVPR 2023 paper Stare at What You See: Masked Image Modeling without Reconstruction:

@article{xue2022stare,
  title={Stare at What You See: Masked Image Modeling without Reconstruction},
  author={Xue, Hongwei and Gao, Peng and Li, Hongyang and Qiao, Yu and Sun, Hao and Li, Houqiang and Luo, Jiebo},
  journal={arXiv preprint arXiv:2211.08887},
  year={2022}
}

• This repo is a modification on the MAE repo. Installation and preparation follow that repo.

• The teacher models in this repo are called from Huggingface. Please install transformers package by running:
  pip install transformers.

Pre-training

To pre-train ViT-base (recommended default) with distributed training, run the following on 8 GPUs:

python -m torch.distributed.launch --nproc_per_node=8 main_pretrain.py \
    --batch_size 128 \
    --model mae_vit_base_patch16 \
    --blr 1.5e-4 \
    --min_lr 1e-5 \
    --data_path ${IMAGENET_DIR} \
    --output_dir ${OUTPUT_DIR} \
    --target_norm whiten \
    --loss_type smoothl1 \
    --drop_path 0.1 \
    --head_type linear \
    --epochs 200 \
    --warmup_epochs 20 \
    --mask_type attention \
    --mask_ratio 0.7 \
    --loss_weights top5 \
    --fusion_type linear \
    --teacher_model openai/clip-vit-base-patch16

• Here the effective batch size is 128 (batch_size per gpu) * 8 (gpus) = 1024. If memory or # gpus is limited, use --accum_iter to maintain the effective batch size, which is batch_size (per gpu) * nodes * 8 (gpus) * accum_iter.
• blr is the base learning rate. The actual lr is computed by the linear scaling rule: lr = blr * effective batch size / 256.
• This repo will automatically resume the checkpoints by keeping a "latest checkpoint".

To train ViT-Large, please set --model mae_vit_large_patch16 and --drop_path 0.2. Currently, this repo supports three teacher models: --teacher_model ${TEACHER}, where ${TEACHER} in openai/clip-vit-base-patch16, openai/clip-vit-large-patch14 and facebook/dino-vitb16.

Fine-tuning

Get our pre-trained checkpoints from here.

To fine-tune ViT-base (recommended default) with distributed training, run the following on 8 GPUs:

python -m torch.distributed.launch --nproc_per_node=8 main_finetune.py \
    --epochs 100 \
    --batch_size 128 \
    --model vit_base_patch16 \
    --blr 3e-4 \
    --layer_decay 0.55 \
    --weight_decay 0.05 \
    --drop_path 0.2 \
    --reprob 0.25 \
    --mixup 0.8 \
    --cutmix 1.0 \
    --dist_eval \
    --finetune ${PT_CHECKPOINT} \
    --data_path ${IMAGENET_DIR} \
    --output_dir ${OUTPUT_DIR}

• Here the effective batch size is 128 (batch_size per gpu) * 8 (gpus) = 1024.
• blr is the base learning rate. The actual lr is computed by the linear scaling rule: lr = blr * effective batch size / 256.

To fine-tune ViT-Large, please set --model vit_large_patch16 --epochs 50 --drop_path 0.4 --layer_decay 0.75 --blr 3e-4.

Linear Probing

Run the following on 8 GPUs:

python -m torch.distributed.launch --nproc_per_node=8 main_linprobe.py \
    --epochs 90 \
    --batch_size 2048 \
    --model vit_base_patch16 \
    --blr 0.025 \
    --weight_decay 0.0 \
    --dist_eval \
    --finetune ${PT_CHECKPOINT} \
    --data_path ${IMAGENET_DIR} \
    --output_dir ${OUTPUT_DIR}

• Here the effective batch size is 2048 (batch_size per gpu) * 8 (gpus) = 16384.
• blr is the base learning rate. The actual lr is computed by the linear scaling rule: lr = blr * effective batch size / 256.