Learned Queries for Efficient Local Attention (CVPR 2022 - Oral)

---

[ Arxiv ]

---

Updates (April 19th):

• QnA was accepted for Oral Presentation at CVPR 2022
• Implementation of QnA layer and other components are available
• QnA-ViT and training code will be released later this month
  • Code went refactoring - under testing and reproducing results

---

Models

Pretrained models can be downloaded from this link.

Model	Params	GFLOPs	Top-1	Warmup
QnA_ViT_tiny	16M	2.5	81.7	5
QnA_ViT_tiny_7x7	16M	2.6	82.0	5
QnA_ViT_small	25M	4.4	83.2	5
QnA_ViT_base	56M	9.7	83.9	20

---

Evaluation

Download the model parameters and copy

CUDA_VISIBLE_DEVICES=0 python3 main.py --eval_only \
    --workdir <MODEL_DIR> \ 
    --config configs/imagenet_qna.py \
    --config.model_name <MODEL_DIR> \ 
    --config.dataset_version 5.1.0  \
    --config.data_dir <DATA_DIR> \
    --config.batch_size <BATCH_SIZE> \
    --config.half_precision=False

Flags:

- workdir : location to the checkpoints directory
- model_name : the model name, e.g., qna_vit_tiny (see table above for model names - use lowercase names only).
- dataset_version : Tensorflow datasets ImageNet dataset version. Mine was (5.1.0),
                     you can change according to your installed version.
- data_dir : the location of the ImageNet directory (need to have the validation set)
- batch_size : the evaluation batch size

---

Citation

Please cite our paper if you find this repo helpful:

@InProceedings{Arar_2022_CVPR,
author = {Arar, Moab and Shamir, Ariel and Bermano, Amit H.},
title = {Learned Queries for Efficient Local Attention},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022}
}