HVT

Scalable Vision Transformers with Hierarchical Pooling arxiv

PaddlePaddle training/validation code and pretrained models for HVT.

The official pytorch implementation is here.

This implementation is developed by PaddleViT.

HVT Model Overview

Update

• Update (2022-04-08): Code is updated.
• Update (2021-12-28): Code is released and ported weights are uploaded.

Models Zoo

Model	Acc@1	Acc@5	#Params	FLOPs	Image Size	Crop_pct	Interpolation	Link
HVT-Ti-1	69.45	89.28	5.7M	0.6G	224	0.875	bicubic	google/baidu
HVT-S-0	80.30	95.15	22.0M	4.6G	224	0.875	bicubic	google/baidu
HVT-S-1	78.06	93.84	22.1M	2.4G	224	0.875	bicubic	google/baidu
HVT-S-2	77.41	93.48	22.1M	1.9G	224	0.875	bicubic	google/baidu
HVT-S-3	76.30	92.88	22.1M	1.6G	224	0.875	bicubic	google/baidu
HVT-S-4	75.21	92.34	22.1M	1.6G	224	0.875	bicubic	google/baidu

*The results are evaluated on ImageNet2012 validation set.

Data Preparation

ImageNet2012 dataset is used in the following file structure:

│imagenet/
├──train_list.txt
├──val_list.txt
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

• train_list.txt: list of relative paths and labels of training images. You can download it from: google/baidu
• val_list.txt: list of relative paths and labels of validation images. You can download it from: google/baidu

Usage

To use the model with pretrained weights, download the .pdparam weight file and change related file paths in the following python scripts. The model config files are located in ./configs/.

For example, assume weight file is downloaded in ./hvt_s0_patch16_224.pdparams, to use the hvt_s0_patch16_224 model in python:

from config import get_config
from hvt import build_hvt as build_model
# config files in ./configs/
config = get_config('./configs/hvt_s0_patch16_224.yaml')
# build model
model = build_model(config)
# load pretrained weights
model_state_dict = paddle.load('./hvt_s0_patch16_224.pdparams')
model.set_state_dict(model_state_dict)

Evaluation

To evaluate model performance on ImageNet2012, run the following script using command line:

sh run_eval_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu.py \
-cfg='./configs/hvt_s0_patch16_224.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-eval \
-pretrained='./hvt_s0_patch16_224.pdparams' \
-amp

Note: if you have only 1 GPU, change device number to CUDA_VISIBLE_DEVICES=0 would run the evaluation on single GPU.

Training

To train the model on ImageNet2012, run the following script using command line:

sh run_train_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu.py \
-cfg='./configs/hvt_s0_patch16_224.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-amp

Note: it is highly recommanded to run the training using multiple GPUs / multi-node GPUs.

Reference

@inproceedings{pan2021scalable,
  title={Scalable vision transformers with hierarchical pooling},
  author={Pan, Zizheng and Zhuang, Bohan and Liu, Jing and He, Haoyu and Cai, Jianfei},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={377--386},
  year={2021}
}