TODO: This README should be modified

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, arxiv

PaddlePaddle training/validation code and pretrained models for ViT.

The official TF implementation is here.

This implementation is developed by PaddleViT.

ViT Model Overview

Update

• Update (2021-09-27): More weights are uploaded.
• Update (2021-08-11): Code is released and ported weights are uploaded.

Models Zoo

Model	Acc@1	Acc@5	#Params	FLOPs	Image Size	Crop_pct	Interpolation	Link
vit_base_patch32_224	80.68	95.61	88.2M	4.4G	224	0.875	bicubic	google/baidu(ubyr)
vit_base_patch32_384	83.35	96.84	88.2M	12.7G	384	1.0	bicubic	google/baidu(3c2f)
vit_base_patch16_224	84.58	97.30	86.4M	17.0G	224	0.875	bicubic	google/baidu(qv4n)
vit_base_patch16_384	85.99	98.00	86.4M	49.8G	384	1.0	bicubic	google/baidu(wsum)
vit_large_patch16_224	85.81	97.82	304.1M	59.9G	224	0.875	bicubic	google/baidu(1bgk)
vit_large_patch16_384	87.08	98.30	304.1M	175.9G	384	1.0	bicubic	google/baidu(5t91)
vit_large_patch32_384	81.51	96.09	306.5M	44.4G	384	1.0	bicubic	google/baidu(ieg3)

*The results are evaluated on ImageNet2012 validation set.

Notebooks

We provide a few notebooks in aistudio to help you get started:

*(coming soon)*

Requirements

• Python>=3.6
• yaml>=0.2.5
• PaddlePaddle>=2.1.0
• yacs>=0.1.8

Data

ImageNet2012 dataset is used in the following folder structure:

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

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 the downloaded weight file is stored in ./vit_base_patch16_224.pdparams, to use the vit_base_patch16_224 model in python:

from config import get_config
from transformer import build_vit as build_model
# config files in ./configs/
config = get_config('./configs/vit_base_patch16_224.yaml')
# build model
model = build_model(config)
# load pretrained weights, .pdparams is NOT needed
model_state_dict = paddle.load('./vit_base_patch16_224.pdparams')
model.set_dict(model_state_dict)

Evaluation

To evaluate ViT model performance on ImageNet2012 with a single GPU, run the following script using command line:

sh run_eval.sh

or

CUDA_VISIBLE_DEVICES=0 \
python main_single_gpu.py \
    -cfg='./configs/vit_base_patch16_224.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \
    -eval \
    -pretrained='./vit_base_patch16_224.pdparams'

Run evaluation using multi-GPUs:

sh run_eval_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg='./configs/vit_base_patch16_224.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \
    -eval \
    -pretrained='./vit_base_patch16_224.pdparams'

Training

To train the ViT model on ImageNet2012 with single GPU, run the following script using command line:

sh run_train.sh

or

CUDA_VISIBLE_DEVICES=0 \
python main_single_gpu.py \
  -cfg='./configs/vit_base_patch16_224.yaml' \
  -dataset='imagenet2012' \
  -batch_size=32 \
  -data_path='/dataset/imagenet' \

Run training using multi-GPUs:

sh run_train_multi.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg='./configs/vit_base_patch16_224.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \

Visualization Attention Map

(coming soon)

Reference

@article{dosovitskiy2020image,
  title={An image is worth 16x16 words: Transformers for image recognition at scale},
  author={Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and others},
  journal={arXiv preprint arXiv:2010.11929},
  year={2020}
}