T2T_ViT

Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet, arxiv

PaddlePaddle training/validation code and pretrained models for T2T-ViT.

The official pytorch implementation is here.

This implementation is developed by PaddleViT.

T2T-ViT Model Overview

Update

• Update (2022-03-24): Code is refactored and bugs are fixed.
• Update (2021-09-27): Model FLOPs and # params are uploaded.
• Update (2021-08-18): Code is released and ported weights are uploaded.

Models Zoo

Model	Acc@1	Acc@5	#Params	FLOPs	Image Size	Crop_pct	Interpolation	Link
t2t_vit_7	71.68	90.89	4.3M	1.0G	224	0.9	bicubic	google/baidu
t2t_vit_10	75.15	92.80	5.8M	1.3G	224	0.9	bicubic	google/baidu
t2t_vit_12	76.48	93.49	6.9M	1.5G	224	0.9	bicubic	google/baidu
t2t_vit_14	81.50	95.67	21.5M	4.4G	224	0.9	bicubic	google/baidu
t2t_vit_19	81.93	95.74	39.1M	7.8G	224	0.9	bicubic	google/baidu
t2t_vit_24	82.28	95.89	64.0M	12.8G	224	0.9	bicubic	google/baidu
t2t_vit_t_14	81.69	95.85	21.5M	4.4G	224	0.9	bicubic	google/baidu
t2t_vit_t_19	82.44	96.08	39.1M	7.9G	224	0.9	bicubic	google/baidu
t2t_vit_t_24	82.55	96.07	64.0M	12.9G	224	0.9	bicubic	google/baidu
t2t_vit_14_384	83.34	96.50	21.5M	13.0G	384	1.0	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 ./t2t_vit_7.pdparams, to use the t2t_vit_7 model in python:

from config import get_config
from t2t_vit import build_t2t_vit as build_model
# config files in ./configs/
config = get_config('./configs/t2t_vit_7.yaml')
# build model
model = build_model(config)
# load pretrained weights
model_state_dict = paddle.load('./t2t_vit_7.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/t2t_vit_7.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-eval \
-pretrained='./t2t_vit_7.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/t2t_vit_7.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

@article{yuan2021tokens,
  title={Tokens-to-token vit: Training vision transformers from scratch on imagenet},
  author={Yuan, Li and Chen, Yunpeng and Wang, Tao and Yu, Weihao and Shi, Yujun and Jiang, Zihang and Tay, Francis EH and Feng, Jiashi and Yan, Shuicheng},
  journal={arXiv preprint arXiv:2101.11986},
  year={2021}
}