Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Rethinking Spatial Dimensions of Vision Transformers, arxiv

PaddlePaddle training/validation code and pretrained models for PiT.

The official pytorch implementation is here.

This implementation is developed by PaddleViT.

drawing

PiT Model Overview

Update

  • Update (2022-03-30): Code is refactored.
  • Update (2021-12-08): Code is updated and ported weights are uploaded.
  • Update (2021-11-13): Code is released.

Models Zoo

Model Acc@1 Acc@5 #Params FLOPs Image Size Crop_pct Interpolation Link
pit_ti 72.91 91.40 4.8M 0.5G 224 0.9 bicubic google/baidu
pit_ti_distill 74.54 92.10 5.1M 0.5G 224 0.9 bicubic google/baidu
pit_xs 78.18 94.16 10.5M 1.1G 224 0.9 bicubic google/baidu
pit_xs_distill 79.31 94.36 10.9M 1.1G 224 0.9 bicubic google/baidu
pit_s 81.08 95.33 23.4M 2.4G 224 0.9 bicubic google/baidu
pit_s_distill 81.99 95.79 24.0M 2.5G 224 0.9 bicubic google/baidu
pit_b 82.44 95.71 73.5M 10.6G 224 0.9 bicubic google/baidu
pit_b_distill 84.14 96.86 74.5M 10.7G 224 0.9 bicubic google/baidu

*The results are evaluated on ImageNet2012 validation set.

Teacher Model Link
RegNet_Y_160 google/baidu

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 ./pit_b_224.pdparams, to use the pit_b_224 model in python:

from config import get_config
from pit import build_pit as build_model
# config files in ./configs/
config = get_config('./configs/pit_b_224.yaml')
# build model
model = build_model(config)
# load pretrained weights
model_state_dict = paddle.load('./pit_b_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/pit_b_224.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-eval \
-pretrained='./pit_b_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 with distillation, run the following script using command line:

sh run_train_multi_distill.sh

or

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu_distill.py \
-cfg='./configs/pit_b_distilled_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.

Finetuning

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

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python main_multi_gpu_distill.py \
-cfg='./configs/pit_b_distilled_384.yaml' \
-dataset='imagenet2012' \
-batch_size=16 \
-data_path='/dataset/imagenet' \
-pretrained='./pit_b_distilled_224.pdparams' \
-amp

Note: use -pretrained argument to set the pretrained model path, you may also need to modify the hyperparams defined in config file.

Reference

@inproceedings{heo2021pit,
    title={Rethinking Spatial Dimensions of Vision Transformers},
    author={Byeongho Heo and Sangdoo Yun and Dongyoon Han and Sanghyuk Chun and Junsuk Choe and Seong Joon Oh},
    booktitle = {International Conference on Computer Vision (ICCV)},
    year={2021},
}