Skip to content

Latest commit

 

History

History

CoaT

README.md

Co-Scale Conv-Attentional Image Transformers, arxiv

PaddlePaddle training/validation code and pretrained models for the model released in ICCV2021: CoaT.

The official PyTorch implementation is here.

This implementation is developed by PaddleViT.

drawing

ConvNeXt Model Overview

Update

  • Update (2022-05-16): Code is released and ported weights are uploaded.

Models Zoo

Model Acc@1 Acc@5 #Params FLOPs Image Size Crop_pct Interpolation Link
CoaT-Lite Tiny 77.51 93.92 5.7M 1.7G 224 0.9 bicubic google/baidu
CoaT-Lite Mini 79.10 94.61 11.0M 2.1G 224 0.9 bicubic google/baidu
CoaT-Lite Small 81.83 95.58 19.8M 4.2G 224 0.9 bicubic google/baidu
CoaT-Lite Medium 83.60 96.73 44.6M 10.5G 224 0.9 bicubic google/baidu
CoaT Tiny 78.45 94.07 7.7M 4.8G 224 0.9 bicubic google/baidu
CoaT Mini 81.27 95.38 14.8M 7.3G 224 0.9 bicubic google/baidu
CoaT Small 82.36 96.21 31.5M 13.3G 224 0.9 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 ./coat_tiny.pdparams, to use the coat_tiny model in python:

from config import get_config
from coat import build_coat as build_model
# config files in ./configs/
config = get_config('./configs/coat_tiny.yaml')
# build model
model = build_model(config)
# load pretrained weights
model_state_dict = paddle.load('./coat_tiny.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/coat_tiny.yaml' \
-dataset='imagenet2012' \
-batch_size=256 \
-data_path='/dataset/imagenet' \
-eval \
-pretrained='./coat_tiny.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/coat_tiny.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{Xu_2021_ICCV,
    author    = {Xu, Weijian and Xu, Yifan and Chang, Tyler and Tu, Zhuowen},
    title     = {Co-Scale Conv-Attentional Image Transformers},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {9981-9990}
}