paper url: https://arxiv.org/abs/2202.03822
We propose a novel plug-in module that can be integrated to many common backbones, including CNN-based or Transformer-based networks to provide strongly discriminative regions. The plugin module can output pixel-level feature maps and fuse filtered features to enhance fine-grained visual classification. Experimental results show that the proposed plugin module outperforms state-ofthe-art approaches and significantly improves the accuracy to 92.77% and 92.83% on CUB200-2011 and NABirds, respectively.
// We move old version to ./v0/
-
install requirements
-
replace folder timm/ to our timm/ folder (for ViT or Swin-T)
pytorch model implementation timm
recommand anaconda
recommand weights and biases
deepspeed // future works
In this paper, we use 2 large bird's datasets to evaluate performance:
- our pretrained model in https://idocntnu-my.sharepoint.com/:f:/g/personal/81075001h_eduad_ntnu_edu_tw/EkypiS-W0SFDkxnHN1Imv5oBPgoRblDgW8wHuVA0c6Ka7Q?e=FhBLDC
- cub200 and nabird dataset: https://idocntnu-my.sharepoint.com/:f:/g/personal/81075001h_eduad_ntnu_edu_tw/EoBb2gijwclEulDGxv_hOtIBeKuV3M6qy3IGIGMhm-jq0g?e=tcg6tm
- resnet50_miil_21k.pth and vit_base_patch16_224_miil_21k.pth are imagenet21k pretrained model (place these file under models/), thanks to https://github.com/Alibaba-MIIL/ImageNet21K/blob/main/MODEL_ZOO.md !!
- Windows10
- Ubuntu20.04
- macOS (CPU only)
- Single GPU Training
- DataParallel (single machine multi-gpus)
- DistributedDataParallel
(more information: https://pytorch.org/tutorials/intermediate/ddp_tutorial.html)
train data and test data structure:
├── tain/
│ ├── class1/
│ | ├── img001.jpg
│ | ├── img002.jpg
│ | └── ....
│ ├── class2/
│ | ├── img001.jpg
│ | ├── img002.jpg
│ | └── ....
│ └── ....
└──
you can directly modify yaml file (in ./configs/)
python main.py --c ./configs/CUB200_SwinT.yaml
model will save in ./records/{project_name}/{exp_name}/backup/
Building model refers to ./models/builder.py
More detail in how_to_build_pim_model.ipynb
comment out main.py line 66
model = torch.nn.DataParallel(model, device_ids=None)
use_amp: True, training time about 3-hours.
use_amp: False, training time about 5-hours.
If you want to evaluate our pretrained model (or your model), please give provide configs/eval.yaml (or costom yaml file is fine)
set yaml (configuration file)
| Key | Value | Description |
|---|---|---|
| train_root | ~ | set value to ~ (null) means this is not in training mode. |
| val_root | ../data/eval/ | path to validation samples |
| pretrained | ./pretrained/best.pt | pretrained model path |
../data/eval/ folder structure:
├── eval/
│ ├── class1/
│ | ├── img001.jpg
│ | ├── img002.jpg
│ | └── ....
│ ├── class2/
│ | ├── img001.jpg
│ | ├── img002.jpg
│ | └── ....
│ └── ....
└──
python main.py --c ./configs/eval.yaml
results will show in terminal and been save in ./records/{project_name}/{exp_name}/eval_results.txt
python heat.py --c ./configs/CUB200_SwinT.yaml --img ./vis/001.jpg --save_img ./vis/001/
If you want to reason your picture and get the confusion matrix, please give provide configs/eval.yaml (or costom yaml file is fine)
set yaml (configuration file)
| Key | Value | Description |
|---|---|---|
| train_root | ~ | set value to ~ (null) means this is not in training mode. |
| val_root | ../data/eval/ | path to validation samples |
| pretrained | ./pretrained/best.pt | pretrained model path |
../data/eval/ folder structure:
├── eval/
│ ├── class1/
│ | ├── img001.jpg
│ | ├── img002.jpg
│ | └── ....
│ ├── class2/
│ | ├── img001.jpg
│ | ├── img002.jpg
│ | └── ....
│ └── ....
└──
python infer.py --c ./configs/eval.yaml
results will show in terminal and been save in ./records/{project_name}/{exp_name}/infer_results.txt
-
Thanks to timm for Pytorch implementation.
-
This work was financially supported by the National Taiwan Normal University (NTNU) within the framework of the Higher Education Sprout Project by the Ministry of Education(MOE) in Taiwan, sponsored by Ministry of Science and Technology, Taiwan, R.O.C. under Grant no. MOST 110- 2221-E-003-026, 110-2634-F-003 -007, and 110-2634-F-003 -006. In addition, we thank to National Center for Highperformance Computing (NCHC) for providing computational and storage resources.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent