TransDA

KO | EN

gitlite — search

#hacktoberfest #ai #typescript #golang #machine-learning #deep-learning #linux #macos #windows #vue #awesome-list #pytorch #openai #mcp

(★ 20)

This is an official implementation for "Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation"

TransDA 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation

This repo contains the supported code and configuration files to reproduce semantic segmentaion results of [TransDA](https://arxiv.org/abs/2203.07988).

### [Paper](https://arxiv.org/abs/2003.00273)

> **Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation**<br>
>
> **Abstract** *After the great success of Vision Transformer variants (ViTs) in computer vision, it has also demonstrated great potential in domain adaptive semantic segmentation. Unfortunately, straightforwardly applying local ViTs in domain adaptive semantic segmentation does not bring in expected improvement. We find that the pitfall of local ViTs is due to the severe high-frequency components generated during both the pseudo-label construction and features alignment for target domains. These high-frequency components make the training of local ViTs very unsmooth and hurt their transferability. In this paper, we introduce a low-pass filtering mechanism, momentum network, to smooth the learning dynamics of target domain features and pseudo labels. Furthermore, we propose a dynamic of discrepancy measurement to align the distributions in the source and target domains via dynamic weights to evaluate the importance of the samples. After tackling the above issues, extensive experiments on sim2real benchmarks show that the proposed method outperforms the state-of-the-art methods.*

## Author
Runfa Chen, Yu Rong, Shangmin Guo, Jiaqi Han, Fuchun Sun, Tingyang Xu, Wenbing Huang

## Citation
If you find this code useful for your research, please cite our paper:

```
@misc{chen2022smoothing,
    title={Smoothing Matters: Momentum Transformer for Domain Adaptive Semantic Segmentation},
    author={Runfa Chen and Yu Rong and Shangmin Guo and Jiaqi Han and Fuchun Sun and Tingyang Xu and Wenbing Huang},
    year={2022},
    eprint={2203.07988},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}
```

### Prerequisites
- Python 3.8
- Pytorch 1.7.1
- torchvision from master
- yacs
- tqdm
- matplotlib
- GCC >= 4.9
- OpenCV
- CUDA >= 10.1
- mmcv >= 1.3.0
- cityscapesscripts

### Datasets

- Download [The GTA5 Dataset]( https://download.visinf.tu-darmstadt.de/data/from_games/ )

- Download [The SYNTHIA Dataset]( http://synthia-dataset.net/download/808/ )

- Download [The Cityscapes Dataset]( https://www.cityscapes-dataset.com/ )

- Generate the label statics file for GTA5 and SYNTHIA Datasets by running 
```
python datasets/generate_gta5_label_info.py -d datasets/gta5 -o datasets/gta5/
python datasets/generate_synthia_label_info.py -d datasets/synthia -o datasets/synthia/
```

### Train & Validate & Test
We provide the script using 4 Tesla V100 GPUs.

TransDA-S
```
bash bin_small.sh
```
TransDA-B
```
bash cls_base.sh
```

### Resume and test from our best checkpoint

* checkpoint:  model_last.pth (You need to save to './pth/CHECKPOINT_FOLDER_NAME/model_last.pth')
* Swin ViT weight can be downloaded from [Swin Transformer for ImageNet Classification]( https://github.com/microsoft/Swin-Transformer ) and export weight = YOUR_WEIGHT_PATH.
* TransDA-S checkpoint can be downloaded from [GTA5toCityscapes](https://drive.google.com/drive/folders/1kU1AAZpiOIVUiM0z8xc3s0yCYKsLty8I?usp=sharing ) and [SYNTHIAtoCityscapes]( https://drive.google.com/drive/folders/142ioi5tTEATp2bVbapTB1k6Z0asTwPFV?usp=sharing ).
```
export base_dir=.
export dataset=_gta5
# export dataset=_synthia
export weight=https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_small_patch4_window7_224.pth
# validate
python test.py -cfg configs/train$dataset.yaml --saveres data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/val resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth MODEL.WEIGHTS $weight SOLVER.DIS binary
# test for submit
python test.py -cfg configs/train$dataset.yaml data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/test resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth  DATASETS.TEST cityscapes_test  MODEL.WEIGHTS $weight SOLVER.DIS binary
```

* TransDA-B checkpoint can be downloaded from [GTA5toCityscapes]( https://drive.google.com/drive/folders/16zLe32ZRU524VCpGGL4-NcGez2fMeS1R?usp=sharing ) and [SYNTHIAtoCityscapes]( https://drive.google.com/drive/folders/1tjqr69Etz-1M8L5Xvdpubbg127Yu32u5?usp=sharing ).
```
export base_dir=.
export dataset=_gta5
# export dataset=_synthia
export weight=https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22k.pth
# validate
python test.py -cfg configs/train$dataset.yaml --saveres data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/val resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth MODEL.WEIGHTS $weight SOLVER.DIS class
# test for submit
python test.py -cfg configs/train$dataset.yaml data_dir $base_dir/datasets OUTPUT_DIR $base_dir/datasets/cityscapes/test resume $base_dir/pth/CHECKPOINT_FOLDER_NAME/model_last.pth  DATASETS.TEST cityscapes_test  MODEL.WEIGHTS $weight SOLVER.DIS class
```

### Visualization results

![Visualization](https://raw.githubusercontent.com/alpc91/TransDA/main/figure/seg3.png)
![UMAP](https://raw.githubusercontent.com/alpc91/TransDA/main/figure/UMAP.png)

### Acknowledgments
This project is based on [FADA](https://github.com/JDAI-CV/FADA) and [Swin Transformer for Semantic Segmentaion](https://github.com/SwinTransformer/Swin-Transformer-Semantic-Segmentation). We thank them for their excellent open-source projects.

TransDA

TransDA 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

TransDA

TransDA 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README