conda create -n vw python=3.8
conda install pytorch=1.13.1 torchvision=0.14.1 pytorch-cuda=11.3 -c pytorch -c nvidia
pip install mmcv==1.7.1
pip install -v -e .
More Utilization: See MMSegmentation Docs.
Swin Transformer
tools/dist_train.sh configs/swin/CONFIG.py NUM_GPUS --work-dir work_dirs/EXP_NAME
ConvNeXt
tools/dist_train.sh configs/convnext/CONFIG.py NUM_GPUS --work-dir work_dirs/EXP_NAME
Single GPU
python tools/test.py path/to/config.py path/to/weights.pth --eval mIoU
Multiple GPUs
tools/dist_test.sh path/to/config.py path/to/weights.pth NUM_GPUS --eval mIoU
| Name | Backbone | crop size |
lr sched |
mIoU | mIoU (ms+flip) |
download |
|---|---|---|---|---|---|---|
| VW | Swin-B | 640x640 | 160k | 52.5 | 53.5 | model |
| VW | Swin-L | 640x640 | 160k | 54.4 | 55.8 | model |
| VW | ConvNeXt-T | 512x512 | 160k | 47.3 | 48.3 | model |
| VW | ConvNeXt-S | 512x512 | 160k | 48.8 | 49.9 | model |
| VW | ConvNeXt-B | 640x640 | 160k | 53.3 | 54.1 | model |
| VW | ConvNeXt-L | 640x640 | 160k | 54.3 | 55.1 | model |
| VW | ConvNeXt-XL | 640x640 | 160k | 54.6 | 55.3 | model |
| Name | Backbone | crop size |
lr sched |
mIoU | mIoU (ms+flip) |
download |
|---|---|---|---|---|---|---|
| VW | ConvNeXt-T | 512x1024 | 160k | 81.3 | 82.1 | model |
| VW | ConvNeXt-S | 512x1024 | 160k | 82.2 | 83.1 | model |
| VW | ConvNeXt-B | 512x1024 | 160k | 83.2 | 83.9 | model |
| VW | ConvNeXt-L | 512x1024 | 160k | 83.4 | 84.1 | model |
| VW | ConvNeXt-XL | 512x1024 | 160k | 83.6 | 84.3 | model |
@inproceedings{yan2023multi,
title={Multi-Scale Representations by Varing Window Attention for Semantic Segmentation},
author={Yan, Haotian and Wu, Ming and Zhang, Chuang},
booktitle={The Twelfth International Conference on Learning Representations},
year={2023}
}