Skip to content

Latest commit

 

History

History

seaformer-seg

SeaFormer

Requirements

  • pytorch 1.5+
  • mmcv-full==1.3.14

Main results

Results on ADE20K

Model Params FLOPs mIoU(ss) Link
SeaFormer-T_512x512_2x8_160k 1.7 0.6 35.0 Baidu Drive, Google Drive
SeaFormer-T_512x512_4x8_160k 1.7 0.6 36.5 Baidu Drive, Google Drive
SeaFormer-S_512x512_2x8_160k 4.0 1.1 38.9 Baidu Drive, Google Drive
SeaFormer-S_512x512_4x8_160k 4.0 1.1 39.5 Baidu Drive, Google Drive
SeaFormer-B_512x512_2x8_160k 8.6 1.8 41.2 Baidu Drive, Google Drive
SeaFormer-B_512x512_4x8_160k 8.6 1.8 41.9 Baidu Drive, Google Drive
SeaFormer-L_512x512_2x8_160k 14.0 6.5 43.0 Baidu Drive, Google Drive
SeaFormer-L_512x512_4x8_160k 14.0 6.5 43.8 Baidu Drive, Google Drive

Results on Cityscapes

Model FLOPs mIoU Link
SeaFormer-S_1024x512_1x8_160k 2.0 71.1 Baidu Drive, Google Drive
SeaFormer-S_1024x1024_1x8_160k 8.0 76.4 Baidu Drive, Google Drive
SeaFormer-B_1024x512_1x8_160k 3.4 72.2 Baidu Drive, Google Drive
SeaFormer-B_1024x1024_1x8_160k 13.7 77.7 Baidu Drive, Google Drive
  • The password of Baidu Drive is seaf

Usage

Please see mmsegmentation for dataset prepare.

clone the repository locally:

git clone https://github.com/fudan-zvg/SeaFormer
cd SeaFormer/seaformer-seg
mkdir -p modelzoos/classification

For training on SeaFormer_Tiny, run:

cp /cls_outdir/last.pth.tar modelzoos/classification/SeaFormer_T.pth
sh tools/dist_train.sh local_configs/seaformer/<config-file> <num-of-gpus-to-use> --work-dir /path/to/save/checkpoint

We use 8 gpus by default. If you use fewer gpus, you will need to increase the batch size to ensure that the total batch size remains the same.

To evaluate, run:

sh tools/dist_test.sh local_configs/seaformer/<config-file> <checkpoint-path> <num-of-gpus-to-use>