安裝

conda create --name env_lian_semask python=3.7
conda activate env_lian_semask
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=10.1 -c pytorch
pip install mmcv-full==1.2.0 -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.6.0/index.html
git clone https://github.com/ananzeng/STAS-Detection-Competition-on-Pathological-Section-Images-of-Lung-Adenocarcinoma-II-Using-Image-Seg.git
pip install -e .

下載預訓練權重(semask_large_fpn_ade20k.pth) 下載STAS權重(iter_60000.pth)

mkdir checkpoint
cp semask_large_fpn_ade20k.pth checkpoint
cp iter_60000.pth work_dirs/semfpn_semask_swin_large_patch4_window12_640x640_80k_ade20k

使用TTA 推論模型 將此行取消註解 將此行註解 輸出的結果位在work_dirs/com_Private_Public/semfpn_semask_swin_large_patch4_window12_640x640_80k_ade20k_ITER60000_1st

python tta_com.py

產生訓練資料 將比賽方提供的資料夾SEG_Train_Datasets做預處理產生ground truth以及ADE20K資料格式的資料集

python make_gt_image.py

進行訓練 將此行註解 將此行取消註解 產生的權重檔案會在./work_dirs/semfpn_semask_swin_large_patch4_window12_640x640_80k_ade20k

python tools/train.py configs/semfpn_semask_swin_large_patch4_window12_640x640_80k_ade20k.py --seed 69 --deterministic