该仓库用于UW-Madison GI Tract Image Segmentation比赛
{0: {0: 0.9931173288276431, 1: 0.0068826711723568796, 'ratio': 144.29242716350245}, 1: {0: 0.9936987980163507, 1: 0.006301201983649225, 'ratio': 157.69988021251598}, 2: {0: 0.9965100078185938, 1: 0.0034899921814062367, 'ratio': 285.53359320623633}}
conda create -n mmseg-kaggle python=3.10 -y
conda activate mmseg-kaggle
# conda install pytorch=1.11.0 torchvision cudatoolkit=11.3 -c pytorch
pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.11.0/index.html
git clone https://github.com/zezeze97/kaggle_segmentation.git
cd {path of project}
pip install -e .
从官网下载好数据集后,放在该项目的input目录下,运行kaggle_segmentation/prepare_data.ipynb
# 训练
bash run.sh train $GPU
# 测试
bash run.sh test $GPU
kaggle_segmentation/inference_demo.ipynb
- 2.5d data: 同一个case,同日的3张slice拼接成一张(stride=2)
- mutilabel问题,最后激活使用sigmoid而不是softmax!!
- 图片case的相关性,更好的建模方式?
- 图片尺寸较小, 可以尝试upernet origin size?(默认1/4大小)
- 实验结果整理
- 增加smp unet decoder 移植 https://github.com/CarnoZhao/Kaggle-UWMGIT/blob/kaggle_tractseg/mmseg/models/segmentors/smp_models.py
- swin transformer v2
- 更好的数据增强方式:https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/331450
- 5张slice拼接成一张,需要修改pretrained ckpts的第一层...