/mfpsg

mask2former psg

Primary LanguagePythonApache License 2.0Apache-2.0

Mask2Former PSG

整体框架还是基于mmdet，其中relation部分借鉴transformers


Install

环境参考 get_started.md

pip install -e .

apex 建议通过下面代码安装

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

panopticapi

pip install git+https://github.com/cocodataset/panopticapi.git

数据准备

下载 coco instance val 2017，用于验证 psg val 的instance map

修改 wqx/main.py 中下列文件路径

if __name__ == '__main__':
    # raw data file
    raw_psg_data='/share/data/psg/dataset/for_participants/psg_train_val.json'
    raw_coco_val_json_file='/share/data/coco/annotations/instances_val2017.json'

    # output file
    output_coco80_val_instance_json = '/share/wangqixun/workspace/bs/psg/psg/data/instances_val2017_coco80.json'
    output_tra_json='/share/wangqixun/workspace/bs/psg/psg/data/psg_tra.json'
    output_val_json='/share/wangqixun/workspace/bs/psg/psg/data/psg_val.json'
    output_val_instance_json='/share/wangqixun/workspace/bs/psg/psg/data/psg_instance_val.json'

执行

python wqx/main.py

(可能)需要的一些预训练权重

分割部分:

mask2former， 预训练权重

关系分类：

transformers， 预训练权重


训练

  • 调整 configs/psg/mask2former_r50.py 中预训练路径、输出路径、tra val 数据路径
# 模型中预训练部分
model['relationship_head']['pretrained_transformers'] = "/path/chinese-roberta-wwm-ext"
load_from = "/path/mask2former_r50_lsj_8x2_50e_coco-panoptic_20220326_224516-11a44721.pth"
# tra 数据部分
data['train']['ann_file'] = 'data/psg_tra.json'
data['train']['img_prefix'] = '/path/psg/dataset/'
data['train']['seg_prefix'] = '/path/psg/dataset/'
# val 数据部分
data['val']['ann_file'] = 'data/psg_val.json'
data['val']['img_prefix'] = '/path/psg/dataset/'
data['val']['seg_prefix'] = '/path/psg/dataset/'
data['val']['ins_ann_file'] = 'data/psg_instance_val.json'
# test 数据部分
data['test']['ann_file'] = 'data/psg_val.json'
data['test']['img_prefix'] = '/path/psg/dataset/'
data['test']['seg_prefix'] = '/path/psg/dataset/'
data['test']['ins_ann_file'] = 'data/psg_instance_val.json'
# 输出路径
work_dir = 'output/v0'
  • 训练咯
# 8卡训练
bash tools/dist_train.sh configs/psg/mask2former_r50.py 8

Submit

  • 调整 wqx/infer_p.py 中 cfg 和 ckp
if __name__ == '__main__':
    get_val_p(
        mode='v0',
        cfg='configs/psg/mask2former_r50.py',
        ckp='output/v0/latest.pth',
    )
  • 推理
mkdir -p /share/wangqixun/workspace/bs/psg/mfpsg/submit/val/submission/panseg
python wqx/infer_p.py
  • 打包
cd submit/v0
zip -r submission.zip submission/