CascadedMaskRCNN
本项目实现了 CascadedMaskRCNN 在 NPU 上的训练, 迁移自 detectron2-0.2.1. detectron2 github链接
CascadedMaskRCNN Detail
本项目对 detectron2-0.2.1 做了如下更改：
- 迁移到 NPU 上
- 使用混合精度训练、测试
- 对于一些操作，固定动态 shape 、使用 NPU 算子优化性能、同时将一些操作转移到 CPU 上进行
Requirements
- NPU 配套的 run 包安装(20211018)
- Python 3.7.5
- PyTorch(NPU20210930 版本)
- apex(NPU20210930 版本)
- torchvision 0.6.0
- decorator
- sympy
- 安装 detectron2
conda create -n cascaded-mask python=3.7.5
conda activate cascaded-mask
pip install torch-1.5.0+ascend.post3.20210930-cp37-cp37m-linux_x86_64.whl
pip install apex-0.1+ascend.20210930-cp37-cp37m-linux_x86_64.whl
pip install -r CascadedMaskRCNN/requirements.txt
source CascadedMaskRCNN/test/env_npu.sh
python -m pip install -e CascadedMaskRCNN
- 下载 COCO 数据集，放在 datasets 中。如已有下载可通过设置环境变量DETECTRON2_DATASETS=“coco 所在数据集路径”进行设置，如 export DETECTRON2_DATASETS=/opt/npu/，则 coco 数据集放在 /opt/npu/ 目录中
- 下载预训练模型 R-50.pkl ,configs/Misc/cascade_mask_rcnn_R_50_FPN_1x.yaml配置文件中MODEL.WEIGHTS 设置为R-50.pkl的绝对路径
Training
cd CascadedMaskRCNN
conda activate cascaded-mask
# training 1p accuracy
bash ./test/train_full_1p.sh --data_path=/opt/npu/
# training 1p performance
bash ./test/train_performance_1p.sh --data_path=/opt/npu/
# training 8p accuracy
bash ./test/train_full_8p.sh --data_path=/opt/npu/
# training 8p performance
bash ./test/train_performance_8p.sh --data_path=/opt/npu/
#test 8p accuracy
bash test/train_eval_8p.sh --data_path=/opt/npu/ --pth_path=./output/model_final.pth
# finetuning 1p
bash test/train_finetune_1p.sh --data_path=/opt/npu/ --pth_path=./output/model_final.pth
Log path: test/output/devie_id/train_device_id.log # training detail log test/output/devie_id/CascadedMaskRCNN_bs128_8p_perf.log # 8p training performance result log test/output/devie_id/CascadedMaskRCNN_bs128_8p_acc.log # 8p training accuracy result log
CascadedMaskRCNN training result
| Acc@1 | FPS | Npu_nums | Epochs | AMP_Type |
|---|---|---|---|---|
| - | 3 | 1 | 12.17 | O2 |
| 36.255 | 25 | 8 | 24.34 | O2 |
公网地址说明
代码涉及公网地址参考 public_address_statement.md