SOLOv2

This implements training of SOLOv2 on the Coco dataset, mainly modified from pytorch/examples.

SOLOv2 Detail

As of the current date, Ascend-Pytorch is still inefficient for contiguous operations. Therefore, SOLOv2 model need to be modified in the following aspects:

Converting tensors with the dynamic shapes into tensors with fixed shapes. (This is the hardest one)
Several operations, like the sum of INT64, are not supported on the NPU, so we modified tensors' dtype when needed
Framework bottlenecks lead to poor performance, so we improve the original code to improve the performance of the model
We used Apex for mmdtection due to the hardware defects of the NPU
...

Requirements

NPU配套的run包安装
Python 3.7.5
PyTorch(NPU版本)
apex(NPU版本)
MMCV v0.2.16

Document and data preparation

下载压缩modelzoo\contrib\PyTorch\cv\instance_segmentation\SOLOv2 文件夹
于npu服务器解压SOLOv2压缩包
下载coco2017数据集
下载预训练模型，下载地址：https://download.pytorch.org/models/resnet50-19c8e357.pth，下载完成后重命名为resnet50.pth
将coco数据集放于SOLOv2/data目录下，目录结构如下：

GFocalV2
├── configs
├── data
│   ├── coco
│       ├── annotations   796M
│       ├── train2017     19G
│       ├── val2017       788M
│   ├── pretrained
│       ├── resnet50.pth  98M

Configure the environment

进入SOLOv2目录，source环境变量
cd SOLOv2
source test/env_npu.sh

配置安装mmcv

cd mmcv
python3 setup.py build_ext
python3 setup.py develop
cd ..
pip list | grep mmcv  # 查看版本和路径

配置安装mmdet

pip install -r requirements/build.txt
pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"
pip install -v -e .

Train MODEL

进入SOLOv2目录下

1p

导入环境变量，修改train_full_1p.sh权限并运行

chmod +x ./test/train_full_1p.sh
bash ./test/train_full_1p.sh --data_path=./data/coco

8p

导入环境变量，修改train_full_8p.sh权限并运行

chmod +x ./test/train_full_8p.sh
bash ./test/train_full_8p.sh --data_path=./data/coco

Eval

修改train_eval_1p.sh权限并运行

chmod +x ./test/train_eval_1p.sh
bash ./test/train_eval_1p.sh --data_path=./data/coco

finetuning

修改train_finetune_1p.sh权限并运行

chmod +x ./test/train_eval_1p.sh
bash ./test/train_finetune_1p.sh --data_path=./data/coco

SOLOv2 training result

Acc@1	FPS	Npu/Gpu_nums	Epochs	AMP_Type	Loss_Scale
18.4	8.1	1p Gpu	1	O1	128.0
34.8	38.3	8p Gpu	12	O1	128.0
18.4	1.1	1p Npu	1	O1	128.0
34.3	6.2	8p Npu	12	O1	128.0