Before running
- install numactl：
apt-get install numactl # for Ubuntu
yum install numactl # for CentOS
- get R-50.pkl：
mkdir -p /root/.torch/models/
wget https://dl.fbaipublicfiles.com/detectron/ImageNetPretrained/MSRA/R-50.pkl
mv R-50.pkl /root/.torch/models/
- ln -s dataset：
mkdir ./dataset
ln -snf path_to_coco ./dataset/coco
- other requirements：
pip3 install torchvision==0.2.1
# other recommended requirements
apex==0.1+ascend.20220315
torch==1.5.0+ascend.post5.20220315
- source env and build：
source test/env_npu.sh
Running
- To train：
# 1p train full
bash test/train_full_1p.sh --data_path=./dataset/
# 1p train perf
bash test/train_performance_1p.sh --data_path=./dataset/
# 8p train full
bash test/train_full_8p.sh --data_path=./dataset/
# 8p train perf
bash test/train_performance_8p.sh --data_path=./dataset/
- To evaluate:
bash test/train_eval_1p.sh --data_path=./dataset/ --weight_path=./model_0044999.pth # for example
Result
1p batch_size == 8，8p batch_size == 64
| NAME | Steps | BBOX-MAP | SEGM-MAP | FPS |
|---|---|---|---|---|
| GPU-1p | 360000 | - | - | 8.7 |
| GPU-8p | 20000 | 29.0 | 25.7 | 55.1 |
| NPU-1p | 400 | - | - | 4.6 |
| NPU-8p | 20000 | 28.8 | 25.7 | 34.8 |
公网地址说明
代码涉及公网地址参考 public_address_statement.md