DAWNBench_Inference

Resnet26d for DAWNBench inference task on ImageNet

We run Resnet26d on Alibaba Cloud ecs.ebman1.26xlarge, which consists of 4 npu core and 104 vCPUs.

The model and inference pipeline are improved by Apsara AI Acceleration(AIACC) team in Alibaba Cloud.

The following instructions show how to achieve the performance that we submitted to DAWNBench step by step.

1. install miniconda and dependencies

   wget https://mirrors.tuna.tsinghua.edu.cn/anaconda/miniconda/Miniconda3-latest-Linux-x86_64.sh
   sh Miniconda3-latest-Linux-x86_64.sh
   conda install python=3.6
   conda install glog=0.3.4
   conda install opencv
   pip install pillow opencv-python
   pip install torch torchvision
   pip install hgai-centos_rel_1.0.4.sp1.whl

2. git clone DAWNBench_Inference code

   git clone https://github.com/ali-perseus/DAWNBench_Inference.git

3. run the following commands to replicate our results submitted to DAWNBench,

   ##resize and crop using torchvision
   ##edit preprocress.py if necessary
   python3 ./preprocress.py
   ##edit build.sh if necessary
   ##build
   sh ./build.sh
   ##run test
   sh ./test.sh

4.Congratulations! the result is as follows:

final inference time : 0.0739278 ms
final Prec@5: 0.93156