How to reduce running time of the Intel resnet50 benchmark?

[eltimen]

Hi all,

I want to launch resnet50 pytorch-cpu benchmark on the my server with Intel CPU. When I run it in Docker following the instructions in readme bash run_offline.sh 8, it runs >5 hours in the Offline scenario. I think the measurements are taking too long for this case. Is this duration expected from this benchmark?

I have also tried the alternative way to launch it using the automation scripts I was able to run the benchmark after fixing some minor problems (like IndentationError at this line) but the measurements also take too long on the my hardware.

I tried decreasing the values of parameters number_cores and *.Offline.target_qps in the file user_default.conf inside the container, but it had no effect on the measurements duration. Can anybody advise me how to reduce the running time of this benchmark or tell me what I'm doing wrong?

[arjunsuresh]

Hi @eltimen, how many cores are there in the system? Also are you trying to do an official submission or just trying out the performance?

Even at 100 QPS, R50 offline scenario should take only 10 minutes. If it's taking longer, it means wrong qps is being used to run the model.