search

sacmehta / EdgeNets

This repository contains the source code of our work on designing efficient CNNs for computer vision
MIT License
411 stars 82 forks source link

the segmentation speed is too slow on Android #28

Closed lawo123 closed 4 years ago

lawo123 commented 4 years ago

I deployed the segmentation model to android by ncnn,the inference cost 400ms,image_size=512*512.How to increase speed？

sacmehta commented 4 years ago

Are you running on CPU? Which task are you running? If its a detection task, you might want to use 300x300 resolution?

lawo123 commented 4 years ago

I run segmentation task on Android gpu,scales = 2.My picture is 1024 ×720,so i retrained model to 512*512 resolution.

sacmehta commented 4 years ago

Could you benchmark different stages (frame reading/processing, model execution, overlay etc.) to understand the bottleneck?

sacmehta commented 4 years ago

I think the issue is solved, so closing now