search

dog-qiuqiu / Yolo-Fastest

:zap: Based on yolo's ultra-lightweight universal target detection algorithm, the calculation amount is only 250mflops, the ncnn model size is only 666kb, the Raspberry Pi 3b can run up to 15fps+, and the mobile terminal can run up to 178fps+
Other
1.98k stars 427 forks source link

inference costs 29ms 1080TI #13

Open DLTensor opened 3 years ago

DLTensor commented 3 years ago

I followed your README.md, modified Makefile to open CUDA CUDNN and GPU, and then run "make" I used ./Yolo-Fastest/VOC/yolo-fastest.weights to run image_yolov3.sh, it costs 29ms on 1080TI GPU. I do not know where is wrong looking forward to your reply

dog-qiuqiu commented 3 years ago

On some GPUs (such as NVIDIA PASCAL: 1080ti, 1070...), Darknet Group convolution is not well supported, which will cause the problem of low training inference efficiency, but it will not appear on the 20 series and 16 series graphics cards, for example The reasoning time for 2080ti is 2ms, and 1660ti is 3ms. It is suspected to be the cause of CUDNN. It is recommended that users in this situation use pytorch for training and inference

DLTensor commented 3 years ago

ok，I  will try Thanks for your reply

---Original--- From: "dog-qiuqiu"<notifications@github.com> Date: Wed, Sep 16, 2020 23:51 PM To: "dog-qiuqiu/Yolo-Fastest"<Yolo-Fastest@noreply.github.com>; Cc: "DLTensor"<1523452426@qq.com>;"Author"<author@noreply.github.com>; Subject: Re: [dog-qiuqiu/Yolo-Fastest] inference costs 29ms 1080TI (#13)

On some GPUs (such as NVIDIA PASCAL: 1080ti, 1070...), Darknet Group convolution is not well supported, which will cause the problem of low training inference efficiency, but it will not appear on the 20 series and 16 series graphics cards, for example The reasoning time for 2080ti is 2ms, and 1660ti is 3ms. It is suspected to be the cause of CUDNN. It is recommended that users in this situation use pytorch for training and inference

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.