faster-rcnn speed and accuracy about channel-pruning HOT 8 CLOSED

from channel-pruning.

Hi,I would like to ask why the speed did not improve it? Even if I used cudnn to speed up, the speed did not improve anything.

from channel-pruning.

you can compare these logs, it indeed improves:
https://github.com/yihui-he/py-faster-rcnn/blob/master/experiments/logs/faster_rcnn_end2end_VGG16_.txt.2017-01-03_15-27-04
https://github.com/yihui-he/py-faster-rcnn/blob/master/experiments/logs/vggx2.txt
https://github.com/yihui-he/py-faster-rcnn/blob/master/experiments/logs/vggx4.txt

from channel-pruning.

Thank you for your reply, but when i use your open source code for VGGx4 test results for an average of 0.148s per picture, VGGx2 is 0.180s, i use cuda8.0, cudnn5.1, gtx1080 machine, is it because i am Careless so I ignored what? thanks for your help.

from channel-pruning.

IDK

from channel-pruning.

https://www.google.com/search?q=IDK

from channel-pruning.

Of course, cudnn is faster...

from channel-pruning.

Hi, yesterday I am wrong, the use of cudnnV7 slower than not used, I experiment on the 1080 ,results are as follows:
use cudnn not use cudnn
VGGx4: mAP 0.6553 speed 0.143s/each picture mAP 0.654 speed 0.089s/each picture
VGGx2: mAP 0.6671 speed 0.158s/each picture mAP 0.6639 speed 0.101s/each picture

from channel-pruning.