Different Results on GPU vs CPU

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

http://pjreddie.com/darknet/

Other

21.68k stars 7.96k forks source link

Different Results on GPU vs CPU #6528

Open VisionEp1 opened 4 years ago

VisionEp1 commented 4 years ago

see last comment- better problem description

VisionEp1 commented 4 years ago

update it happens also with the normal pre trained yolov4 file yolov4.cfg.txt

i used the most recent opencv version + the most recent darknet.

windows with cmake (one time i disabled cuda+cudnn via cmake the other time i didnt)

4-5% confidence difference seems huge to me

VisionEp1 commented 4 years ago

or with lower confidence. even more different

VisionEp1 commented 4 years ago

after days of testing.

avx is the issue after forcing is_avx to -1 everything works perfect. might be linked with see issue: #5793

truongbv1 commented 5 months ago

Hi, @AlexeyAB I see another bug that causes different results on GPU and CPU GPU=1 - working properly GPU=0 - not working properly with GPU=0, you missed the case of RELU6 in float activate(float x, ACTIVATION a), if use RELU6 activation func then all return 0