Darknet: mAP count stopped working (cudnn error)

[dkubatin]

Good day, @AlexeyAB !
After the last commits mAP ceased to be considered.

• Videocard: RTX 2080Ti
• CUDA Version: 10.2
• OpenCV version: 3.2.0
• cuDNN: 7

He rolled back to the commit from Feb 20 and everything works.

[AlexeyAB]

This seems like a hardware error message. Try again with the last commit.

[dkubatin]

mAP is still not considered.
I checked all the commits, the error is in the March 11 commit.
Link to commit: commit

_It was at this commit that this error appeared.
Checked on two devices._

[AlexeyAB]

Do you get the same error message in all cases?
Do you use the same directory on HDD for all commits?

Checked on two devices.

On 2 different computers?
Or on 1 computer with 2 GPUs?

[dkubatin]

Do you get the same error message in all cases?

Yes.

Do you use the same directory on HDD for all commits?

No, on different computers, with different drives

[AlexeyAB]

Show content of obj.data file

[dkubatin]

Show content of obj.data file

classes=1
train=/home/dkubatin/learn_plate_20/train.txt
valid=/home/dkubatin/learn_plate_20/test.txt
names=/home/dkubatin/learn_plate_20/metadata/classes.names
backup=/home/dkubatin/learn_plate_20/backup

[AlexeyAB]

Can you successfully run object detection on video or calc_anchors?

./darknet detector calc_anchors obj.data -num_of_clusters 9 -width 416 -height 416

[dkubatin]

calc_anchors works, but the demo on the video does not work, writes: _Video-stream stopped!_

[AlexeyAB]

the demo on the video does not work, writes: Video-stream stopped!

Can you show screenshot?

[AlexeyAB]

Set in obj.data file

train=/home/dkubatin/learn_plate_20/test.txt

run calc_anchors and show me screenshot of result

[dkubatin]

[AlexeyAB]

Try to run mAP calculation with flag -cuda_debug_sync

[dkubatin]

He doesn’t count with the flag either, however, after the pull of recent changes, the map calculation works both with and without the flag.

[AlexeyAB]

Is problem solved?

[dkubatin]

Yes!