Darknet: YOL) compiles and runs, but no predictions

[ramanpreet9]

I compiled yolo with GPU, CUDA, and opencv on.
It compiles correctly. it runs on the sample images like dog.jpg etc as described in wiki. However, I am not getting any predictions!

My output:
Case 1 and 2 (GPU ON, CUDA ON/OFF, opencv on
Loading weights from yolo.weights...Done!
data/dog.jpg: Predicted in 0.056 seconds.

and it hangs here

case 3: GPU ON, CUDA, OPENCV OFF
data/dog.jpg: Predicted in 0.17 seconds.
Not compiled with OpenCV, saving to predictions.png instead

but the predictions have no BB.

One thing i noticed was that the link for yolo weights is ~200MB but the wiki page says its 1 GB.

[konbick]

I am having exactly the same problem and also the same observation about the file size.

Were you able to solve this problem meanwhile?

[ramanpreet9]

The yolo wights have an issue as i mentioned in my post. If you use the small or tiny weights, it works fine. but, until the author responds, you cannot use the full weights.

[konbick]

Man, thanks a lot for your post! You saved me a lot of headache!

[konbick]

Oh, no. It doesn't work for me even with the tiny weights...

[ramanpreet9]

try both tiny and small. it worked for me.

[konbick]

Should I use a different cfg? Could you let me know your command when trying to infere a test image? Thanks

[ramanpreet9]

./darknet yolo test cfg/yolov1/yolo-small.cfg /weights/yolo-small.weights
data/dog.jpg -thresh 0.1

On Mon, Mar 19, 2018 at 5:00 PM, konbick [e-mail] wrote:

Should I use a different cfg? Could you let me know your command when
trying to infere a test image? Thanks

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/pjreddie/darknet/issues/486#issuecomment-374142233,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AMbPE_l0L7C6NdHqMWlBGVtIjjht4Ns6ks5tf3OSgaJpZM4SPV7y
.

[ramanpreet9]

Yes, you need to use the relevant config file small.weights with small.cfg etc

[konbick]

If I use the following command, it works well for me, too

./darknet detect cfg/yolo.cfg weights/yolo.weights data/horses.jpg

[kjcx100]

I had the same problem as u,I test the cmd “./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg” but no response。Then I changed to "./darknet detect cfg/yolov2.cfg yolo.weights data/dog.jpg" ,It's works。I'm wondering why！！！

[csbrown]

I am having the same problem, except I cannot detect the dog for any combination of weights and architectures.

CUDA 8.0
K80
CUDNN 7.1
Nvidia drivers 375.66

GPU=1, CUDNN=1

[csbrown]

Upgraded my nvidia drivers and cuda. No better.

CUDA 9.1
K80
CUDNN 7.1
Nvidia drivers 390.67

GPU=1, CUDNN=1

[csbrown]

Lowered the -thresh option to 0.01, and it detects everything, everywhere at a confidence of 5%. This is weird, but I suppose this is because 0%s cause issues with cross-entropy loss, so the minimum confidence is 5%..... okay. If I set the -thresh option to 0.06, I detect like a gazillion people, one for each grid cell, it seems. Going to add a cli option for -hier_thresh and see what that does so I can get a handle on all this to dig deeper to debug.

[csbrown]

Turned off CUDNN and it seems to work. What version of CUDNN is everyone using?

[zhangyi6]

yeah! I have the same problem, and I turned off CUDNN, and it worked, but I don't know why, is the version of CUDNN wrong?
CUDA -> 8.0
NVIDIA drive -> 375.26
(seem like I have two version of CUDNN, 5.1.10 and 7.2.1)

[grosmar]

Any update?

I have the same issue with Cuda 10, if I disable CUDNN it works

[Tzuya14]

I encountered the same problem while working on GCP and I can confirm that it works every time while disabling cudnn but I also made it work with cudnn enabled just by creating a new IDENTICAL virtual machine with same cuda and cudnn, which is a bit weird. This may show that is not a problem with the weights.

[ministrudels]

I encountered this problem using the NVIDIA GPU Cloud Image for Deep Learning, Data Science, and HPC image on GCP. Turns out, it was because I was using the NVIDIA Tesla V100 GPU. I wasnt having this issue on a new instance with the NVIDIA Tesla P100.

[SirojbekSafarov]

opencv=1
cuda=1
cudnn=1

same error, please modify cfg/yolov3.cfg,

Testing

batch=1
subdivisions=1

Training

batch=256

subdivisions=64

then, detector result display ok