Hi, I have a question, what the num_class and num_classes mean?

[shuxiao9058]

As titled, for example, I want to train only 1 type of object, how can I set the numbers in train and test protoxt files? 2 or 1?

[eric612]

You can set classes = 1 , and prev convolution channel = (5+1)*3 , if loss function is yolov3

[shuxiao9058]

I'm a beginner with DL, now I work with yolo lite prototxt you provide.
Thanks for your reply.

[shuxiao9058]

How can I set the prev convolution channel? Do you have any document about this? @eric612

[eric612]

You can search the text in test and train prototxt and replace
num_output: 75 to num_output: 18

[shuxiao9058]

Thanks, I will have a try. How about the num_output with yolov2?

I have search on the net, and got:
pjreddie/darknet#561
pjreddie/darknet#323
https://www.bbsmax.com/A/VGzlVjQy5b/

So if trained with yolov2:
coords=4
num_class=1
(classes+ coords+ 1)* (NUM)=(1+ 4+ 1)*5=30

Is it right? and the NUM is the num in RegionLoss layer.

[shuxiao9058]

I trained with yolov3-lite, but got, what does this mean.

I0905 05:47:23.277906 21668 yolov3_layer.cpp:344] avg_noobj: 9.66286e-05 avg_obj: -nan avg_iou: -nan avg_cat: -nan recall: -nan recall75: -nan

[eric612]

It means there has no ground truth box match one of the yolo loss , so it will return null