Darknet: Is anyone who trained your own data?

[springkim]

Detection is well good using Pretrained model(yolov3.weights).

But, When i use my trained model it was not working. (super bad location)

I trained model around 10000 iterations. and i got 0.02 avg loss.

I don't know why it wans't work.

My data works well at YOLO v2. I mean, the data quality is very good.

BTW, pjreddie wrote max_batches = 500200 in yolov3.cfg.

Do i need to more iteration? like this? 500200 ??? It's too many...

I think the train interface not changed.

[AlexeyAB]

I trainded Yolov3 5000 iterations and result is well: https://github.com/AlexeyAB/darknet/issues/504#issuecomment-377290060

There are several differences for training Yolo v3, you should change classes and filters in the 3 places in cfg-file - read more: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

[abeyang00]

Whats the minimum size of the ROIs to be set in annotation text file?
I think mine isnt training well :/

[eyebies]

Mine seems to train okay but I have a lot of NANS

[AlexeyAB]

@abeyang00 Size of truth bounded box should not be less than 1/1000 of image width or height.

[yjdeveloper]

@AlexeyAB I have trained YOLOv2 model using Google Colab. After 250 iterations, my avg loss reaches 0.6XXXX and doesn't decrease further. Is my network performing well?