AlexeyAB
commented
6 years ago Hi,
- On what cfg file based your cfg-file:
yolo-voc.cfgoryolo-voc.2.0.cfg, or some other? - What command line do you use for training?
- Did you test every your image using Yolo_mark, i.e. can you see correct boxes for each images in your dataset using: https://github.com/AlexeyAB/darknet
- Did you use OpenCV 3.3.0, CUDA 8.0 and cuDNN 5.1?
Hey,
While training on my own data, I encountered some strange behavior. Every time Obj and No Obj are going to zero really fast. The rest of the training scores appears to be ok, but after any number of steps the result is always the same - no detection.
It is so strange that both avg loss and Recall are looking right.
I tought that this is probably a problem with my data, but when I trained the network based on the data from this article https://timebutt.github.io/static/how-to-train-yolov2-to-detect-custom-objects/ the results are the same.
Additionaly if am trying to run detector on CPU it produces some silly output (a lot of results some with score > 100%)
For comparison, I have trained the net on the original repository. After 2000 steps, the detector is working properly and the Obj value still increases.
Currently I am training network on my own data but the statistics also look good.
I am using Ubuntu 16.10 with two GTX 1080. I tried run this on both cards at once, single card and on CPU. Every time training does not produce any valuable effect.
If i can attach some more information please give me know.