search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.72k stars 7.96k forks source link

about training my own dataset #1342

Open franklu323 opened 6 years ago

franklu323 commented 6 years ago

hi, I'm trying to use my own dataset to train the yolov3. I have 1900 image in total and 5 classes and I set batch=64 and sub=32. right now, it seems right in training. but I am confused that for each iteration, I got some nan in avg iou, class.... it like capture I don't know that my training is going into the right way or not?? and what is the 'count' for. because in my training, count is like only 1 or 2. can you help me out?? Is the 'count' also will infect the result or just the iterations are still too low (right now it just 430). I searched that 'count' means that the number of TP samples. so if the meaning is right, is my count number like really low or not?? if it's really low, how can I improve my count?? thank you so much @AlexeyAB

AlexeyAB commented 6 years ago

Training goes well

https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

Note: If during training you see nan values for avg (loss) field - then training goes wrong, but if nan is in some other lines - then training goes well.

franklu323 commented 6 years ago

hi @AlexeyAB , I am still confused that 'count' meaning and my count number is ok or less. I stopped my training at 700 iterations (just want to have a look), and it seems can detect my training dataset. I am also wondering, my dataset is from the videoes frame and I only chose the first 100 frames pre video (for example) to do the training, when I use the rest video frame, it can be detected but with some error. but when I download a related image from the internet. my training network cannot detect anything. can you help me to sort it out, like what caused this problem? thank you so much.

AlexeyAB commented 6 years ago

@franklu323 Hi, Your count number is well. But you should use more images and images should be more diverse.

Train more than 4000 iterations and do this recommendations: https://github.com/AlexeyAB/darknet#how-to-improve-object-detection