Open athulvingt opened 4 years ago
It depends on the size of your dataset. Train 4x times more iterations than you have images.
I have 15,000 images 114,000+ annotations of 34 classes. so is it 15,000 x 4 or 114,000 x 4
60 000 - 70 000
does the image aspect ratio has something to do with training or inference? I did a ocr test on vehicle number plate, the number was "PC \n 111 AC" , it recognised only '11' of '111' , but when i cropped and took only the the lower part ( only '111AC ', excluding 'PC'), it recognised correctly
Yolo is great and really admire your work. I have few doubts, You have mentioned that ideal iterations is class*2000. But, config files in your repo the max_batch is 500,000 for a 80 class yolo network