Open smgrz opened 2 years ago
@smgrz see this answer: https://github.com/WongKinYiu/yolor/issues/103#issuecomment-927617539
hi did you solve it successfully? Could you please help me to fix a problem?
@Mobu59 I solved the problem of training only one class by following the answer in https://github.com/WongKinYiu/yolor/issues/103#issuecomment-927617539. If that's your problem, that should do it.
@Mobu59 I solved the problem of training only one class by following the answer in #103 (comment). If that's your problem, that should do it. I want to know what the format of the data set should be ,I make two folders: dataset/images/train, dataset/images/val,and put txts which including imgnames and xywh respectively,but when I train, cannot find dataset
Take a look at how coco's dataset is organized: https://cocodataset.org/#download I downloaded the val2017 images and labels and organized my dataset the same way. From root directory, it looks something like this:
58 0.389578 0.416103 0.038594 0.163146
62 0.127641 0.505153 0.233312 0.222700
62 0.934195 0.583462 0.127109 0.184812
56 0.604656 0.632547 0.087500 0.241385
Then train with the yaml pointing to your dataset txt files (the ones directly in your root). It should look something like this: https://github.com/WongKinYiu/yolor/blob/main/data/coco.yaml
Good luck!
Take a look at how coco's dataset is organized: https://cocodataset.org/#download I downloaded the val2017 images and labels and organized my dataset the same way. From root directory, it looks something like this:
- val2017.txt: list of paths to validation images
- images/val2017: directory containing all images listed in txt file above
- labels/val2017: directory containing list of txt files with the same name as images, but only with .txt extension rather than .jpg. These txt files are themselves a list of the class number with xywh. For example:
58 0.389578 0.416103 0.038594 0.163146 62 0.127641 0.505153 0.233312 0.222700 62 0.934195 0.583462 0.127109 0.184812 56 0.604656 0.632547 0.087500 0.241385
Then train with the yaml pointing to your dataset txt files (the ones directly in your root). It should look something like this: https://github.com/WongKinYiu/yolor/blob/main/data/coco.yaml
Good luck!
Thanks for your help!I‘ll try it tomorrow
is other loss function or activation function needed? or changing filters and classes to 18 and 1 is enough?
Training works with only one class, but when I take the trained model in detect.py the class (e.g. person) is not recognized!
Do I need to change some parameters somewhere?