search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.77k stars 7.96k forks source link

average loss can not be less than 2.xx #4923

Open hamidfathi1998 opened 4 years ago

hamidfathi1998 commented 4 years ago

hey, I want to make a custom object detection and I have 32 classes. I edited yolov3.cfg file like be AlexeyAB part ( How to train (to detect your custom objects): ) and use darknet53.conv.74 for the convolutional layers that is my project details ( data cfg file ) plateNumber log image

AlexeyAB commented 4 years ago

average loss can not be less than 2.xx

This is normal

hamidfathi1998 commented 4 years ago

average loss can not be less than 2.xx

This is normal this is not good results average loss should be less than 0.06! how to get better results you can see my data

AlexeyAB commented 4 years ago

It doesn't matter what is the absolute value of avg loss.

Look at the mAP.

stephanecharette commented 4 years ago

that is my project details ( data cfg file ) plateNumber log image

Heh...take a look at this: https://www.ccoderun.ca/programming/ml/iranian_plates.html

hamidfathi1998 commented 4 years ago

that is my project details ( data cfg file ) plateNumber log image

Heh...take a look at this: https://www.ccoderun.ca/programming/ml/iranian_plates.html

yeh I see that but my detect so be different I add other plate (taxi, government plate, ...)

hamidfathi1998 commented 4 years ago

It doesn't matter what is the absolute value of avg loss.

Look at the mAP.

thanks my training model was be stoped and I starting training, but my average loss value is 2000 and mAP value begins with zero (first training model mAP value is 50 percent ) that is correct ???

stephanecharette commented 4 years ago

From the few images I saw on the link you provided, I think your images are much too small to train a network. Random example: 177.jpg, your entire image is only 133x41, so the individual letters on the plates are only 3-5 pixels wide. You'll need larger images.