training data. - Githubissues

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

training data. #1627

Open BlcaKHat opened 6 years ago

BlcaKHat commented 6 years ago

I am training data for detecting helmet for 1000 images. I am training on CPU , as i don't have GPU. I left overnight and average loss is 4.8. any suggestion on when should i stop the training or how much time it will take . train

AlexeyAB commented 6 years ago

Train not less than 4000 iterations.

BlcaKHat commented 6 years ago

alex, after 850 iteration, the average loss become 0.0001. but no detection from weights . shall I keep training ?

BlcaKHat commented 6 years ago

tr_out tr_ter

@AlexeyAB shall I keep training, it is not detecting anything, just showing inference time on image.

BlcaKHat commented 6 years ago

@AlexeyAB , my PC restarted after 3700 rep. and still no detection. what you suggest.? shall i train it more ?

BlcaKHat commented 6 years ago

cfg.txt obj_name.txt weight darknet53.conv.74

BlcaKHat commented 6 years ago

i used these files to train. where did i went wrong

AlexeyAB commented 6 years ago

Train yolov3-tiny.cfg without darknet53.conv.74 : https://github.com/AlexeyAB/darknet#how-to-train-tiny-yolo-to-detect-your-custom-objects
What params do you use in the Makefile?
How many training images do you have?
What anchors can you get by this command? ./darknet detector calc_anchors data/obj.data -num_of_clusters 6 -width 416 -height 416
Can you show your bad.list and bad_label.list files if they exist?
Check your dataset by using: https://github.com/AlexeyAB/Yolo_mark

BlcaKHat commented 6 years ago

anchors.txt bad_label_list.txt makefile.txt makefile.txt

number of classes = 1.( Helmet) number of images = 700 ( source : openImage)

calculating k-means++ ... avg IoU = 48.48 %

BlcaKHat commented 6 years ago

@AlexeyAB how do i check my data. and what is bad_label.list.?

AlexeyAB commented 6 years ago

@BlcaKHat

Your anchors are totaly wrong, because anchors can't be larger than network resolution 416x416: 41105.2578,39776.2500, 134823.9062,45003.3398, 234606.1562,43598.8945, 77814.1484,156373.4531, 348901.2500,79930.6875, 224374.4688,151252.9844
Therefore your dataset is totaly wrong, because you have such anchors, and because you have too many errors in the bad_label_list.txt. The file bad_label_list.txt contains wrong labels from your dataset with x,y,w,h >0 or <= 0
Open your dataset in the https://github.com/AlexeyAB/Yolo_mark and see that all your labels are wrong

BlcaKHat commented 6 years ago

@AlexeyAB is it necessary to rename the image files according to class names ?

BlcaKHat commented 6 years ago

the name of the images are different from class name, is it a reason for bad_label ?

AlexeyAB commented 6 years ago

No.

Look inside bad_label.list file:

Wrong label: data/obj/00001bcc92282a38.txt - j = 0, x = 385.279999, y = 308.823914, width = 186.240005, height = 58.304977
Wrong label: data/obj/0003d1891fd14575.txt - j = 0, x = 222.720001, y = 172.190445, width = 177.919998, height = 139.544418

Your x = 385.279999 is higher than 1. But it should be higher than 0 and lower than 1: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

- float values relative to width and height of image, it can be equal from (0.0 to 1.0]

BlcaKHat commented 6 years ago

yeah.. I downloaded the dataset from openImage. and they're giving autonation file like this. what you suggest?, shall I do it manually for all 7k images .? or is there a way to convert that is this format.?

AlexeyAB commented 6 years ago

You should write your own script for convertion annotations files to the Yolo format.