How long should yolo train for?

[pjryan513]

I am training on about 300 jpg images with about 5 bounding boxes per image on a Tesla K20c. I read that for most people around 200 images they trained for about 10-12 hours on a fairly good gpu. I have been training for close to 24-hours now. Is something wrong or should I just be patient?

[TheMikeyR]

You don't get the detections you want? The training is endless, you can test it on the go and stop training when it doesn't improve anymore or overfit the data.

[pjryan513]

The tutorial here says that after it is done training I should get a yolo-obj_final.weights in the section about training your own data. In the section below the document mentions that after 2000 iteration you should be good. The problem is that the training is still going can I only have up to weights 900, the document said that every 100 iteration it should save so I was expecting to stop once I got weight 2000. Here is my /backup folder

yolo-phallusia_100.weights  yolo-phallusia_600.weights
yolo-phallusia_200.weights  yolo-phallusia_700.weights
yolo-phallusia_300.weights  yolo-phallusia_800.weights
yolo-phallusia_400.weights  yolo-phallusia_900.weights
yolo-phallusia_500.weights  yolo-phallusia.backup

Is something going wrong or do I just not understand how yolo trains?

[ahsan856jalal]

you are referring AlexeyAb's tutorial and using pjreddie's version. That tutorial is subjected to AlexeyAB's fork of darknet which saves weights after 100 iterations . Moreover, he said 2000 iterations per class.

[youqisc]

@pjryan513 You can simply use your backup/yolo-phallusia.backup for testing.

The author setup that way in detector.c, you won't see the next weight file until 10000 iteration, and you can modify it in the detector.c file and recompile.

[KellyChenTue]

Hello, I have a related question here, somehow the model saves only '.backup', I don't see anything like '_100.weights', '_200.weights' that is saved. anyone has clue, what's going on?

many thanks!

[xzessmedia]

Hello, I have a related question here, somehow the model saves only '.backup', I don't see anything like '_100.weights', '_200.weights' that is saved. anyone has clue, what's going on?

many thanks!

i recommend to read before you write! if its your default behaviour, you are propably in the wrong job! since you compile darknet yourself, with a bit of research which is fundamental in development you would be able to answer yourself, since you can read anything in the source ;)

there is nothing to add to @pjryan513 answer!

@pjryan513 You can simply use your backup/yolo-phallusia.backup for testing.

The author setup that way in detector.c, you won't see the next weight file until 10000 iteration, and you can modify it in the detector.c file and recompile.

and by the way people work hard and release their code to the public as a gift - as a contribution for a greater purpose! you should be grateful by not using an issue tracker as a discussion board. many people take the issue tracker as an indicator if it can be used for production. so by abusing the issue tracker as a board, you damage the health of a repository! Unfortunately about 90% of the people don't think about that since they are green behind their ears and their eyes and ears are closed.

honor your work and especially the work of other people! master your tools!

[saman202]

https://link.medium.com/erxZEfUB9ib

[aic2022]

Hi, I used yolov3 darknet to object detection in google colab "!git clone https://github.com/AlexeyAB/darknet" ,i changed the number of classes for all yolo conv layers, and the filter, i changed also the max batch and 80%, 90%, but the output was: detections_count = 0, unique_truth_count = 0
class_id = 0, name = hand , ap = 0.00% (TP = 0, FP = 0)

for conf_thresh = 0.25, precision = nan, recall = nan, F1-score = nan for conf_thresh = 0.25, TP = 0, FP = 0, FN = 0, average IoU = 0.00 %

IoU threshold = 50 %, used 101 Recall-points mean average precision (mAP@0.50) = 0.000000, or 0.00 % Total Detection Time: 0 Seconds why all results = 00?

[naveenvj25]

Hi, I used yolov3 darknet to object detection in google colab "!git clone https://github.com/AlexeyAB/darknet" ,i changed the number of classes for all yolo conv layers, and the filter, i changed also the max batch and 80%, 90%, but the output was: detections_count = 0, unique_truth_count = 0 class_id = 0, name = hand , ap = 0.00% (TP = 0, FP = 0)

for conf_thresh = 0.25, precision = nan, recall = nan, F1-score = nan for conf_thresh = 0.25, TP = 0, FP = 0, FN = 0, average IoU = 0.00 %

IoU threshold = 50 %, used 101 Recall-points mean average precision (mAP@0.50) = 0.000000, or 0.00 % Total Detection Time: 0 Seconds why all results = 00?

I too have the same issue? where you able to solve it?