Create pre-trained weights for detection without darknet / torch

Question

I want to train my data and create a weights file to train my YOLOV3 network. The objects in my database are not in COCO's classes so I do not want to use their weights file. I also have a limitation - I can not use darknet and I also prefer not to use the torch library.

The files I have: (according https://github.com/AlexeyAB/Yolo_mark/issues/60#issuecomment-401854885)

yolov3-custom.cfg according my classes
obj.names with my class's names
train.txt + test.txt with list of image's location
folder with images+ labels yolo format (object-class, x, y, width, height)
obj.data

What can I do?

I think you have no option except to use darknet like in [this question](https://stackoverflow.com/questions/57898577/how-to-reduce-number-of-classes-in-yolov3-files). — AbdelAziz AbdelLatef, Nov 18 '20 at 01:34
Why you cannot use darknet? maybe this issue can be fixed so you can train your weights — Alexander Riedel, Nov 18 '20 at 10:01

score 1 · Answer 1 · answered Nov 20 '20 at 14:43

1

You can still use the pre-trained weights on ImageNet if you want to start with pre-trained weights. If you have different classes than the COCO dataset that's no problem. You can define your own classes, and start training with the pre-trained weights. During training, the weights will be updated to detect your custom classes.

You can use these weights for YOLOv3: "https://pjreddie.com/media/files/yolov3.weights"

answered Nov 20 '20 at 14:43

Y.Ynot

337
5
15

I started training my custom with the pre-trained weights of yolov3.weights. In the beginning the loss was high but after all fixed I trained my custom according to https://github.com/qqwweee/keras-yolo3 – HELO Nov 23 '20 at 13:53

score 0 · Accepted Answer · answered Nov 23 '20 at 13:57

0

I trained my custom according to github.com/qqwweee/keras-yolo3

answered Nov 23 '20 at 13:57

HELO

67
8

I started training my custom with the pre-trained weights of yolov3.weights. In the beginning the loss was high but after all fixed I trained my custom – HELO Nov 23 '20 at 13:57

Create pre-trained weights for detection without darknet / torch

2 Answers2