Tiny yolov3
purvikpatel opened this issue · 2 comments
Can I use tiny yolov3 cfg file with Gaussian Yolov3?
And if yes what changes I have to make to train images with tiny yolov3 cfg file?
@purvikpatel yes, you can. Clone this repo, then change layer name [yolo] to [Gaussian_yolo] in the config file and recalculate the filter numbers according to the gaussian layer. In normal yolo format, the calculation is:
filters = (4 box coordinates + objectness score + number of classes) * number of anchor boxes
eg. for general yolo architecture training on coco dataset (80 classes):
filters = (4 + 1 + 80) * 3 = 255
For gaussian yolo, the only change is the number of box coordinates, it's not 4 for gaussian yolo, but it's 8. So the calculation for the same configuration on coco dataset is:
filters = (8 + 1 + 80) * 3 = 267
I hope it helps :)
@canyilmaz90
Nice explanation!
Thanks.