Giters

AlexeyAB / Yolo_mark

GUI for marking bounded boxes of objects in images for training neural network Yolo v3 and v2

Home Page:https://github.com/AlexeyAB/darknet

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Question about native resolution of the network for yolov4.conv.137 file and anchors size validity

acobo opened this issue · comments

commented

Hello all,
I will appreciate your help with this question: most tutorials and examples for retraining yolov4 use pretrained weights from the yolov4.conv.137 file .... which is the original network used to obtain those weights? the "complete" weights files at the model zoo all of them specify the native resolution for training (416, 512, 608 ...), but I cannot find it for this partial weights file.
I understand that using the same resolution in my dataset , in the network for retraining, and in the pretrained wedights should give the best results?

Additionally, I am wondering what happend to the anchors if the size of my dataset is not the same as the size of the network. I have calculated my anchors for a 608x608px dataset, if the network's size is 416 and all images are resized, are the anchors still valid?

Thanks in advance for your help!
ADolfo.