search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.76k stars 7.96k forks source link

random=1 can't use #5054

Open ShaneHsieh opened 4 years ago

ShaneHsieh commented 4 years ago

I can not train to use random. CFG:yolo_v3_tiny_pan3_aa_ae_mixup_scale_giou.cfg.txt

I use the newest version,command line use

darknet.exe detector train 20191230Person\Person.data yolo_v3_tiny_pan3_aa_ae_mixup_scale_giou.cfg yolov3-tiny.conv.15

It will show error: Error: Different size of input layers: 368 x 368, 736 x 736

I has checked code , it is [route] layer print when resize_network. But i don't know how to fix it,

ShaneHsieh commented 4 years ago

I found some issue.

When random = 1 , will to do resize network(function:resize_network). But when it resize [maxpool] layer it used antialiasing=1, it doesn't resize the convolutional layer that maxpool layer made. And when it resize the next layer , it will get the maxpool's output size rather than the convolutional layer output size.

AlexeyAB commented 4 years ago

antialiasing=1 is deprecated. Just remove antialiasing=1 from cfg-file.

ShaneHsieh commented 4 years ago

OK ,it is work. thank you.

ShaneHsieh commented 4 years ago

But the performance and mAP will be same when i remove antialiasing=1?

antialiasing=1 is mean?

AlexeyAB commented 4 years ago

Yes