how to solve the error when ./train_darknet_v3.sh about caffe-yolo9000 HOT 8 CLOSED

Your solver.prototxt has some problem. If you use stagelr, you should set
lr_policy: "multifixed"

from caffe-yolo9000.

Thank you for your answer，that is the problem！If I set lr_policy: "multifixed" I must compile your caffe not use my own. So with your tip, I solve it.

from caffe-yolo9000.

I have a problem when I train my own dataset. The log shows as follows:
I0517 10:51:16.226722 13269 region_loss_layer.cpp:512] avg_noobj: 0.473491 avg_obj: -nan avg_iou: -nan avg_cat: -nan recall: -nan class_count: 0
....
and the "-nan" can not be stoped even though I changed the initial base_lr.
So did you encounter the same question and can you help me to solve it?

from caffe-yolo9000.

from caffe-yolo9000.

@oppo62258801 Maybe the data label is bad, you should protect it. Decrease the base lr may also help you.

from caffe-yolo9000.

@choasup
you mean decrease the following values?
stagelr: 0.0001
stagelr: 0.001
stagelr: 0.0001
stagelr: 0.00001

from caffe-yolo9000.

@Designbook1
I have tried to decrease the stagelr. And I even changed the lr_policy from to multifixed to fixed. They all don not work.
My dataset has four classes, but there is the "-nan" problem. But when I try to change four classes to only one class, it still does not work.

from caffe-yolo9000.

@oppo62258801 hi friend, did you resolve the nan issue? I encountered this issue.
I want to use yolo9000 to detect text box in natural scene pictures, but I got below issue, do you have any idea? thanks.
Region Avg IOU: -nan, Class: -nan, Obj: -nan, No Obj: 0.518544, Avg Recall: -nan, count: 0
34: 345.566132, 340.394623 avg, 0.000000 rate, 0.152061 seconds, 34 images
Loaded: 0.000026 seconds

from caffe-yolo9000.