[Training Issue] Multi-scale training on YOLOv2

[ryusaeba]

Model : yolov2

Backbone : Darknet19

Questions : Hi @eric612 ,

When using the multi-scale for YOLOv2 training, it will pop the following message. concat_layer.cpp:42] Check failed: top_shape[j] == bottom[i]->shape(j) (19 vs. 13) All inputs must have the same shape, except at concat_axis. The shape mismatch will have 19 v.s. X, which is depends on the random pick from which one resolution. If disable the multi-scaling feature, the issue will be gone.

The same Annotation layer used in YOLOv3 won't have this error.

Have you ever met this problem before?

[eric612]

The reorg layer would not dynamic change blob size when training , it may make some dim error .

https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/layers/reorg_layer.cpp

it was copied from here https://github.com/gklz1982/caffe-yolov2/blob/master/src/caffe/layers/reorg_layer.cpp

[ryusaeba]

Hi @eric612 , Thanks for the quick response, so how do you workaround this when using multi-scale trick for YOLOv2?

[eric612]

You can reference this code , and modify the reorg layer source code https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/layers/yolov3_layer.cpp#L287-L289