Eric Liu
Eric Liu
It is recommend to set resize width and height with 0 (keep original size) when produce lmdb .
Not in training , It's here , https://github.com/eric612/MobileNet-YOLO/blob/master/data/VOC0712/create_data.sh#L14
可以參考以下連結 https://github.com/eric612/MobileNet-YOLO/issues/47 https://github.com/PiyalGeorge/MobileNet-YOLO/tree/master/models/yolov3
看不出來確切原因，但我直覺應該是資料方面的問題，建議你可以在訓練與測試過程中打印出來圖跟bounding box
另外我看你的solver ，test只有6張，似乎有點少，如果連training的圖也是這麼少，應該是這個問題
因為我GPU記憶體不夠，所以test batch size我都只用1 ，大於1的我沒測試過，問題應該出在這行 https://github.com/eric612/MobileNet-YOLO/blob/master/src/caffe/layers/yolov3_detection_output_layer.cpp#L344-L387 裡面的bbox 都塞到第一個batch裡，改掉就會正常
training iter size 建議設 32
1. https://github.com/eric612/MobileNet-YOLO/issues/9 2. Caffe已經是沒有再更新的framework ，裡面使用的cuda與相關矩陣的運算已經過時了
It is a good works , I will try to implement the layers , thanks .
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