dog-qiuqiu / FastestDet

:zap: A newly designed ultra lightweight anchor free target detection algorithm, weight only 250K parameters, reduces the time consumption by 10% compared with yolo-fastest, and the post-processing is simpler
BSD 3-Clause "New" or "Revised" License
768 stars 131 forks source link

转成ncnn之后模型精度差好多 #28

Closed 523997931 closed 1 year ago

523997931 commented 1 year ago

大神您好, 我这边训练了一个自己的3类的检测模型,直接评测的情况尚可,但是转换ncnn模型并且用大神给的example去进行推理,发现检测结果差别很大,不知道这个是什么原因呢

523997931 commented 1 year ago

验证了一下,好像转onnx之后精度就变了,不太清楚为什么,就是用test脚本直接做的导出

523997931 commented 1 year ago

我这边看了一下,发现转之前的输出是82222(输出是3类),但是转onnx之后输出变成了852222(coco的类别数80类),感觉应该是这个原因,不知道这是怎么导致的,为什么输出维度转换之后发生了变化