aovoc / nnieqat-pytorch

A nnie quantization aware training tool on pytorch.
MIT License
238 stars 40 forks source link

量化后对于小目标检测得分特别低 #5

Closed northeastsquare closed 2 years ago

northeastsquare commented 4 years ago

你好,我用darknet训练了yolov3后,转换到nnie,用nnie sample,修改类别为两类(人,人头),跑起来后,发现对于大目标person预测正常,但是对于小目标,人头,得分特别低,很多只有0.0?, 我把阈值设置为0.01才把这些框取出来。而在darknet转到caffe后,得分在0.3以上,一切正常。 请问这种现象,是nnie的量化方式导致的吗?使用您的代码,训练后,这种现象会好吗?谢谢

MuhammadAsadJaved commented 3 years ago

@northeastsquare Hi, Have you tried yolov3 quantization using this method?

northeastsquare commented 3 years ago

have not

aovoc commented 3 years ago

yolov5在coco上进行量化实验,使用nnieqat量化包能够提升小目标检测效果,实验结果见readme.

northeastsquare commented 3 years ago

谢谢,@aovoc 如果在Hi3559上不使用nnieqat量化包的结果在coco上跑一下,和使用过nnieqat的结果进行对比,更能看到效果

northeastsquare commented 3 years ago

其实我也不确定,对于有小目标,要使用0.01的阈值获取result,是因为 caffe模型使用ruyi studio量化,我哪里做错了?其实我也是再四检测类别数量、通道、yolo输出层的设置; 还是因为海思量化过程本身导致的。

tensorflowt commented 3 years ago

其实我也不确定,对于有小目标,要使用0.01的阈值获取result,是因为 caffe模型使用ruyi studio量化,我哪里做错了?其实我也是再四检测类别数量、通道、yolo输出层的设置; 还是因为海思量化过程本身导致的。

请问,你有尝试作者的方法到海思的芯片上吗?