Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

trt model conf is negative #39

Open
lxzatwowone1 opened this issue Aug 19, 2022 · 18 comments
Open

trt model conf is negative #39

lxzatwowone1 opened this issue Aug 19, 2022 · 18 comments

Comments

@lxzatwowone1
Copy link

I train a single class model by yolov7 and conver trt.
python3 yolov7/export.py --grid --simplify
python3 T_F_Y_S/export.py -o xxx.onnx -e xxx.engine -p fp16 --end2end
when I use trt model inference, the boxes and class are same as pth model, but conf is negative and always in [-0.4,-0.6]
can you help me,thanks!!!

@Linaom1214
Copy link
Owner

Linaom1214 commented Aug 19, 2022

I train a single class model by yolov7 and conver trt. python3 yolov7/export.py --grid --simplify python3 T_F_Y_S/export.py -o xxx.onnx -e xxx.engine -p fp16 --end2end when I use trt model inference, the boxes and class are same as pth model, but conf is negative and always in [-0.4,-0.6] can you help me,thanks!!!

provide your model file (onnx model)

@lxzatwowone1
Copy link
Author

I train a single class model by yolov7 and conver trt. python3 yolov7/export.py --grid --simplify python3 T_F_Y_S/export.py -o xxx.onnx -e xxx.engine -p fp16 --end2end when I use trt model inference, the boxes and class are same as pth model, but conf is negative and always in [-0.4,-0.6] can you help me,thanks!!!

prevoid your model file (onnx model)

好的,感谢,您意思是我转onnx时候有问题么,prevoid 这个单词我没明白,看您地址是西安,索性我就用中文了TT

@Linaom1214
Copy link
Owner

Linaom1214 commented Aug 19, 2022

I train a single class model by yolov7 and conver trt. python3 yolov7/export.py --grid --simplify python3 T_F_Y_S/export.py -o xxx.onnx -e xxx.engine -p fp16 --end2end when I use trt model inference, the boxes and class are same as pth model, but conf is negative and always in [-0.4,-0.6] can you help me,thanks!!!

prevoid your model file (onnx model)

好的,感谢,您意思是我转onnx时候有问题么,prevoid 这个单词我没明白,看您地址是西安,索性我就用中文了TT

提供一下您的模型文件, 我看一下 linaom1214@163.com

@lxzatwowone1
Copy link
Author

我大概找到了问题,pth转onnx的时候,也就是
python3 yolov7/export.py --grid --simplify这一步后 onnxruntime推理输出的置信度score已经不对了,看起来好像都减半了但是细看也不是减半,但是用您的代码转也并没有得到和onnx同样的置信度结果,百思不得其解TT

@lxzatwowone1
Copy link
Author

新的发现!
image
zhe'bu'f这部分代码貌似有问题,我在尽力修改,稍有成效

@Linaom1214
Copy link
Owner

end2end 和 不包含 nms的模型差距很大吗? 之前我测试的是好的

@lxzatwowone1
Copy link
Author

您是用几类模型测的

@Linaom1214
Copy link
Owner

您是用几类模型测的

80 类 coco

@lxzatwowone1
Copy link
Author

官方那个不行的

@Linaom1214
Copy link
Owner

官方那个不行的

????

@Linaom1214
Copy link
Owner

我重新测试了一次,多目标、单目标 代码都是好的呀,端到端和不包含nms的模型是可以对齐的, 也许是多Batch的原因,我能想到的就只有这了,等我搞完论文好好研究一下。

@lxzatwowone1
Copy link
Author

感谢 我这边还在尝试

@Linaom1214
Copy link
Owner

感谢 我这边还在尝试

确实是 多batch的原因 我在更新动态batch的时候,发现转换会出错, 您 那边有进展吗

@ZhuYoukang320
Copy link

我也遇到分数为负数的情况,检查了很多次代码都没问题。后来偶然发现是FP16时才会出现负的分数,FP32结果是正常的

@Linaom1214
Copy link
Owner

我也遇到分数为负数的情况,检查了很多次代码都没问题。后来偶然发现是FP16时才会出现负的分数,FP32结果是正常的

FP16 会有一定的精度损失

@jiguanglu
Copy link

fp16的score与原始pt推理的score的和基本上是1,很奇怪。

@EdisionWew
Copy link

我也遇到分数为负数的情况,检查了很多次代码都没问题。后来偶然发现是FP16时才会出现负的分数,FP32结果是正常的

你这个问题解决了吗,我这边也遇到了非负的情况,可以交流下吗

@ZhuYoukang320
Copy link

我也遇到分数为负数的情况,检查了很多次代码都没问题。后来偶然发现是FP16时才会出现负的分数,FP32结果是正常的

你这个问题解决了吗,我这边也遇到了非负的情况,可以交流下吗

没有解决呢,可能是tensorrt nms插件的问题吧。FP16的话我最后是放弃了端对端nms的方式,自己做后处理

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

5 participants