Xiangrong Chen

> 您好， 我用官方给的yolov5s.pt转为tnn，测试结果正常。 但把自己训练的pt转为tnn后，精度下降很多（跟pytorch版本对比），这是哪里的原因呢？ [https://github.com/ppogg/shufflev2-yolov5](https://github.com/ppogg/shufflev2-yolov5) 试一下这个把兄弟,我也想知道这个模型转TNN或者MNN会不会崩

Sure~

https://github.com/ppogg/Retinaface_Ghost/blob/master/weights/mobilev3_Final.pth here~

Hi friends, this link has all the data sets you want, I trained based on these data： [https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB) 

You can try to open the data set to see the format, you will find that his **train.txt** contains this format: 0--Parade/0_Parade_marchingband_1_465.jpg 345 211 4 4 331 126 3 3...

Hello, my friend, I think what you said is that evaluate.py needs to be used _mobilev3_txt_ and _ground_truth_ In fact, _mobilev3_txt_ is the key point position after model prediction, and...

Yes, it is indeed a data problem. Are you your own data set?

OK, I get it. May be the label.txt is not imported. Is the path checked? You can print out the content of label.txt to see if it is loaded or...

Thanks for using YOLOv5 Lite！In fact, this is caused by an overly high version of pytorch. Please follow the requirements.txt to install BTW，the torch version in my env is 1.8.0.

Here you are~