float16 溢出问题

Closed this issue 3 days ago · 2 comments

qingyang-jiandao commented a month ago

只要模型里面有数值超过65519，就当作inf处理，造成最后推理结果错误，麻烦解决一下这个bug

nihui commented a month ago

可以禁用整个模型的fp16推理
https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-produce-wrong-result#disable-fp16

也可以针对容易溢出的层单独禁用fp16推理
https://github.com/Tencent/ncnn/wiki/layer-feat-mask#disable-fp16-for-certain-layer-to-fix-overflow

qingyang-jiandao commented a month ago

好的，谢谢了