Pytorch 32 onnx model will change the solution every time the output results

Normally onnx The output of the model will change every time , This is an extraordinary , Difficult topics . because , Once the model parameters are exported , In theory, it will not change . and ,paddle、pytorch、tensorflow When the framework exports the model , It is also the default setting model mode The value is eval, That is, set parameters without updating . If not personally experienced , See with your own eyes , Bloggers will not believe that there is such a fantastic thing . It can only be said , I know everything I've met . Normal use standard ai Framework export deployment model , Such a thing can never happen .

The reason why bloggers encounter this kind of thing , Because the early operation show is too 6. Reach out MMSegmentation in , Train in your own way mmseg In the model , then And the magic changes the network to deploy .【 Because bloggers use the magic change network , You can only use it yourself pytorch Built in torch Model transfer onnx, After export onnx The model output becomes unstable . Generally speaking ,mmseg、mmdet All models of the series should use mmdeploy Export 】. stay libtorch Deploy , The output of the model is normal every time 【 Any test , Input unchanged , Output unchanged 】. And more recently , Convert the model to onnx After deployment , Whether it's c++ Next or python Next , For the same input data ,onnxruntime The result of each output is different every time .

1、 State description

The blogger's model is saved as pth（libtorch Format ） when , The output of each load is fixed . And in the conversion to onnx After deployment , For the same input data , There is a slight deviation in the results of each output .