Different Outputs in Python and Torchscript C++

[Keunhoi_An]

Maybe you’re suffering the same problem as me.

Why libtorch cannot get 'running_mean' and 'running_var' of BatchNormalization2D in .pt file?

I’ve found that the result of libtorch is very different from pytorch result. I save my model as .pt file using pytorch and load by libtorch, and all parameters are successfully copied which I double checked by loading it by pytorch again. After checking the module.parameter() in libtorch, I’ve found that all of BatchNormalization2D modules of libtorch do not contain running_mean and running_var. And also, after the first conv layer, all of parameters are not the same as original value. I’m …