I trained Dense Net in Pytorch, set model.eval() and moved the model to C++ following the tutorial. But in C++ when i do the inference, the result is different from the result in Pytorch(In Python), and I already do the same normalization.
It is scary. For me, having the assumption of using python for the training/studying and C++ for the inference with same results and easy model sharing was one of the main points to choose PyTorch over other frameworks.
@xiaofanwei Could you share with us the steps you take to reproduce this problem? I can help debug this.