the .pth model maybe have some operator which can't appropriate haisi arm.
for example:
when you write some operator named like:
weight = self.aaa(abc)[:,:,0,0]; # u easy to see that i use the slicing with a function together .
you know,it's easy to understanding for pth,python,onnx,pytorch..
but ,the point is that haisi .om can't understanding this operation,haisi don't know some complicated operation to together used.acctully,what ever complicated or easy operation, i think if you use some operation at one python line, even it easy to understanding for python,but the haisi arm can't understanding exactly. if you ignore this point, maybe you not received some bug info.everything going to continue.But the final result is not same as the pth's result.
so for solve this problem.you have to write python code detailed. for example:
weight = self aaa(abc);
weight = weight.squeeze(-1);
weight = weight.squeeze(-1);
done..