About converting torch.max(x,dim=1)[0] in C++

C++
1

Hi, All

I have inquiry about torch.max(x,dim=1)[0] conversion to C++.
But in torch::max(x)[0], there is only one argument allowed. How do I make sure the conversion is the same and apply it to dim=1 please ?

Any suggestion is welcome ?

Thanks,

2

Hi,

There are multiple overloads of max: https://pytorch.org/cppdocs/api/function_namespaceat_1ac66d13a86d2ee2c974cd7d22de2ec409.html?highlight=max
In particular you want the one where you can specify the dim.
Note also that [0] won’t work with the tuple<> output and you will need to use std::get<0>(the_tuple) IIRC.

3

Thanks a lot.

I modified the code to be torch::max(x,1 false) but it is still incorrect
“No viable conversion from returned value of type ‘std::tuple<Tensor, Tensor>’ to function return type ‘torch::Tensor’”

4

See the second part of my comment above: you need to unpack the tuple values using std::get<0>().

5

Thanks for the guide.

std::get<0>(max(x,1,false)) seems to work.

There is another with torch.max(x1,1)[1].data
In this case, is
std::get<1>(max(x1,1,false)) equivalent to torch.max(x1,1)[1].data
because std::get<1>(max(x1,1,false)).data() is also valid.

Just to confirm

6

Hi,

In general, you should never use .data. So you should change the python version of that code first :smiley:

7

May I ask what is the best alternative for .data option ?
Will it cause issues in c++ ?

Thanks

8

It will cause issues on the python side as well.
If you want a Tensor with the same content but that does not share autograd history, you should use .detach() (and the same c++ function when porting).