Seeding a torch::Generator in C++

C++
1

In Python, we create and seed a random generator like this

  gen = torch.Generator()
  gen.manual_seed( 42 )

to be used for example in

  x = torch.multinomial( x, num_samples=1, generator=gen )

Now, I would like to do the same thing in C++. So, I tried

  auto gen = torch::Generator();
  gen.manual_seed( 42 );  // Does not exist !!!

but the method manual_seed does not exist.
I found another method called set_current_seed but it crashes at execution!

  auto gen = torch::Generator();
  gen.set_current_seed( 42 );  // Crashes !!!

Any idea ?

1 Like
2

I don’t know why your code crashes, so you would need to try to get the stacktraces etc. as the same code is used e.g. here with an additional lock.

3

It appears that in C++ the torch::Generator() class is just an empty wrapper.
The set_current_seed method is defined as

  void set_current_seed(uint64_t seed) { impl_->set_current_seed(seed); }

and of course, when the impl_ pointer is null, the app craches!

I finally found a default generator equivalent to the Python one.

So, the following code in Python

  gen = torch.Generator().manual_seed( 42 )
  p = torch.rand( 5, generator=gen )
  print(p)
  x = torch.multinomial(p, num_samples=10, replacement=True, generator=gen)
  print(x)

prints exactly the same result as the following C++ code

  auto gen = at::detail::createCPUGenerator( 42 );
  auto p = torch::rand( {5}, gen );
  std::cout << p << std::endl;
  auto x = torch::multinomial( p, /*n_samples*/10, /*replac*/true, gen );
  std::cout << x << std::endl;

outputs

tensor([0.8823, 0.9150, 0.3829, 0.9593, 0.3904])
tensor([3, 1, 0, 3, 2, 1, 2, 0, 2, 3])
1 Like
4

Unfortunately, and surprisingly, if the number of samples is 1 the Python version and C++ version does NOT give the same result! So, to solve that, I ask two samples and keep the first one…

5

I think the non-at::detail:: (because stuff in a library’s detail:: namespace is likely not under any API stability guarantees) method of doing this is with at::make_generator<Impl>, where for CPU generators the at::CPUGeneratorImpl seems like the type to use. E.g.

auto gen = at::make_generator<at::CPUGeneratorImpl>(/*seed*/);
auto ten = torch::randn({2, 3, 4}, gen);

This seems to work nicely and for PyTorch 2.4.1+cpu is matching up with the Python way:

ten = torch.randn((2, 3, 4), generator=torch.Generator())
1 Like