None Default Arguments

[rusty1s]

Hello and thanks for this great repository!

Can you give me a hint on how to implement default arguments? I tried the following so far but it does not work:

at::Tensor torch_func(at::Tensor tensor1, at::Tensor tensor2={}) {
  ...
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
  m.def("torch_func", &torch_func, py::arg("tensor1"), py::arg("tensor2")={});
}

torch_func(tensor1, None)

I would be very grateful for help.

• OS: macOS
• PyTorch version: 0.4.0
• How you installed PyTorch (conda, pip, source): source
• Python version: 3.6
• CUDA/cuDNN version: -
• GPU models and configuration: -
• GCC version (if compiling from source): 4.9

[goldsborough]

Hi, we don't currently provide a way of creating a "void"/"None" Tensor in our C++ extension API. Would it be enough for you to use something like at::empty(torch::CPU(at::kDouble), {0})); as the default argument and then check with .numel() == 0 whether or not it is empty? This would work both in the function prototype and for the py::arg() = ... .

Let me know whether this would be sufficient for you.

[rusty1s]

Thanks! python setup.py install now works perfectly.

However, calling the function in Python with a None argument gives the following error message:

incompatible function arguments. The following argument types are supported:
1. (tensor1: at::Tensor, tensor2 = tensor([], dtype=torch.float64)) -> at::Tensor

Invoked with: tensor([2, 2]), None

[goldsborough]

We don't have any conversion from None to a C++ type (it's questionable what it would convert to), but you can just omit the argument. I assume that's the default argument behavior you want? Or do you specifically want to pass None? On Apr 26, 2018, at 18:58, Matthias Fey <notifications@github.com<mailto:notifications@github.com>> wrote: Thanks! python setup.py install now works perfectly. However, calling the function in Python with a None argument gives the following error message: incompatible function arguments. The following argument types are supported: 1. (tensor1: at::Tensor, tensor2 = tensor([], dtype=torch.float64)) -> at::Tensor Invoked with: tensor([2, 2]), None — You are receiving this because you commented. Reply to this email directly, view it on GitHub<#5 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AGIcm7TFFmZu8AKHezCc4mRq3kdYVacvks5tsnskgaJpZM4TmVvM>.

[rusty1s]

Yeah, I see. Actually, I wanted to pass None, but I guess I'm fine with the proposed solution (unless you know of an easy way to allow for None arguments). Thanks again!

[goldsborough]

I will keep the possibility of None converting to an undefined Tensor in mind as we expand the extension API. Thanks for posting the issue! I'll mark it as resolved.

[apaszke]

@goldsborough how about we expose torch::optional and let Nones bind to that?

[goldsborough]

Yes that's a good idea, I'll do that

[goldsborough]

@rusty1s you can now pass None if you use at::optional<at::Tensor> in C++ (and at::nullopt or nullptr as the C++ default type)