Error in forward() Expected a value of type 'Tuple[Tensor, Tensor]'

C++
1

Hello,
trying to test libtorch with Qt following this useful answer from PyTorch forum, but when it comes to forward I’m catching the following exception:

forward() Expected a value of type 'Tuple[Tensor, Tensor]' for argument 'inputs' but instead found type 'Tensor'.
Position: 1
Declaration: forward(__torch__.detectron2.export.caffe2_modeling.Caffe2RetinaNet self, (Tensor, Tensor) inputs) -> ((Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor))

  • The loaded model is a detectron2 exported via caffe_tracer.export_torchscript().
  • Caffe2Tracer input was just an image of torch.Size([3, 128, 80])
  • The input tensor used to forward the network has shape [3, 128, 80], matching the tensor using to export the model.
  • The input is then unqueezed to shape [1, 3, 128, 80]

My question:
How can I make a proper input tensor (Tuple[Tensor, Tensor]) from my image?

Thanks.

1 Like
2

I’m also struggling with this problem. Have you found a solution? I will appreciate any suggestions.

3

You need to use toTuple() with the result of forward instead of toTensor(). And then use toTensor() with each tuple’s element. Below is sample code.

auto tuple = model.forward(input_to_net).toTuple();
auto output = tuple->elements()[0].toTensor();