How to retrieve number of outputs of a model?

C++
1

Hi!

I tried m_ptrModule->get_method(“forward”).graph()->outputs().size();
But that returns 1 also for a model with two outputs.

Any help would be appreciated.

Regards,

Roos

2

Hi Roos,

Very likely, the Python inheritance of TorchScript shows, i.e. that functions pretending to be returning multiple values actually return a tuple.

So

import torch

@torch.jit.script
def d(x):
    return x,x

print(d.graph)

gives

graph(%x.1 : Tensor):
  %3 : (Tensor, Tensor) = prim::TupleConstruct(%x.1, %x.1)
  return (%3)

So a proper check would check whether auto tt = graph()->output()->type()->cast<TupleType>() and then check tt->elements().size() or somesuch. The elements() also has the types of the returns if it matters.

Best regards

Thomas

3

Hi Thomas,

Thanks for your answer. Unfortunately your solution does not compile, e.g.
'output() is not a member a torch::jit::Module and more.

In the end I solved it by passing dummy input through the module like this:

    auto outputs = m_ptrModule->forward(inputs);
    if (outputs.isTuple())
    {
        m_nNumOutputs = outputs.toTuple()->elements().size();
    }
    else
    {
        m_nNumOutputs = 1;
    }

But personally I think this is a hack. I can ask the number of inputs:

m_ptrModule->get_method("forward").num_inputs();

But not the number of outputs…

Greetings,

Roos

4

Sorry, output() is a shorthand for asserting outputs.size()==1 and then using outputs[0] - but for node, not graph… The other bits should still work.

Besteht regards

Thomas

5

Hi Thomas,

Thank you. Now it works:

    auto outputs = m_ptrModule->get_method("forward").graph()->outputs();
    auto asTuple = outputs[0]->type()->cast<c10::TupleType>();
    if (asTuple)
    {
        m_nNumOutputs = asTuple->elements().size();
    }
    else
    {
        m_nNumOutputs = 1;
    }

Greetings,

Roos

1 Like
6

Just try

auto outputs = m_ptrModule->get_method("forward").graph()->outputs().size();