Why can't I add default-argument to my lambda function parameters (or is there an alternative way to do it)?

Question

So I have a method with simple lambda functions I use to update my weights and I want to try different functions but I also want to have default parameter for the decay:

void ema_update(int i, const NetImpl& mdl, void (&updwp)(torch::Tensor& w, const torch::Tensor& w1, double decay = 0.999)  = [](torch::Tensor& w, const torch::Tensor& w1, double decay) {
    w.set_data(w.data().detach() * decay + w1.detach() * (1. - decay));
    }, void (&updw)(torch::Tensor& w, const torch::Tensor& w1, double decay = 0.999) = [](torch::Tensor& w, const torch::Tensor& w1, double decay) {
        w = w.detach() * decay;
        w += w1.detach() * (1. - decay);
        }) {
    updw(layers[i].cnvtr1->weight, mdl.layers[i].cnvtr1->weight);
    updw(layers[i].cnvtr2->weight, mdl.layers[i].cnvtr2->weight);
    updw(layers[i].cnvtr3->weight, mdl.layers[i].cnvtr3->weight);
    updw(layers[i].lin1->weight, mdl.layers[i].lin1->weight);
    updw(layers[i].lin2->weight, mdl.layers[i].lin2->weight);
    updw(layers[i].lin3->weight, mdl.layers[i].lin3->weight);

    updw(layers[i].cnv1->weight, mdl.layers[i].cnv1->weight);
    updw(layers[i].cnv2->weight, mdl.layers[i].cnv2->weight);
    updw(layers[i].cnv3->weight, mdl.layers[i].cnv3->weight);
    updw(layers[i].cnv4->weight, mdl.layers[i].cnv4->weight);
    updw(layers[i].rnnresh, mdl.layers[i].rnnresh);
    if (layers[i].mha->in_proj_weight.defined())
        updw(layers[i].mha->in_proj_weight, mdl.layers[i].mha->in_proj_weight);
    if (layers[i].mha->k_proj_weight.defined())
        updw(layers[i].mha->k_proj_weight, mdl.layers[i].mha->k_proj_weight);
    if (layers[i].mha->q_proj_weight.defined())
        updw(layers[i].mha->q_proj_weight, mdl.layers[i].mha->q_proj_weight);
    if (layers[i].mha->v_proj_weight.defined())
        updw(layers[i].mha->v_proj_weight, mdl.layers[i].mha->v_proj_weight);
    for (size_t pi = 0; pi < layers[i].trans->decoder.ptr()->parameters().size(); ++pi)
        updwp(layers[i].trans->decoder.ptr()->parameters()[pi], mdl.layers[i].trans->decoder.ptr()->parameters()[pi].data());//torch::nn::init::xavier_uniform_(p).cuda();
    for (size_t pi = 0; pi < layers[i].trans->encoder.ptr()->parameters().size(); ++pi)
        updwp(layers[i].trans->encoder.ptr()->parameters()[pi], mdl.layers[i].trans->encoder.ptr()->parameters()[pi].data());
    for (size_t pi = 0; pi < layers[i].rnn1->all_weights().size(); ++pi)
        updwp(layers[i].rnn1->all_weights()[pi], mdl.layers[i].rnn1->all_weights()[pi].data());
}

Here I specify all the layers I need to update and the small lambdas are default parameters - however I can't set default parameters on the function pointer prototypes for the decay.

MSVC says:

error C2383: 'updwp': default-arguments are not allowed on this symbol

It doesn't matter if I use reference or pointer to the function.

I'm open to alternative suggestions to make both lambdas and decay default parameters.

Also it was originally:

void ema_update(int i, const NetImpl& mdl, double decay = 0.999, void (&updwp)(torch::Tensor& w, const torch::Tensor& w1)  = [](torch::Tensor& w, const torch::Tensor& w1) {
    w.set_data(w.data().detach() * decay + w1.detach() * (1. - decay));
    }, void (&updw)(torch::Tensor& w, const torch::Tensor& w1) = [](torch::Tensor& w, const torch::Tensor& w1) {
        w = w.detach() * decay;
        w += w1.detach() * (1. - decay);
        })

But that doesn't work either.

Self contained example:

void f(void (*pf)(int &a, double b=0.77) = [] (int &a, double b){
     a *= b;
}) {
    int a = 9;
    pf(a);
}

void f1(double b=0.77, void (*pf)(int &a) = [] (int &a){
    a *= b;
}) {
    int a = 9;
    pf(a);
}

https://wandbox.org/permlink/XmZtzsxmcwgUIbJy

I also tried std functions:

#include <functional>



void f3(double b=0.77, std::function<void (double b, int &a)> fa = [] (double b, int &a){
     a *= b;
}) {
    int a = 9;
    std::function f = std::bind(fa, b);
    f(a);
}

https://wandbox.org/permlink/Jvv3Dw2gSemkyaxt

You can in C++20, but please don't, it's ridiculously unreadable. Use a free function if you must. — Passer By
– Passer By, Commented Mar 8 at 11:38
Do you ever plan on calling the lambda with and without default parameters? Anyway better to make a struct with all the parameters in it and pass an instance of that. Would make the code a lot more readable too — Pepijn Kramer
– Pepijn Kramer, Commented Mar 8 at 11:40
Wait, ignore my previous comment. You mean you want the function pointer parameter to have default arguments? That's never intended, but you can wrap it like auto wrapped = [pf, b](int& a) { return pf(a, b); }; — Passer By
– Passer By, Commented Mar 8 at 11:44
@PasserBy Yes. Yeah that's the solution I've found - I was overcomplicating it :). — AnArrayOfFunctions
– AnArrayOfFunctions, Commented Mar 8 at 11:46
@PepijnKramer No - I only call with the default parameter passed/or omitted by the caller. — AnArrayOfFunctions
– AnArrayOfFunctions, Commented Mar 8 at 11:53

AnArrayOfFunctions · Accepted Answer · 2025-03-08 12:04:25Z

0

I've found a solution - which is the simplest (maybe not the most elegant):

I need to pass by reference because in the original code a is not simple integer but an object with getter and setter I use inside the lambda.

#include <functional>

void f4(double b=0.77, std::function<void (double b, int &a)> fa= [] (double b, int &a){
     a *= b;
}) {
    int a = 9;
    auto f = [fa, b] (int &a) {fa(b, a);};
    f(a);
}

https://wandbox.org/permlink/uKwqe1reWwwEJAGC

edited Mar 8 at 12:04

answered Mar 8 at 11:46

AnArrayOfFunctions

3,9984 gold badges38 silver badges77 bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

Pepijn Kramer Mar 8 at 11:58

Two remarks, why do you give your lambda an "in/out parameter" (its IMO not very idiomatic, that's what you have retrun values for). And instead of using a function pointer, use a std::function<void(double,int&> that way you don't have to explain to your callers they can only pass in lambdas without captures.

AnArrayOfFunctions Mar 8 at 12:01

I need in/out because I operate with 2 objects that I need to modify (in the original code). They have get/set methods. But otherwise thanks I agree std::function is better. (Here is the simplified version)

Eljay Mar 8 at 14:10

@AnArrayOfFunctions • Avoid in/out parameters for value objects. Use move-in or copy-by-value parameters, and return values for value objects. (Reserve in/out parameters for entity objects.) Strive make calculating functions pure, not ones that mutate their caller's arguments in situ.

Collectives™ on Stack Overflow

Why can't I add default-argument to my lambda function parameters (or is there an alternative way to do it)?

1 Answer 1

3 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

3 Comments

Your Answer

Sign up or log in

Post as a guest

Related