[mlpack-git] [mlpack] Implementation of primal-dual interior point SDP solver (#391)

[Ryan Curtin]

> @@ -189,5 +187,40 @@ void AugLagrangianFunction<LRSDPFunction>::Gradient(
>    gradient = 2 * s * coordinates;
>  }
>  
> +// Template specializations for function and gradient evaluation.
> +// Note that C++ does not allow partial specialization of class members,
> +// so we have to go about this in a somewhat round-about way.
> +template <>
> +inline double AugLagrangianFunction<LRSDPFunction<SDP<arma::sp_mat>>>::Evaluate(
> +    const arma::mat& coordinates) const
> +{
> +  return EvaluateImpl(function, coordinates, lambda, sigma);
> +}

Hm, right, it actually has to be more complex than this then...

```
template<typename FunctionType>
inline double AugLagrangianFunction<FunctionType>::Evaluate(
    const arma::mat& coordinates,
    const boost::enable_if<IsLRSDPFunctionType<FunctionType>::result>::result* = 0)
{
  // no more need for EvaluateImpl()
}
```

and then

```
template<typename FunctionType>
struct IsLRSDPFunctionType
{
  static const bool result = false;
};

template<>
struct IsLRSDPFunctionType<LRSDP<SDP<mat>>>
{
  static const bool result = true;
};

template<>
struct IsLRSDPFunctionType<LRSDP<SDP<sp_mat>>>
{
  static const bool result = true;
};
```

but at this point the metaprogramming is getting insane enough that maybe your solution is cleaner and easier to read... :)  If there were many more types of `SDP<...>` I'd suggest it more seriously, but that's not the case...

We wouldn't need to change the type signature on `AugLagrangianFunction`; we'd need to declare the function above as an extra member.  Anyway, reconsidering the extra complexity of what I just suggested, I'm now against my own idea.  Well, it was fun to think about, at least...

---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/pull/391/files#r23553960
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20150126/ccdef644/attachment.html>