Availability of extra data parameters for optimization

[JohanneMontano]

Good day,

I have a complicated loss function for my problem so the derivatives are a bit convoluted and annoying to work with hence I am considering to use autodiff. However, this loss function requires real world samples at any time t, but these parameters are not the target of optimization. I was aiming to use backward mode so I can get all the partial derivatives in one go. I have looked at the documentation and was wondering if something like this was possible,

// Samples derived in real-time go inside somehow
struct Params {// Insert dynamic data structure here, like a vector or something};

// The function that depends on parameters for which derivatives are needed
var f(var a,  var b, var c, var d, var e, const Params& params)
{
    // Insert complicated loss here, e.g.,
   for (const auto& param: params) {
        Params new_params = params - offset_tensor;
        alpha = new_params / a - b*...
 }
   return loss_val;
}

Also just as a side question, are there any suggestions for new users like me to make sure that the autodiff routine runs as fast as possible?

Thank you very much in advance