crossprod.R

require(autodiffr)
ad_setup() # to ensure it is established

ores <- function(x){
    x # Function will be the parameters. ofn is sum of squares
}

ores <- ad.variant(ores)

ofn <- function(x){
    res <- ores(x) # returns a vector of residual values
    # print(res)
    val <- as.vector(crossprod(res))
    # val <- crossprod(res)
    # as.numeric can't be used as it is equivalent to as.double
    # but when using autodiff w.r.t the function,
    # the argument will be of some extension of the float number,
    # so as.numeric (as.double) will cause problems.
    # I'm wondering whether ad.variant should treat as.numeric (as.double) as as.vector.
    # BTW, things like as.numeric shouldn't be needed because I try to make autodiffr adaptable
    # to various kinds of input/output,
    # at least for user interface functions like grad/jacobian/hessian,
    # so if a 1x1 matrix doesn't work, it should be counted as an issue.
    val
}

ofn <- ad.variant(ofn)

## Now try to generate the gradient function
ogr <- grad(ofn, mode = "forward") ## there is some problems with the default reverse mode
print(ogr) # this will be more or less meaningless link to Julia function
x0 <- c(1,2,3)
print(ofn(x0)) # should be 14
print(ogr(x0)) # should be 2, 4, 6