InexactError() about forwarddiff.jl HOT 15 CLOSED

try nu = zeros(eltype(para), point). For AD to work, the code need to take generic input types. zeros without an explicit types defaults to using Float64.

from forwarddiff.jl.

Thanks for your quick answer, it works!
Because I am a Matlab user and new to Julia, not quite understand the subtle difference. I command zeros(point) and thenu = zeros(eltype(para), point), and found they are all Float64 type. I found myself always confused with Julia type, even in simple case. :(

from forwarddiff.jl.

Behind the scenes, ForwardDiff will provide a vector of a different type (not Float64) to compute the derivatives.

from forwarddiff.jl.

Hi mlubin,

But it didn't work when using forwarddiff_hessian ...
Error is

`hessian` has no method matching hessian(::Array{FADHessian{Float64,6},1})
in g at C:\Users\genwei\.julia\v0.3\ForwardDiff\src\typed_fad\FADHessian.jl:406

from forwarddiff.jl.

I'm not too familiar with this code, I'll let @scidom comment on this.

from forwarddiff.jl.

@mlubin, @genwei007 I am on an island without laptop (on holidays), I will leave this issue open and will comment on it upon my return.

from forwarddiff.jl.

Any updates? I am running into similar problems.

from forwarddiff.jl.

@marcusps, also with forwarddiff_hessian?

from forwarddiff.jl.

Hi @marcusps, still haven't managed to find a bit of time to look into this, hopefully will get it off my todo list soon.

from forwarddiff.jl.

Not with the Hessian, but with a gradient, @mlubin. I'll try to get a minimal test case written tomorrow.

from forwarddiff.jl.

Hi mlubin and scidom,
Just wonder if other packages like ReverseDiffSource or DualNumbers can work for my previous syntax?

Thanks.

from forwarddiff.jl.

DualNumbers can be used for Jacobians but not Hessians. ReverseDiffSource could possibly work.

from forwarddiff.jl.

This is issue has been resolved after doing what @mlubin suggested. Sorry that it took months before I sat down to check it out. The problem was exactly that the routines of ForwardDiff don't work with generic input times, so when you defined zeros you have to explicitly tell what type of elements the resulting nu vector will hold. I will close this issue now. For the sake of completeness, here is the operational code for your example in Julia 0.4:

using ForwardDiff

point = 2 # can be used within function
y = [0:(point - 1);]
function GPCM(para)
  a = para[1]
  d = para[2]
  tau = para[3:4]
  t = para[5]
  nu = zeros(eltype(para), point)
  for k = 1:point
    nu[k] = exp(a.*(y[k].*(t-d) - sum(tau[1:k]) )) [1]
  end
  de =  sum(nu)
  p = nu ./ de
end

para = [1.,0.,0.,0.,1.]
Jac = forwarddiff_jacobian(GPCM, Float64, fadtype=:typed)
Jac(para)

from forwarddiff.jl.

Hi scidom,

How about 'forwarddiff_hessian' which is not applicable in the present case... Still got errors. Thanks.

from forwarddiff.jl.

Hi @genwei007, let me check this and will let you now if there is an obvious fix.

from forwarddiff.jl.