WIP: Multiple Properties codes

[andresrossb]

Updating current linear codes to allow for multiple properties.

[cortner]

mutiple properties - ACE.jl doesn't know about species and doesn't want to know about species :)

[cortner]

   # stage 3: get the gradients
   fill!(g, zero(eltype(g)))
   for iX = 1:length(cfg)
      @inbounds @fastmath for iA = 1:length(basis1p)
         g[iX] += _real(dAco[iA] * dA[iA, iX])
      end
   end

might become something like

   # stage 3: get the gradients
   fill!(g, zero(eltype(g)))
   for iP = 1:num_properties, iX = 1:length(cfg)
      for iA = 1:length(basis1p)
         g[iP, iX] += _real(dAco[iA][iP] * dA[iA, iX])
      end
   end

[cortner]

For testing: check that evaluate, grad_config, grad_params, grad_params_config give same answers with

• one model, N properties
• N models, 1 property. Doing 2, 3, properties should be enough.

[cortner]

this ready to be merged?

[cortner]

we may need to rebase onto new main

[cortner]

@andresrossb Please pull this branch, I've rebased onto latest mater

[cortner]

I think we might be ready to review it tomorrow?

[cortner]

I think this is very close to merging, thanks for your work on this. Any questions for clarification, just add below the comments, or here.

[cortner]

Hi Andres - when would be a good moment to rebase your PR onto the latest MAIN branch? I'd like to do it asap so they don't diverge too much, but will wait until you have committed everything and don't work on it for a few minutes. Ping me in zulip?

[cortner]

please take another look at my questions and let's try and talk later

[cortner]

from my end the only thing left to do is add some documentation.

[cortner]

strange - a lot of tests are failing for me right now, I have no idea why

[cortner]

but I think this was caused in one of my PRs - I have no idea how I missed this.

[cortner]

most but not all tests are fixed. There is a nasty failure of grad_config_params cause by this:

function grad_params_config(m::LinearACEModel, cfg::AbstractConfiguration) 
   dB = acquire_dB!(m.basis, cfg)
   return [grad_params_config!(dB, m, cfg) for i = 1:length(m.c[1])]
end

Why was it necessary to wrap this in a vector? That makes its behaviour inconsistent with the scalar case and also inconsistent with what we discussed I thought?

I think this is the last thing we need to resolve before merging.

[cortner]

I fixed the FIO tests and merged into your branch. So now it should be possbile to get all tests to pass.

[cortner]

finally - thanks for your work. I won't delete the branch for now since we want to now look into the differentiation issue.