Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

initial ooc linear learn fn #89

Merged
merged 4 commits into from
Oct 21, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion Project.toml
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
name = "PotentialLearning"
uuid = "82b0a93c-c2e3-44bc-a418-f0f89b0ae5c2"
authors = ["CESMIX Team"]
version = "0.2.6"
version = "0.2.7"

[deps]
AtomsBase = "a963bdd2-2df7-4f54-a1ee-49d51e6be12a"
Expand Down
96 changes: 92 additions & 4 deletions src/Learning/linear-learn.jl
Original file line number Diff line number Diff line change
Expand Up @@ -226,7 +226,8 @@ Fit energies and forces using weighted least squares.
function learn!(
lp::CovariateLinearProblem,
ws::Vector,
int::Bool
int::Bool;
λ::Real=0.0
)
@views B_train = reduce(hcat, lp.B)'
@views dB_train = reduce(hcat, lp.dB)'
Expand All @@ -249,11 +250,11 @@ function learn!(

βs = Vector{Float64}()
try
βs = (A'*Q*A) \ (A'*Q*b)
βs = (A'*Q*A + λ*I) \ (A'*Q*b)
catch e
println(e)
println("Linear system will be solved using pinv.")
βs = pinv(A'*Q*A)*(A'*Q*b)
βs = pinv(A'*Q*A + λ*I)*(A'*Q*b)
end

# Update lp.
Expand All @@ -263,7 +264,7 @@ function learn!(
else
lp.β .= βs
end

end


Expand All @@ -282,4 +283,91 @@ function learn!(
return learn!(lp, ws, int)
end

function assemble_matrices(lp, ws)
@views B_train = reduce(hcat, lp.B)'
@views dB_train = reduce(hcat, lp.dB)'
@views e_train = lp.e
@views f_train = reduce(vcat, lp.f)

@views A = [B_train; dB_train]
@views b = [e_train; f_train]

W = Diagonal([ws[1] * ones(length(e_train));
ws[2] * ones(length(f_train))])

A, W, b
end

function ooc_learn!(
lb::InteratomicPotentials.LinearBasisPotential,
ds_train::PotentialLearning.DataSet;
ws = [30.0,1.0],
symmetrize::Bool = true,
λ::Union{Real,Nothing} = 0.01,
reg_style::Symbol = :default,
AtWA = nothing,
AtWb = nothing
)

basis_size = length(lb.basis)

if isnothing(AtWA) || isnothing(AtWb)
AtWA = zeros(basis_size,basis_size)
AtWb = zeros(basis_size)

W = zeros(1,1)

configs = get_system.(ds_train)

for config in ds_train
ref_energy = get_values(get_energy(config))
ref_forces = reduce(vcat,get_values(get_forces(config)))

sys = get_system(config)
global_descrs = reshape(sum(compute_local_descriptors(sys,lb.basis)),:,1)'
force_descrs = stack(reduce(vcat,compute_force_descriptors(sys,lb.basis)))'

A = [global_descrs; force_descrs]
b = [ref_energy; ref_forces]
if size(W)[1] != size(A)[1]
W = Diagonal( [ws[1]*ones(length(ref_energy));
ws[2]*ones(length(ref_forces))])
end

AtWA .+= A'*W*A
AtWb .+= A'*W*b
end
else
AtWA = deepcopy(AtWA)
AtWb = deepcopy(AtWb)
end

if symmetrize
AtWA = Symmetric(AtWA)
end

if !isnothing(λ)
if reg_style == :default
reg_matrix = λ*Diagonal(ones(size(AtWA)[1]))
AtWA += reg_matrix
end

if reg_style == :scale_thresh || reg_style == :scale
for i in 1:size(AtWA,1)
reg_elem = AtWA[i,i]*(1+λ)
if reg_style == :scale_thresh
reg_elem = max(reg_elem,λ)
end#
AtWA[i,i] = reg_elem
end
end

end

β = AtWA \ AtWb
println("condition number of AtWA: $(cond(AtWA))")

lb.β .= β

AtWA, AtWb
end
Loading