From eee8dce5898581dce53d36291947f80af7238197 Mon Sep 17 00:00:00 2001 From: Liwei Zhang <68692847+zhanglw0521@users.noreply.github.com> Date: Thu, 1 Oct 2026 12:50:44 +0200 Subject: [PATCH] fitting acceleration --- src/fit.jl | 91 ++++++++++++++++++++++++------------ src/utils/transformations.jl | 25 +++++++++- test/test_transformations.jl | 39 ++++++++++++++++ 3 files changed, 125 insertions(+), 30 deletions(-) diff --git a/src/fit.jl b/src/fit.jl index 444e23f..7f47808 100644 --- a/src/fit.jl +++ b/src/fit.jl @@ -34,7 +34,8 @@ function fit!(model::AbstractModel, Rs::Union{Vector{PState{T}},Vector{Vector{PS for (j,R) in enumerate(Rs) valset = partial_md(R,ps,st)[1] for (i, (l1,l2)) in enumerate(LLset) - A_all[i][(2l1+1)*(2l2+1)*(j-1)+1:(2l1+1)*(2l2+1)*j,:] = flat(valset[i]) + first_row = (2l1 + 1) * (2l2 + 1) * (j - 1) + 1 + flat!(A_all[i], first_row, valset[i]) end end @@ -48,39 +49,60 @@ function fit!(model::AbstractModel, Rs::Union{Vector{PState{T}},Vector{Vector{PS println("Fitting the $(Dict_Int2Orbs[l1])$(Dict_Int2Orbs[l2]) blocks ...") println() - RMSE = 0 - # get the design matrix A for the (l1,l2) block, and delete the corresponding part in A_all A = popat!(A_all, 1) num = size(A)[2] # number of basis A = [A; λ*Γ[i]] - - for kk = 1 : size(model.ps.dot[i].W,1) - ii, jj = k2ij(kk, n_orbs1[l1+1], n_orbs2[l2+1]) - # println("Fitting the ($ii,$jj)-th $(Dict_Int2Orbs[l1])$(Dict_Int2Orbs[l2]) block ...") - - Yij = [ get_Y(Ys[t], n_orbs1, n_orbs2, l1, l2, ii, jj) for t = 1:length(Ys) ] - - # construct Y - Y = zeros(Float64, length(Ys)*length(Yij[1])) - for k in 1:length(Ys) - Y[(k-1)*length(Yij[1])+1:k*length(Yij[1])] = Yij[k] + output_count = size(model.ps.dot[i].W, 1) + block_size = (2l1 + 1) * (2l2 + 1) + data_rows = length(Ys) * block_size + if solver isa ACEfit.QR && iszero(solver.lambda) + Y = zeros(Float64, data_rows + num, output_count) + for kk in 1:output_count + ii, jj = k2ij(kk, n_orbs1[l1+1], n_orbs2[l2+1]) + for sample in eachindex(Ys) + row_start = (sample - 1) * block_size + 1 + rows = row_start:(row_start + block_size - 1) + Y[rows, kk] .= vec(get_Y( + Ys[sample], n_orbs1, n_orbs2, + l1, l2, ii, jj, + )) + end end - Y = [Y; zeros(num)] - # solve for C[kk] - C = ACEfit.solve(solver, A, Y)["C"]; - @set! model.ps.dot.$(layer_set[i]).W[kk,:] = C - # list of potential solvers: ACEfit: QR, LSQR, RRQR, SKLEARN_BRR, SKLEARN_ARD, BLR, TruncatedSVD... - - RMSE += norm((A*C-Y)[1:end-num])^2/(length(Y)-num) + # Solve all orbital pairs with one QR factorization and reuse the + # disposable augmented matrix as the factorization workspace. + A_data = copy(view(A, 1:data_rows, :)) + C = qr!(A) \ Y + residual = A_data * C - view(Y, 1:data_rows, :) + model.ps.dot[i].W .= transpose(C) + RMSE = sqrt(sum(abs2, residual) / (data_rows * output_count)) + else + # Preserve the established one-right-hand-side interface for + # solvers that do not support matrix-valued targets. + mean_squared_error = 0.0 + for kk in 1:output_count + ii, jj = k2ij(kk, n_orbs1[l1+1], n_orbs2[l2+1]) + Y = zeros(Float64, data_rows + num) + for sample in eachindex(Ys) + row_start = (sample - 1) * block_size + 1 + rows = row_start:(row_start + block_size - 1) + Y[rows] .= vec(get_Y( + Ys[sample], n_orbs1, n_orbs2, + l1, l2, ii, jj, + )) + end + C = ACEfit.solve(solver, A, Y)["C"] + @set! model.ps.dot.$(layer_set[i]).W[kk, :] = C + mean_squared_error += + norm((A * C - Y)[1:data_rows])^2 / data_rows + end + RMSE = sqrt(mean_squared_error / output_count) + GC.gc() end - # println("RMSE = $(sqrt(RMSE/length(LLset)))") - println("RMSE = $(sqrt(RMSE/size(model.ps.dot[i].W,1)))") + println("RMSE = $RMSE") println() - - GC.gc() end model = (TP <: On_Model) ? TP(model.model, model.ps, model.st, model.n_orbs, true) : TP(model.model, model.ps, model.st, model.n_orbs1, model.n_orbs2, true) @@ -124,7 +146,8 @@ function fit!(model::AbstractModel, Rs::Union{Vector{PState{T}},Vector{Vector{PS for (j,R) in enumerate(Rs) valset = partial_md(R,ps,st)[1] for (i, (l1,l2)) in enumerate(LLset) - A_all[i][(2l1+1)*(2l2+1)*(j-1)+1:(2l1+1)*(2l2+1)*j,:] = flat(valset[i]) + first_row = (2l1 + 1) * (2l2 + 1) * (j - 1) + 1 + flat!(A_all[i], first_row, valset[i]) end end println("Finish constructing A") @@ -215,7 +238,12 @@ function fit!(model::Density_Model, Rs, Ys; solver = ACEfit.SKLEARN_BRR(), λ = if length(Rs[key]) == 0 || length(Ys[key]) == 0 continue end - model.Models[key] = fit!(model.Models[key], identity.(pop!(Rs,key)), assemble_Y( identity.(pop!(Ys,key)), get_norbs(model.Models[key])...); solver = solver, λ = λ, reg = reg, GC_switcher = GC_switcher) + states = identity.(pop!(Rs, key)) + targets = identity.(pop!(Ys, key)) + model.Models[key] = fit!( + model.Models[key], states, targets; + solver=solver, λ=λ, reg=reg, + ) end else for key in keys(model.Models) @@ -224,7 +252,12 @@ function fit!(model::Density_Model, Rs, Ys; solver = ACEfit.SKLEARN_BRR(), λ = if length(Rs[key]) == 0 || length(Ys[key]) == 0 continue end - model.Models[key] = fit!(model.Models[key], identity.(pop!(Rs,key)), assemble_Y( identity.(pop!(Ys,key)), get_norbs(model.Models[key])...); solver = solver, λ = λ, reg = reg, GC_switcher = GC_switcher) + states = identity.(pop!(Rs, key)) + targets = identity.(pop!(Ys, key)) + model.Models[key] = fit!( + model.Models[key], states, targets; + solver=solver, λ=λ, reg=reg, + ) end end if !isfitted(model) @@ -358,4 +391,4 @@ end # return DM, train_set # end -end # module \ No newline at end of file +end # module diff --git a/src/utils/transformations.jl b/src/utils/transformations.jl index 98cea49..f1f31c7 100644 --- a/src/utils/transformations.jl +++ b/src/utils/transformations.jl @@ -4,7 +4,7 @@ using RepLieGroups.O3: ClebschGordan, wigner_D_indices, adjoint, Rot3DCoeffs, di import RepLieGroups.O3: _mrange import EquivariantModels: coco_dot -export eigen_retraction, svd_retraction, flat, k2ij +export eigen_retraction, svd_retraction, flat, flat!, k2ij # The transformation matrix from complex SHs to real SHs function ctran(L) @@ -39,6 +39,29 @@ function flat(a) return tmp end +function flat!(destination::Matrix, first_row::Int, values) + size(destination, 2) == length(values) || throw(DimensionMismatch( + "the destination and values must have the same number of columns", + )) + isempty(values) && return destination + + value_length = length(first(values)) + first_row >= 1 || throw(BoundsError(destination, first_row)) + first_row + value_length - 1 <= size(destination, 1) || + throw(BoundsError(destination, first_row:(first_row + value_length - 1))) + + destination_rows = size(destination, 1) + @inbounds for column in eachindex(values) + value = values[column] + length(value) == value_length || throw(DimensionMismatch( + "all values must have the same flattened length", + )) + destination_offset = first_row + (column - 1) * destination_rows + copyto!(destination, destination_offset, value, 1, value_length) + end + return destination +end + function k2ij(k, n, m) i = div(k-1, m) + 1 j = k - (i-1)*m diff --git a/test/test_transformations.jl b/test/test_transformations.jl index badcd39..4cb96c8 100644 --- a/test/test_transformations.jl +++ b/test/test_transformations.jl @@ -1,7 +1,9 @@ module TestCouplingTransformations using ACEdensitymatrix +using ACEfit using LinearAlgebra +using Random using SparseArrays using Test @@ -30,6 +32,43 @@ const ADM = ACEdensitymatrix end end end + + @testset "allocation-free feature assembly" begin + values = [ + [1.0 2.0; 3.0 4.0], + [-1.0 0.5; 0.25 2.0], + ] + reference = ADM.flat(values) + destination = fill(NaN, size(reference, 1) + 2, size(reference, 2)) + + ADM.flat!(destination, 2, values) + + @test destination[2:(end - 1), :] == reference + @test all(isnan, destination[[1, end], :]) + @test_throws DimensionMismatch ADM.flat!( + zeros(size(reference, 1), size(reference, 2) + 1), 1, values, + ) + end + + @testset "batched QR matches repeated solves" begin + rng = MersenneTwister(1234) + data_matrix = randn(rng, 18, 6) + regularization = 1e-4 * Matrix{Float64}(I, 6, 6) + design_matrix = vcat(data_matrix, regularization) + data_targets = randn(rng, 18, 5) + targets = vcat(data_targets, zeros(6, 5)) + + batched = qr!(copy(design_matrix)) \ targets + repeated = hcat([ + ACEfit.solve( + ACEfit.QR(), design_matrix, targets[:, column], + )["C"] + for column in axes(targets, 2) + ]...) + + @test batched ≈ repeated rtol=1e-12 atol=1e-12 + @test data_matrix * batched ≈ data_matrix * repeated + end end end