Error when trying to use BLR fitting

[od-qmul]

I get this error:
length(model.basis) = 2396
┌─────────┬──────────┬───────┬──────┬────────┬───────┐
│ Type │ #Configs │ #Envs │ #E │ #F │ #V │
├─────────┼──────────┼───────┼──────┼────────┼───────┤
│ default │ 3598 │ 75929 │ 3598 │ 0 │ 0 │
├─────────┼──────────┼───────┼──────┼────────┼───────┤
│ total │ 3598 │ 75929 │ 3598 │ 0 │ 0 │
│ missing │ 0 │ 0 │ 0 │ 227787 │ 21588 │
└─────────┴──────────┴───────┴──────┴────────┴───────┘
[ Info: Assembling linear problem.
[ Info: - Creating feature matrix with size (3598, 2396).
[ Info: - Beginning assembly with processor count: 8.
Progress: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| Time: 0:04:00
[ Info: - Assembly completed.
[ Info: Assembling full weight vector.
Progress: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| Time: 0:00:01
ERROR: LoadError: MethodError: no method matching getindex(::Nothing, ::String)
Stacktrace:
[1] acefit!(model::ACE1x.ACE1Model, raw_data::Vector{Atoms}; solver::ACEfit.BLR, weights::Dict{String, Dict{String, Float64}}, energy_key::String, force_key::String, virial_key::String, smoothness::Int64, prior::Nothing, repulsion_restraint::Bool, restraint_weight::Float64, export_lammps::Nothing, export_json::Nothing, verbose::Bool)
@ ACEpotentials ~/.julia/packages/ACEpotentials/ZGCA0/src/model.jl:79
[2] top-level scope
@ ~/ubc_work/TiO2_ace/fit_TiO2_ace.jl:23
in expression starting at /home/ubc_oadicks/ubc_work/TiO2_ace/fit_TiO2_ace.jl:23

When I run this code:

1 using Distributed
2 addprocs(7, exeflags="--project=$(Base.active_project())")
3 @Everywhere using ACEpotentials
4
5 using ACEpotentials
6 data_file = "./datasets/train_set_tio2.xyz"
7 datatrain = read_extxyz(data_file)
8 model = acemodel(elements = [:Ti, :O],
9 order = 3,
10 totaldegree = 12,
11 rcut = 8.0,
12 Eref = [:Ti =>-1622.7019449, :O =>-434.20893827]) #Energies in eV
13 @show length(model.basis);
14
15 weights = Dict("E" => 10.0, "F" => 1.0 , "V" => 0.0 )
16
17 #solver = ACEfit.LSQR(damp = 1e-2, atol = 1e-6);
18 solver = ACEfit.BLR(factorization = "svd", tol = 1e-6);
19
20 #P = smoothness_prior(model; p = 4) # (p = 4 is in fact the default)
21
22 #acefit!(model, datatrain; solver=solver, prior = P);
23 acefit!(model, datatrain; solver=solver);
24
25 @info("Training Error Table")
26 ACEpotentials.linear_errors(datatrain, model; weights=weights);
27
28 @info("Test Error Table")
29 test_data_file = "./datasets/test_set_tio2.xyz"
30 datatest = read_extxyz(test_data_file)
31 ACEpotentials.linear_errors(datatest, model; weights=weights);
32
33 export2lammps("./TiO2_potential_v5.yace", model)
34 export2json("./TiO2_potential_v5.json", model)

When I use LSQR there are no errors.
Cholesky factorization failed

[cortner]

On a plane and can't use browser for some reason, so a limited answer ... It seems that the LSQ system is extremely ill-conditioned hence cholesky fails. I don't really understand though why it is using Cholesky when you say you want SVD and with that this issue should not occur. Is it possible that the way you select the solver is wrong syntax?  What happens if you use just `solver = ACEfit.BLR()` Either @wcwitt or I need to look at that. Message ID: ***@***.***>

[od-qmul]

Thanks. Sorry I should have been clearer. It converges when I use

solver = ACEfit.LSQR(damp = 1e-2, atol = 1e-6);
P = smoothness_prior(model; p = 4) # (p = 4 is in fact the default)
acefit!(model, datatrain; solver=solver, prior = P);

I switched to BLR to see if I could reduce the error.

When I initially tried
solver = ACEfit.BLR()
and used the default parameters i get this error:
ERROR: LoadError: PosDefException: matrix is not positive definite; Cholesky factorization failed.

So I switched to factorization = "svd" as stated in the source code

[wcwitt]

Is the @Everywhere in your script actually capitalized?

[od-qmul]

It is, but it doesn't seem to be an issue, the code runs properly with another solver

[wcwitt]

Ah - wait, what ACEfit version do you have? It might be the bug that was fixed by v0.1.4.

[od-qmul]

This is the version of ACEpotentials I have

Status ~/.julia/environments/v1.9/Project.toml
[3b96b61c] ACEpotentials v0.6.3

[wcwitt]

Thanks - can you please look in your Manifest for the ACEfit version?

[od-qmul]

The Manifest says it is v0.1.4

25 [[deps.ACEfit]]
26 deps = ["Distributed", "IterativeSolvers", "LinearAlgebra", "LowRankApprox", "Optim", "ParallelDataTransfer", "ProgressMeter", "SharedArrays", "StaticArrays"]
27 git-tree-sha1 = "f90a60d5d05f1343bdde77a7603e0c1ae23c681b"
28 uuid = "ad31a8ef-59f5-4a01-b543-a85c2f73e95c"
29 version = "0.1.4"

[wcwitt]

Found it: factorization = "svd" needs to be factorization = :svd.

[od-qmul]

Thanks that worked well