Merge pull request #304 from ACEsuit/dev-0.10

EquivariantTensors + J1.12 Port

Author: jameskermode

.github/workflows/CI.yml

@@ -26,9 +26,10 @@ jobs:
         version:
           # - '1.10'  (dramatic failures => to be understood)
           - '1.11'
+          - '1.12'
           # - 'nightly'
-        python-version: 
-          - '3.8'
+        python-version:
+          - '3.11'
         os:
           - ubuntu-latest
         arch:
@@ -144,8 +145,8 @@ jobs:
           Pkg.pkg"registry add https://github.com/ACEsuit/ACEregistry"
         shell: bash -c "julia --color=yes {0}"
       - uses: actions/setup-python@v2
-        with: 
-          python-version: ${{ matrix.python-version }}
+        with:
+          python-version: '3.11'
       - uses: julia-actions/julia-buildpkg@v1
       - uses: julia-actions/julia-docdeploy@v1
         env:

CLAUDE.md

@@ -1,6 +1,6 @@
 name = "ACEpotentials"
 uuid = "3b96b61c-0fcc-4693-95ed-1ef9f35fcc53"
-version = "0.9.1" 
+version = "0.10.0"
 
 [deps]
 ACEfit = "ad31a8ef-59f5-4a01-b543-a85c2f73e95c"
@@ -9,12 +9,13 @@ AtomsBase = "a963bdd2-2df7-4f54-a1ee-49d51e6be12a"
 AtomsBuilder = "f5cc8831-eeb7-4288-8d9f-d6c1ddb77004"
 AtomsCalculators = "a3e0e189-c65a-42c1-833c-339540406eb1"
 AtomsCalculatorsUtilities = "9855a07e-8816-4d1b-ac92-859c17475477"
+BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
 Bumper = "8ce10254-0962-460f-a3d8-1f77fea1446e"
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 ChunkSplitters = "ae650224-84b6-46f8-82ea-d812ca08434e"
 DynamicPolynomials = "7c1d4256-1411-5781-91ec-d7bc3513ac07"
 EmpiricalPotentials = "38527215-9240-4c91-a638-d4250620c9e2"
-EquivariantModels = "73ee3e68-46fd-466f-9c56-451dc0291ebc"
+EquivariantTensors = "5e107534-7145-4f8f-b06f-47a52840c895"
 ExtXYZ = "352459e4-ddd7-4360-8937-99dcb397b478"
 Folds = "41a02a25-b8f0-4f67-bc48-60067656b558"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
@@ -54,25 +55,26 @@ AtomsBase = "0.5"
 AtomsBuilder = "0.2.0"
 AtomsCalculators = "0.2"
 AtomsCalculatorsUtilities = "0.1"
-Bumper = "0.6" 
+BenchmarkTools = "1.6.3"
+Bumper = "0.7"
 ChunkSplitters = "3.0"
-EquivariantModels = "0.0.6"
+EquivariantTensors = "0.3"
 ExtXYZ = "0.2.0"
 Interpolations = "0.15"
 PrettyTables = "1.3, 2.0"
 Reexport = "1"
 StaticArrays = "1"
 YAML = "0.4"
-Lux = "0.5"
-LuxCore = "0.1"
+Lux = "1.25"
+LuxCore = "1"
 RepLieGroups = "0.1.1"
-Optimisers = "0.3.4"
-Polynomials4ML = "0.3" 
-Zygote = "0.6"
-julia = "1.10, 1.11"
+Optimisers = "0.3.4, 0.4"
+Polynomials4ML = "0.5" 
+Zygote = "0.6, 0.7"
+julia = "1.10, 1.11, 1.12"
 SparseArrays = "1.10"
 NamedTupleTools = "0.13, 0.14"
-SpheriCart = "0.1"
+SpheriCart = "0.2"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

Project.toml

@@ -12,15 +12,34 @@
 - Version 0.6.x uses `ACE1.jl` as a backend. It is mature and suitable for linear models with few species. This is not longer actively developed, but critical bugfixes can still be provided.  [[docs-v0.6]](https://acesuit.github.io/ACEpotentials.jl/v0.6/)
 - Version 0.7.x is reserved
 - Version 0.8.x and onwards provides a new and much more flexible implementation, and integrates with the [AtomsBase](https://github.com/JuliaMolSim/AtomsBase.jl) ecosystem. Most but not all features from 0.6.x have been ported to this re-implementation. Usability should be the same or improved for most end-users. For developers this provides a much more flexible framework for experimentation. [[docs-v0.8]](https://acesuit.github.io/ACEpotentials.jl/dev/)
-- Version 0.9 onwards is technically compatible with Julia 1.10, but some unit tests show unexplained increases in fit accuracy, hence we highly recommend to use it only with Julia 1.11.
+- Version 0.9.0 onwards is technically compatible with Julia 1.10, but some unit tests show unexplained increases in fit accuracy, hence we highly recommend to use it only with Julia 1.11.
+- Version 0.10 migrates internally to `EquivariantTensors.jl` for improved maintainability, flexibility for model building and different execution backends. User-facing API remains unchanged; existing code should mostly continue to work. Note however that support for the fast evaluator used for some Hessian calculations has been (temporarily) dropped. This series of minor releases is compatible with Julia 1.12
 
 ## Contributing
 
 Contributions are very welcome. Until clear guidelines and practices are established, we recommend to open an issue where the bugfix or enhancement can be discussed, before starting a pull request. We will do our best to respond in a timely manner.
 
-## Quick Start 
+## Quick Start for v0.10.x
 
-- Install Julia 1.11 
+- Versions after v0.10.x are registered in the General registry and are compatible with Julia 1.12
+- Install Julia 1.12
+- Create new folder a.g. `acetutorial`; Open a shell
+- Create a new project in `acetutorial` and install `ACEpotentials.jl`
+```
+julia --project=. 
+] 
+add ACEpotentials
+```
+- Install the Julia tutorials (this installs two Jupyter notebook tutorials)
+```julia-repl
+using ACEpotentials
+ACEpotentials.copy_tutorial()
+```
+- Work through the tutorials.
+
+## Quick Start for v0.9.x
+
+- Install Julia 1.11
 - Create new folder a.g. `acetutorial`; Open a shell
 - Create a new project in `acetutorial` and install `ACEpotentials.jl`
 ```

README.md

@@ -24,7 +24,6 @@ ACE models are defined in terms of body-ordered invariant features of atomic env
 ### Key Dependencies
 
 * [`Polynomials4ML.jl`](https://github.com/ACEsuit/Polynomials4ML.jl) : basic kernels for embeddings and tensors
-* [`EquivariantModels.jl`](https://github.com/ACEsuit/EquivariantModels.jl) : tools for equivariant model building
 * [`RepLieGroups.jl`](https://github.com/ACEsuit/RepLieGroups.jl) : coupling coefficients for equivariant tensors
 * [`ACEfit.jl`](https://github.com/ACEsuit/ACEfit.jl) : unified interface to various regression algorithms
 * [`AtomsBase.jl`](https://github.com/JuliaMolSim/AtomsBase.jl) : community interface for atomic structures / systems

docs/src/index.md

@@ -62,38 +62,44 @@ omp_result = ACEfit.solve(solver_omp, Wt .* At, Wt .* yt, Wv .* Av, Wv .* yv);
 # We can select the final model, a model with 500 active parameters, and a model with a validation error within 1.3 times the minimum validation error.
 # We can use the `ACEfit.asp_select` function to select the desired models from the result.
 
-asp_final = set_parameters!( deepcopy(model), 
+asp_final = set_parameters!( deepcopy(model),
                   ACEfit.asp_select(asp_result, :final)[1]);
-asp_size_50  = set_parameters!( deepcopy(model), 
+asp_size_50  = set_parameters!( deepcopy(model),
                   ACEfit.asp_select(asp_result, (:bysize, 50))[1]);
-asp_error13  = set_parameters!( deepcopy(model), 
+asp_error13  = set_parameters!( deepcopy(model),
                   ACEfit.asp_select(asp_result, (:byerror, 1.3))[1]);
 
-pot_final = fast_evaluator(asp_final; aa_static = false);
-pot_50 = fast_evaluator(asp_size_50; aa_static = true);
-pot_13 = fast_evaluator(asp_error13; aa_static = true);
+# NOTE: fast_evaluator temporarily disabled due to incompatibility with
+# upstream SparseACEbasis (requires major refactoring)
+# pot_final = fast_evaluator(asp_final; aa_static = false);
+# pot_50 = fast_evaluator(asp_size_50; aa_static = true);
+# pot_13 = fast_evaluator(asp_error13; aa_static = true);
 
-err_13 = ACEpotentials.compute_errors(test_data,  pot_13);
-err_50 = ACEpotentials.compute_errors(test_data,  pot_50);
-err_fin = ACEpotentials.compute_errors(test_data, pot_final);
+# Use the models directly for now
+err_13 = ACEpotentials.compute_errors(test_data,  asp_error13);
+err_50 = ACEpotentials.compute_errors(test_data,  asp_size_50);
+err_fin = ACEpotentials.compute_errors(test_data, asp_final);
 
 
 # Similarly, we can compute the errors for the OMP models.
 
-omp_final = set_parameters!( deepcopy(model), 
+omp_final = set_parameters!( deepcopy(model),
                   ACEfit.asp_select(omp_result, :final)[1]);
-omp_50  = set_parameters!( deepcopy(model), 
+omp_50  = set_parameters!( deepcopy(model),
                   ACEfit.asp_select(omp_result, (:bysize, 50))[1]);
-omp_13  = set_parameters!( deepcopy(model), 
+omp_13  = set_parameters!( deepcopy(model),
                   ACEfit.asp_select(omp_result, (:byerror, 1.3))[1]);
 
-pot_fin = fast_evaluator(omp_final; aa_static = false);
-pot_50 = fast_evaluator(omp_50; aa_static = true);
-pot_13 = fast_evaluator(omp_13; aa_static = true);
+# NOTE: fast_evaluator temporarily disabled due to incompatibility with
+# upstream SparseACEbasis (requires major refactoring)
+# pot_fin = fast_evaluator(omp_final; aa_static = false);
+# pot_50 = fast_evaluator(omp_50; aa_static = true);
+# pot_13 = fast_evaluator(omp_13; aa_static = true);
 
-err_13 = ACEpotentials.compute_errors(test_data,  pot_13);
-err_50 = ACEpotentials.compute_errors(test_data,  pot_50);
-err_fin = ACEpotentials.compute_errors(test_data, pot_fin);
+# Use the models directly for now
+err_13 = ACEpotentials.compute_errors(test_data,  omp_13);
+err_50 = ACEpotentials.compute_errors(test_data,  omp_50);
+err_fin = ACEpotentials.compute_errors(test_data, omp_final);
 
 
 # Finally, we can visualize the results along the solution path.

docs/src/tutorials/asp.jl

@@ -26,13 +26,9 @@ include("analysis/potential_analysis.jl")
 include("descriptor.jl")
 
 
-# TODO: to be completely rewritten
+# TODO: to be completely rewritten or retired 
 # include("export.jl")
 
-# Experimental 
-# TODO: this is basically the UFACE interface which we need to revive
-# include("experimental.jl")
-
 
 # ----------------- Exports that seem important to make the tutorials work. 
 

src/ACEpotentials.jl

@@ -1,6 +1,6 @@
 
-import LuxCore: AbstractExplicitLayer, 
-                initialparameters, 
+import LuxCore: AbstractLuxLayer,
+                initialparameters,
                 initialstates
 using StaticArrays: SMatrix, SVector
 using Random: AbstractRNG
@@ -27,7 +27,7 @@ const SPL_OF_SVEC{DIM, T} =
       }
 
 
-mutable struct LearnableRnlrzzBasis{NZ, TPOLY, TT, TENV, T} <: AbstractExplicitLayer
+mutable struct LearnableRnlrzzBasis{NZ, TPOLY, TT, TENV, T} <: AbstractLuxLayer
    _i2z::NTuple{NZ, Int}
    polys::TPOLY
    transforms::SMatrix{NZ, NZ, TT}
@@ -41,7 +41,7 @@ mutable struct LearnableRnlrzzBasis{NZ, TPOLY, TT, TENV, T} <: AbstractExplicitL
 end
 
 
-mutable struct SplineRnlrzzBasis{NZ, TT, TENV, LEN, T} <: AbstractExplicitLayer
+mutable struct SplineRnlrzzBasis{NZ, TT, TENV, LEN, T} <: AbstractLuxLayer
    _i2z::NTuple{NZ, Int}
    transforms::SMatrix{NZ, NZ, TT}
    envelopes::SMatrix{NZ, NZ, TENV}