ADA-research · thijssnelleman · Jan 20, 2025 · Jan 14, 2025 · Jan 14, 2025 · Jan 14, 2025
diff --git a/.flake8 b/.flake8
@@ -34,7 +34,6 @@ ignore = D300  # Conflicts with docstring-quotes rule
          S404  # Just a warning about the usage of subprocess
          S603  # Flags only false positives as of now
 exclude =
-    sparkle/Components/AutoFolio/,
     sparkle/Components/ablationAnalysis-0.9.4/,
     sparkle/Components/runsolver/,
     sparkle/Components/smac2-v2.10.03-master-778/

diff --git a/.github/workflows/publish_pypi.yml b/.github/workflows/publish_pypi.yml
@@ -0,0 +1,33 @@
+name: Upload Python Package to PyPI when a Release is Created
+
+on: 
+  workflow_dispatch:
+
+  release:
+    types: [created]
+
+
+jobs:
+  pypi-publish:
+    name: Publish release to PyPI
+    runs-on: ubuntu-latest
+    environment:
+      name: release
+      url: https://pypi.org/p/sparkleai
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.x"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install setuptools wheel build
+      - name: Build package
+        run: |
+          python -m build 
+      - name: Publish package distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
diff --git a/.gitmodules b/.gitmodules
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -10,6 +10,7 @@ Notable changes to Sparkle will be documented in this file.
 
 ### Changed
 - CI pipeline for unittest now tests for Python 3.10, 3.11 and 3.12
+- Replaced AutoFolio with ASF-lib for algorithm selection [SPRK-336]
 
 ## [0.9.2] - 07/01/2025
 

diff --git a/Documentation/source/index.md b/Documentation/source/index.md
@@ -98,7 +98,7 @@ Asside from several package dependencies, Sparkle's package / CLI relies on a fe
 Other dependencies are handled by the Conda environment, but if that is not an option for you please ensure you have the following:
 
 - [libnuma](https://anaconda.org/esrf-bcu/libnuma) and [numactl](https://anaconda.org/brown-data-science/numactl) for [Runsolver](http://www.cril.univ-artois.fr/~roussel/runsolver/) compilation which sparkle uses to measure solvers meta data. This is restricted to Linux based systems.
-- [Swig](https://anaconda.org/conda-forge/swig/) 4.0.2 for [SMAC3](https://github.com/automl/SMAC3), which is in turn used by [AutoFolio](https://github.com/automl/AutoFolio).
+- [Swig](https://anaconda.org/conda-forge/swig/) 4.0.2 for [SMAC3](https://github.com/automl/SMAC3).
 
 For detailed installation instructions see the documentation: https://ada-research.github.io/Sparkle/
 

diff --git a/Documentation/source/platform.md b/Documentation/source/platform.md
@@ -203,14 +203,21 @@ The options below are exclusive to `srun` and are thus discouraged:
 
 ---
 
-`selector`
-> aliases: `selector`
+`selector_class`
+> aliases: `selector_class`
 >
-> values: Path.
+> values: Class.
 >
-> Note: Currently only AutoFolio is supported by Sparkle. This setting is soon to be deprecated for a more flexible solution.
+> Description: The ASF-lib Algorithm selector class to use.
+
+---
+
+`selector_model`
+> aliases: `selector_model`
+>
+> values: Model.
 >
-> Description: The Algorithm selector to use.
+> Description: The sklearn model to use for algorithm selection.
 
 ---
 

diff --git a/Examples/selection.md b/Examples/selection.md
@@ -55,10 +55,10 @@ To make sure feature computation and solver performance computation are done bef
 sparkle jobs
 ```
 
-Now we can construct a portfolio selector, using the previously computed features and the objective value results of running the solvers. We can specify an objective to select on with the `--objective` flag, but if we do not, Sparkle will default to the first objective specified in the Settings file. The `--selector-timeout` argument determines for how many seconds we will train our selector for. We can set the flag `--solver-ablation` for actual marginal contribution computation later.
+Now we can construct a portfolio selector, using the previously computed features and the objective value results of running the solvers. We can specify an objective to select on with the `--objective` flag, but if we do not, Sparkle will default to the first objective specified in the Settings file. We can set the flag `--solver-ablation` for actual marginal contribution computation later.
 
 ```bash
-sparkle construct portfolio selector --selector-timeout 1000 --solver-ablation
+sparkle construct portfolio selector --solver-ablation
 sparkle jobs  # Wait for the constructor to complete its computations
 ```
 
@@ -138,7 +138,7 @@ sparkle jobs
 Now we can train a selector based on these features.
 
 ```bash
-sparkle construct portfolio selector --selector-timeout 1000
+sparkle construct portfolio selector
 sparkle jobs  # Wait for the computation to be done
 ```
 

diff --git a/Examples/selection.sh b/Examples/selection.sh
@@ -44,9 +44,9 @@ sparkle run solvers --performance-data
 
 sparkle jobs
 
-# Now we can construct a portfolio selector, using the previously computed features and the objective value results of running the solvers. We can specify an objective to select on with the `--objective` flag, but if we do not, Sparkle will default to the first objective specified in the Settings file. The `--selector-timeout` argument determines for how many seconds we will train our selector for. We can set the flag `--solver-ablation` for actual marginal contribution computation later.
+# Now we can construct a portfolio selector, using the previously computed features and the objective value results of running the solvers. We can specify an objective to select on with the `--objective` flag, but if we do not, Sparkle will default to the first objective specified in the Settings file. We can set the flag `--solver-ablation` for actual marginal contribution computation later.
 
-sparkle construct portfolio selector --selector-timeout 1000 --solver-ablation
+sparkle construct portfolio selector --solver-ablation
 sparkle jobs  # Wait for the constructor to complete its computations
 
 ### Generate a report
@@ -106,7 +106,7 @@ sparkle jobs
 
 # Now we can train a selector based on these features.
 
-sparkle construct portfolio selector --selector-timeout 1000
+sparkle construct portfolio selector 1000
 sparkle jobs  # Wait for the computation to be done
 
 # And generate the report. When running on the PTN/PTN2 data sets, you can compare the two to see the impact of different feature extractors.

diff --git a/README.md b/README.md
@@ -86,7 +86,7 @@ Asside from several package dependencies, Sparkle's package / CLI relies on a fe
 Other dependencies are handled by the Conda environment, but if that is not an option for you please ensure you have the following:
 
 - [libnuma](https://anaconda.org/esrf-bcu/libnuma) and [numactl](https://anaconda.org/brown-data-science/numactl) for [Runsolver](http://www.cril.univ-artois.fr/~roussel/runsolver/) compilation which sparkle uses to measure solvers meta data. This is restricted to Linux based systems.
-- [Swig](https://anaconda.org/conda-forge/swig/) 4.0.2 for [SMAC3](https://github.com/automl/SMAC3), which is in turn used by [AutoFolio](https://github.com/automl/AutoFolio).
+- [Swig](https://anaconda.org/conda-forge/swig/) 4.0.2 for [SMAC3](https://github.com/automl/SMAC3).
 
 For detailed installation instructions see the documentation: https://ada-research.github.io/Sparkle/
 

diff --git a/setup.py b/setup.py
@@ -23,15 +23,14 @@
           "tabulate==0.9.0",
           "pytermgui==7.7.2",
           "tqdm==4.66.5",
+          "smac==2.2.0",
           "RunRunner==0.1.9.1",
+          "asf-lib==0.0.1.15",
+          # ASF-lib requirements
+          "xgboost==2.1.3",
           # Reporting packages
           "plotly==5.24.1",
           "kaleido==0.2.1",
-          # Autofolio packages
-          "xgboost==2.0.3",
-          "scikit-learn==1.3.2",
-          "liac-arff==2.5.0",
-          "smac==2.2.0",
       ],
       include_package_data=True,
       entry_points={"console_scripts": ["sparkle=sparkle.CLI._cli_:main"], },)
diff --git a/sparkle/CLI/compute_marginal_contribution.py b/sparkle/CLI/compute_marginal_contribution.py
@@ -7,6 +7,7 @@
 
 import tabulate
 
+from sparkle.solver import Selector
 from sparkle.CLI.help import global_variables as gv
 from sparkle.CLI.help import logging as sl
 from sparkle.platform.settings_objects import SettingState
@@ -60,14 +61,16 @@ def compute_selector_performance(
     selector_performance_data.add_solver("portfolio_selector")
     selector_performance_data.csv_filepath =\
         actual_portfolio_selector.parent / "performance.csv"
-    selector = gv.settings().get_general_sparkle_selector()
+    selector = Selector(gv.settings().get_selection_class(),
+                        gv.settings().get_selection_model())
 
     schedule = {}
     for instance in performance_data.instances:
         # We get the performance for an instance by infering the model predicition
         # for the instance.
-        feature_vector = feature_data.get_instance(instance)
-        schedule[instance] = selector.run(actual_portfolio_selector, feature_vector)
+        schedule[instance] = selector.run(actual_portfolio_selector,
+                                          instance,
+                                          feature_data)
 
     schedule_performance = selector_performance_data.schedule_performance(
         schedule, target_solver="portfolio_selector", objective=objective)

diff --git a/sparkle/CLI/construct_portfolio_selector.py b/sparkle/CLI/construct_portfolio_selector.py
@@ -4,9 +4,16 @@
 import argparse
 from pathlib import Path
 
+# from asf import selectors
+
+# from sklearn.base import ClassifierMixin, RegressorMixin
+# from asf.predictors import SklearnWrapper, AbstractPredictor
+# from asf.selectors.abstract_model_based_selector import AbstractModelBasedSelector
+
 import runrunner as rrr
 from runrunner.base import Runner
 
+from sparkle.solver import Selector
 from sparkle.platform.settings_objects import SettingState
 from sparkle.structures import PerformanceDataFrame, FeatureDataFrame
 from sparkle.types import resolve_objective
@@ -24,8 +31,6 @@ def parser_function() -> argparse.ArgumentParser:
                     "solver performances.")
     parser.add_argument(*ac.RecomputePortfolioSelectorArgument.names,
                         **ac.RecomputePortfolioSelectorArgument.kwargs)
-    parser.add_argument(*ac.SelectorTimeoutArgument.names,
-                        **ac.SelectorTimeoutArgument.kwargs)
     parser.add_argument(*ac.ObjectiveArgument.names,
                         **ac.ObjectiveArgument.kwargs)
     parser.add_argument(*ac.SelectorAblationArgument.names,
@@ -66,7 +71,6 @@ def main(argv: list[str]) -> None:
 
     # Process command line arguments
     args = parser.parse_args(argv)
-    selector_timeout = args.selector_timeout
     flag_recompute_portfolio = args.recompute_portfolio_selector
     solver_ablation = args.solver_ablation
 
@@ -86,14 +90,14 @@ def main(argv: list[str]) -> None:
     run_on = gv.settings().get_run_on()
 
     print("Start constructing Sparkle portfolio selector ...")
-    selector = gv.settings().get_general_sparkle_selector()
+    selector = Selector(gv.settings().get_selection_class(),
+                        gv.settings().get_selection_model())
 
     judge_exist_remaining_jobs(
         gv.settings().DEFAULT_feature_data_path,
         gv.settings().DEFAULT_performance_data_path)
 
-    # Selector (AutoFolio) cannot handle cutoff time less than 2, adjust if needed
-    cutoff_time = max(gv.settings().get_general_target_cutoff_time(), 2)
+    cutoff_time = gv.settings().get_general_target_cutoff_time()
 
     performance_data = PerformanceDataFrame(gv.settings().DEFAULT_performance_data_path)
     feature_data = FeatureDataFrame(gv.settings().DEFAULT_feature_data_path)
@@ -107,10 +111,8 @@ def main(argv: list[str]) -> None:
 
     # Selector is named after the solvers it can predict, sort for permutation invariance
     solvers = sorted([s.name for s in gv.settings().DEFAULT_solver_dir.iterdir()])
-    selection_scenario_path = (
-        gv.settings().DEFAULT_selection_output
-        / gv.settings().DEFAULT_general_sparkle_selector.name
-        / "_".join(solvers))
+    selection_scenario_path =\
+        gv.settings().DEFAULT_selection_output / selector.name / "_".join(solvers)
 
     # Update latest scenario
     gv.latest_scenario().set_selection_scenario_path(selection_scenario_path)
@@ -125,13 +127,11 @@ def main(argv: list[str]) -> None:
         sys.exit()
 
     selector_path.parent.mkdir(exist_ok=True, parents=True)
-
     selector_run = selector.construct(selector_path,
                                       performance_data,
                                       feature_data,
                                       objective,
                                       cutoff_time,
-                                      selector_timeout,
                                       run_on=run_on,
                                       sbatch_options=sbatch_options,
                                       base_dir=sl.caller_log_dir)
@@ -158,7 +158,6 @@ def main(argv: list[str]) -> None:
                                              feature_data,
                                              objective,
                                              cutoff_time,
-                                             selector_timeout,
                                              run_on=run_on,
                                              sbatch_options=sbatch_options,
                                              base_dir=sl.caller_log_dir)

diff --git a/sparkle/CLI/core/run_portfolio_selector_core.py b/sparkle/CLI/core/run_portfolio_selector_core.py
@@ -10,7 +10,7 @@
 
 from sparkle.CLI.help import global_variables as gv
 from sparkle.structures import PerformanceDataFrame, FeatureDataFrame
-from sparkle.solver import Solver
+from sparkle.solver import Solver, Selector
 from sparkle.types import SolverStatus
 
 
@@ -95,9 +95,11 @@ def call_solver_solve_instance(
 
     # Run portfolio selector
     print("Sparkle portfolio selector predicting ...")
-    selector = gv.settings().get_general_sparkle_selector()
+    selector = Selector(gv.settings().get_selection_class(),
+                        gv.settings().get_selection_model())
     predict_schedule = selector.run(args.selector,
-                                    feature_data.get_instance(str(args.instance)))
+                                    args.instance,
+                                    feature_data)
 
     if predict_schedule is None:  # Selector Failed to produce prediction
         sys.exit(-1)

diff --git a/sparkle/CLI/help/argparse_custom.py b/sparkle/CLI/help/argparse_custom.py
@@ -490,13 +490,6 @@ def __init__(self: ArgumentContainer, names: list[str], kwargs: dict[str, Any])\
                               "help": "configuration budget per configurator run in "
                                       "seconds (wallclock)"})
 
-SelectorTimeoutArgument = \
-    ArgumentContainer(names=["--selector-timeout"],
-                      kwargs={"type": int,
-                              "default": Settings.DEFAULT_portfolio_construction_timeout,
-                              "help": "Cuttoff time (in seconds) for the algorithm"
-                                      "selector construction"})
-
 SolutionVerifierArgument = \
     ArgumentContainer(names=["--solution-verifier"],
                       kwargs={"type": str,

diff --git a/sparkle/Components/AutoFolio/.gitignore b/sparkle/Components/AutoFolio/.gitignore
diff --git a/sparkle/Components/AutoFolio/CITATION.cff b/sparkle/Components/AutoFolio/CITATION.cff