Skip to content

AutoGluon and AMLB have a mismatch in dependencies for certain tasks #709

@israel-cj

Description

@israel-cj

We are creating a new openml suite (some tasks contain textual features), and we have some AutoGluon (v.1.2 and v1.3) errors when handling them. Those tasks have been successfully handled by lightautoml and flaml, the tasks are:

Error #2 (occurs 10 times): Error message: NoResultError: /usr/lib/python3.9/distutils/core.py Affected task IDs (10): openml.org/t/363580, openml.org/t/363588, openml.org/t/363576, openml.org/t/363536, openml.org/t/363579, openml.org/t/363590, openml.org/t/363528, openml.org/t/363728, openml.org/t/363530, openml.org/t/363719

AutoGluon seems to have finished fitting and predicting, and then something on AMLB's side failed. This is the error:

AutoGluon training complete, total runtime = 985.45s ... Best model: WeightedEnsemble_L2 | Estimated inference throughput: 142925.9 rows/s (2500 batch size) INFO:sklearnex: sklearn.neighbors.KNeighborsClassifier.predict_proba: running accelerated version on CPU [DEBUG] [frameworks.shared.caller:13:22:16.905] Result from subprocess: {'error_message': '/usr/lib/python3.9/distutils/core.py', 'models_count': 0} [ERROR] [amlb.benchmark:13:22:16.905] /usr/lib/python3.9/distutils/core.py Traceback (most recent call last): File "/bench/amlb/benchmark.py", line 798, in run meta_result = self.benchmark.framework_module.run( File "/bench/frameworks/AutoGluon/__init__.py", line 15, in run return run_autogluon_tabular(dataset, config) File "/bench/frameworks/AutoGluon/__init__.py", line 32, in run_autogluon_tabular return run_in_venv( File "/bench/frameworks/shared/caller.py", line 169, in run_in_venv raise NoResultError(res.error_message) amlb.results.NoResultError: /usr/lib/python3.9/distutils/core.py

What I have tried and did not work, based on the help from Nick:

  • distutils was deprecated in Python 3.10 (PEP 632) and removed in Python 3.12. So I tried running with Python 3.10 and still persistent
  • It seems like torch tries to import triton, which then leads to some weird imports which eventually gets to distutils. Nick mentioned triton isn’t used by AutoGluon to his knowledge, so maybe try uninstalling. It did not work.

Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions