Update Python (3.13) and PyTorch (2.9) versions (#1121)

Also:

- drop Python 3.9 (EOL)
- drop PyTorch 2.5
- update some tests
- for torch < 2.9, downgrade to triton 3.4
- fix indentation error in GP classes docstrings (similar to #1080)

Author: BenjaminBossan

.github/workflows/testing.yml

@@ -20,8 +20,8 @@ jobs:
     strategy:
       fail-fast: false  # don't cancel all jobs when one fails
       matrix:
-        python_version: ['3.9', '3.10', '3.11', '3.12']
-        torch_version: ['2.5.1+cpu', '2.6.0+cpu', '2.7.1+cpu', '2.8.0+cpu']
+        python_version: ['3.10', '3.11', '3.12', '3.13']
+        torch_version: ['2.6.0+cpu', '2.7.1+cpu', '2.8.0+cpu', '2.9.0+cpu']
         os: [ubuntu-latest]
 
     steps:
@@ -31,18 +31,17 @@ jobs:
       with:
         python-version: ${{ matrix.python_version }}
     - name: Install dependencies
-      # TODO remove triton 3.1 install if torch < 2.6 no longer supported, see #1093
       # TODO remove triton 3.2 install if torch < 2.7 no longer supported
       run: |
         python -m pip install --upgrade pip
         python -m pip install '.[test,docs,dev,extended]'
         python -m pip install pytest-pretty 
         python -m pip install torch==${{ matrix.torch_version }} -f https://download.pytorch.org/whl/torch
         TORCH_VERSION_MAJOR_MINOR=$(python -c "import torch; v=torch.__version__.split('+')[0]; print('.'.join(v.split('.')[:2]))")
-        if [[ $(echo "$TORCH_VERSION_MAJOR_MINOR < 2.6" | bc -l) -eq 1 ]]; then
-          python -m pip install "triton==3.1"
-        elif [[ $(echo "$TORCH_VERSION_MAJOR_MINOR < 2.7" | bc -l) -eq 1 ]]; then
+        if [[ $(echo "$TORCH_VERSION_MAJOR_MINOR < 2.7" | bc -l) -eq 1 ]]; then
           python -m pip install "triton==3.2"
+        elif [[ $(echo "$TORCH_VERSION_MAJOR_MINOR < 2.9" | bc -l) -eq 1 ]]; then
+          python -m pip install "triton==3.4"
         fi
         python -m pip list
     - name: Install skorch

README.rst

@@ -233,10 +233,10 @@ instructions for PyTorch, visit the `PyTorch website
 <http://pytorch.org/>`__. skorch officially supports the last four
 minor PyTorch versions, which currently are:
 
-- 2.5.1
 - 2.6.0
 - 2.7.1
 - 2.8.0
+- 2.9.0
 
 However, that doesn't mean that older versions don't work, just that
 they aren't tested. Since skorch mostly relies on the stable part of

docs/user/installation.rst

@@ -93,10 +93,10 @@ instructions for PyTorch, visit the `PyTorch website
 <http://pytorch.org/>`__. skorch officially supports the last four
 minor PyTorch versions, which currently are:
 
-- 2.5.1
 - 2.6.0
 - 2.7.1
 - 2.8.0
+- 2.9.0
 
 However, that doesn't mean that older versions don't work, just that
 they aren't tested. Since skorch mostly relies on the stable part of

pyproject.toml

@@ -15,7 +15,7 @@ dynamic = ["version", "readme"]
 description = "scikit-learn compatible neural network library for pytorch"
 license = "BSD-3-Clause"
 license-files = ["LICENSE"]
-requires-python = ">=3.9"
+requires-python = ">=3.10"
 dependencies =[
     "numpy>=1.13.3",
     "scikit-learn>=0.22.0",
@@ -31,7 +31,6 @@ authors = [
     {name = "skorch Developers"}
 ]
 classifiers = [                                 #note:classifiers are recommended for packages to be visible on PyPI
-  "Programming Language :: Python :: 3.9",
   "Programming Language :: Python :: 3.10",
   "Programming Language :: Python :: 3.11",
   "Programming Language :: Python :: 3.12",

skorch/classifier.py

@@ -42,11 +42,13 @@
 """
 
 def get_neural_net_clf_doc(doc):
-    indentation = "    "
     # dedent/indent roundtrip required for consistent indention in both
     # Python <3.13 and Python >=3.13
-    # Because <3.13 => not automatic dedent, but it is the case in >=3.13
-    doc = neural_net_clf_doc_start + " " + textwrap.indent(textwrap.dedent(doc.split("\n", 5)[-1]), indentation)
+    # Because <3.13 => no automatic dedent, but it is the case in >=3.13
+    indentation = "    "
+    doc = textwrap.indent(textwrap.dedent(doc.split("\n", 5)[-1]), indentation)
+
+    doc = neural_net_clf_doc_start + " " + doc
     pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + neural_net_clf_additional_text + doc[end:]
@@ -253,11 +255,13 @@ def predict(self, X):
       is used by ``predict`` and ``predict_proba`` for classification."""
 
 def get_neural_net_binary_clf_doc(doc):
-    indentation = "    "
     # dedent/indent roundtrip required for consistent indention in both
     # Python <3.13 and Python >=3.13
-    # Because <3.13 => not automatic dedent, but it is the case in >=3.13
-    doc = neural_net_binary_clf_doc_start + " " + textwrap.indent(textwrap.dedent(doc.split("\n", 5)[-1]), indentation)
+    # Because <3.13 => no automatic dedent, but it is the case in >=3.13
+    indentation = "    "
+    doc = textwrap.indent(textwrap.dedent(doc.split("\n", 5)[-1]), indentation)
+
+    doc = neural_net_binary_clf_doc_start + " " + doc
     pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + neural_net_binary_clf_criterion_text + doc[end:]

skorch/probabilistic.py

@@ -8,6 +8,7 @@
 
 import pickle
 import re
+import textwrap
 
 import gpytorch
 import numpy as np
@@ -471,7 +472,6 @@ def _predict(self, X):
     Module : gpytorch.models.ExactGP (class or instance)
       The module needs to return a
       :class:`~gpytorch.distributions.MultivariateNormal` distribution.
-
 """
 
 exact_gp_regr_criterion_text = """
@@ -483,7 +483,6 @@ def _predict(self, X):
     criterion : gpytorch.mlls.ExactMarginalLogLikelihood
       The objective function to learn the posterior of of the GP regressor.
       Usually doesn't need to be changed.
-
 """
 
 exact_gp_regr_batch_size_text = """
@@ -492,7 +491,6 @@ def _predict(self, X):
       Mini-batch size. For exact GPs, it must be set to -1, since the exact
       solution cannot deal with batching. To make use of batching, use
       :class:`.GPRegressor` in conjunction with a variational strategy.
-
 """
 
 # this is the same text for exact and approximate GP regression
@@ -505,37 +503,40 @@ def _predict(self, X):
       None. There is no default train split for GP regressors because random
       splitting is typically not desired, e.g. because there is a temporal
       relationship between samples.
-
 """
 
 # this is the same text for all GPs
 gp_likelihood_attribute_text = """
-
     likelihood_: torch module (instance)
       The instantiated likelihood.
-
 """
 
 
 def get_exact_gp_regr_doc(doc):
     """Customizes the net docs to avoid duplication."""
+    # dedent/indent roundtrip required for consistent indention in both
+    # Python <3.13 and Python >=3.13
+    # Because <3.13 => no automatic dedent, but it is the case in >=3.13
+    indentation = "    "
+    doc = textwrap.indent(textwrap.dedent(doc.split("\n", 5)[-1]), indentation)
+
     params_start_idx = doc.find('    Parameters\n    ----------')
     doc = doc[params_start_idx:]
-    doc = exact_gp_regr_doc_start + " " + doc
+    doc = exact_gp_regr_doc_start + doc
 
-    pattern = re.compile(r'(\n\s+)(module .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(module .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + exact_gp_regr_module_text + doc[end:]
 
-    pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + exact_gp_regr_criterion_text + doc[end:]
 
-    pattern = re.compile(r'(\n\s+)(batch_size .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(batch_size .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + exact_gp_regr_batch_size_text + doc[end:]
 
-    pattern = re.compile(r'(\n\s+)(train_split .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(train_split .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + gp_regr_train_split_text + doc[end:]
 
@@ -672,7 +673,6 @@ def fit(self, X, y=None, **fit_params):
     Module : gpytorch.models.ApproximateGP (class or instance)
       The GPyTorch module; in contrast to exact GP, the return distribution does
       not need to be Gaussian.
-
 """
 
 gp_regr_criterion_text = """
@@ -684,25 +684,30 @@ def fit(self, X, y=None, **fit_params):
     criterion : gpytorch.mlls.VariationalELBO
       The objective function to learn the approximate posterior of of the GP
       regressor.
-
 """
 
 
 def get_gp_regr_doc(doc):
     """Customizes the net docs to avoid duplication."""
+    # dedent/indent roundtrip required for consistent indention in both
+    # Python <3.13 and Python >=3.13
+    # Because <3.13 => no automatic dedent, but it is the case in >=3.13
+    indentation = "    "
+    doc = textwrap.indent(textwrap.dedent(doc.split("\n", 5)[-1]), indentation)
+
     params_start_idx = doc.find('    Parameters\n    ----------')
     doc = doc[params_start_idx:]
-    doc = gp_regr_doc_start + " " + doc
+    doc = gp_regr_doc_start + doc
 
-    pattern = re.compile(r'(\n\s+)(module .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(module .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + gp_regr_module_text + doc[end:]
 
-    pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + gp_regr_criterion_text + doc[end:]
 
-    pattern = re.compile(r'(\n\s+)(train_split .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(train_split .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + gp_regr_train_split_text + doc[end:]
 
@@ -744,7 +749,6 @@ def __init__(
     Module : gpytorch.models.ApproximateGP (class or instance)
       The GPyTorch module; in contrast to exact GP, the return distribution does
       not need to be Gaussian.
-
 """
 
 gp_binary_clf_criterion_text = """
@@ -756,21 +760,26 @@ def __init__(
     criterion : gpytorch.mlls.VariationalELBO
       The objective function to learn the approximate posterior of of the GP
       binary classification.
-
 """
 
 
 def get_gp_binary_clf_doc(doc):
     """Customizes the net docs to avoid duplication."""
+    # dedent/indent roundtrip required for consistent indention in both
+    # Python <3.13 and Python >=3.13
+    # Because <3.13 => no automatic dedent, but it is the case in >=3.13
+    indentation = "    "
+    doc = textwrap.indent(textwrap.dedent(doc.split("\n", 5)[-1]), indentation)
+
     params_start_idx = doc.find('    Parameters\n    ----------')
     doc = doc[params_start_idx:]
-    doc = gp_binary_clf_doc_start + " " + doc
+    doc = gp_binary_clf_doc_start + doc
 
-    pattern = re.compile(r'(\n\s+)(module .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(module .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + gp_binary_clf_module_text + doc[end:]
 
-    pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+){1,99}')
+    pattern = re.compile(r'(\n\s+)(criterion .*\n)(\s.+|.){1,99}')
     start, end = pattern.search(doc).span()
     doc = doc[:start] + gp_binary_clf_criterion_text + doc[end:]
 

skorch/tests/test_net.py

@@ -511,16 +511,11 @@ def test_pickle_save_and_load_mixed_devices(
             # 1. one for the failed load
             # 2. for switching devices on the net instance
             # remove possible future warning about weights_only=False
-            # TODO: remove filter when torch<=2.4 is dropped
-            w_list = [
-                warning for warning in w.list
-                if "weights_only=False" not in warning.message.args[0]
-            ]
-            assert len(w_list) == 2
-            assert w_list[0].message.args[0] == (
+            assert len(w.list) == 2
+            assert w.list[0].message.args[0] == (
                 'Requested to load data to CUDA but no CUDA devices '
                 'are available. Loading on device "cpu" instead.')
-            assert w_list[1].message.args[0] == (
+            assert w.list[1].message.args[0] == (
                 'Setting self.device = {} since the requested device ({}) '
                 'is not available.'.format(load_dev, save_dev))
 
@@ -3085,46 +3080,10 @@ def test_torch_load_kwargs_forwarded_to_torch_load(
         del call_kwargs['map_location']  # we're not interested in that
         assert call_kwargs == expected_kwargs
 
-    def test_torch_load_kwargs_auto_weights_false_pytorch_lt_2_6(
+    def test_torch_load_kwargs_auto_weights_true(
             self, net_cls, module_cls, monkeypatch, tmp_path
     ):
-        # Same test as
-        # test_torch_load_kwargs_auto_weights_only_false_when_load_params but
-        # without monkeypatching get_default_torch_load_kwargs. The default is
-        # weights_only=False.
         # See discussion in 1063.
-        from skorch._version import Version
-
-        # TODO remove once torch 2.5.0 is no longer supported
-        if Version(torch.__version__) >= Version('2.6.0'):
-            pytest.skip("Test only for torch < v2.6.0")
-
-        net = net_cls(module_cls).initialize()
-        net.save_params(f_params=tmp_path / 'params.pkl')
-        state_dict = net.module_.state_dict()
-        expected_kwargs = {"weights_only": False}
-
-        mock_torch_load = Mock(return_value=state_dict)
-        monkeypatch.setattr(torch, "load", mock_torch_load)
-        net.load_params(f_params=tmp_path / 'params.pkl')
-
-        call_kwargs = mock_torch_load.call_args_list[0].kwargs
-        del call_kwargs['map_location']  # we're not interested in that
-        assert call_kwargs == expected_kwargs
-
-    def test_torch_load_kwargs_auto_weights_true_pytorch_ge_2_6(
-            self, net_cls, module_cls, monkeypatch, tmp_path
-    ):
-        # Same test as
-        # test_torch_load_kwargs_auto_weights_false_pytorch_lt_2_6 but
-        # with weights_only=True, since it's the new default
-        # See discussion in 1063.
-        from skorch._version import Version
-
-        # TODO remove once torch 2.5.0 is no longer supported
-        if Version(torch.__version__) < Version('2.6.0'):
-            pytest.skip("Test only for torch >= 2.6.0")
-
         net = net_cls(module_cls).initialize()
         net.save_params(f_params=tmp_path / 'params.pkl')
         state_dict = net.module_.state_dict()
@@ -4332,16 +4291,6 @@ def test_compile_missing_dunder_in_prefix_arguments(
             ).initialize()
 
     def test_fit_and_predict_with_compile(self, net_cls, module_cls, data):
-        if not hasattr(torch, 'compile'):
-            pytest.skip(reason="torch.compile not available")
-
-        # python 3.12 requires torch >= 2.4 to support compile
-        # TODO: remove once we remove support for torch < 2.4
-        from skorch._version import Version
-
-        if Version(torch.__version__) < Version('2.4.0') and sys.version_info >= (3, 12):
-            pytest.skip(reason="When using Python 3.12, torch.compile requires torch >= 2.4")
-
         # use real torch.compile, not mocked, can be a bit slow
         X, y = data
         net = net_cls(module_cls, max_epochs=1, compile=True).initialize()
@@ -4362,13 +4311,6 @@ def test_binary_classifier_with_compile(self, data):
         # because of a failing isinstance check
         from skorch import NeuralNetBinaryClassifier
 
-        # python 3.12 requires torch >= 2.4 to support compile
-        # TODO: remove once we remove support for torch < 2.4
-        from skorch._version import Version
-
-        if Version(torch.__version__) < Version('2.4.0') and sys.version_info >= (3, 12):
-            pytest.skip(reason="When using Python 3.12, torch.compile requires torch >= 2.4")
-
         X, y = data[0], data[1].astype(np.float32)
 
         class MyNet(nn.Module):