Update CI to use PyTorch 1.13 (#911)

* Update CI to use PyTorch 1.13

Remove 1.9.

* Revert removal of test matrix exclusions

They still don't seem to work :-/

Also, add fail-fast=false strategy for GH workflow.

* Fix error in testing.yml

* CI: Add pip list to get overview of packages

* Use bool instead of np.bool

Remove DeprecationWarning

* Skip accelerate test until there is a new release

.github/workflows/testing.yml

@@ -14,9 +14,10 @@ jobs:
 
     runs-on: ${{ matrix.os }}
     strategy:
+      fail-fast: false  # don't cancel all jobs when one fails
       matrix:
         python_version: ['3.7', '3.8', '3.9', '3.10']
-        torch_version: ['1.9.1+cpu', '1.10.2+cpu', '1.11.0+cpu', '1.12.0+cpu']
+        torch_version: ['1.10.2+cpu', '1.11.0+cpu', '1.12.1+cpu', '1.13.0+cpu']
         os: [ubuntu-latest]
         exclude:
           - python_version: '3.10'
@@ -36,6 +37,7 @@ jobs:
         python -m pip install -r requirements-dev.txt
         python -m pip install -r requirements.txt
         python -m pip install torch==${{ matrix.torch_version }} -f https://download.pytorch.org/whl/torch_stable.html
+        python -m pip list
     - name: Install skorch
       run: |
         python -m pip install .

README.rst

@@ -237,22 +237,21 @@ instructions for PyTorch, visit the `PyTorch website
 <http://pytorch.org/>`__. skorch officially supports the last four
 minor PyTorch versions, which currently are:
 
-- 1.9.1
 - 1.10.2
 - 1.11.0
-- 1.12.0
+- 1.12.1
+- 1.13.0
 
 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
 the PyTorch API, older PyTorch versions should work fine.
 
-In general, running this to install PyTorch should work (assuming CUDA
-11.1):
+In general, running this to install PyTorch should work:
 
 .. code:: bash
 
     # using conda:
-    conda install pytorch cudatoolkit==11.1 -c pytorch
+    conda install pytorch cudatoolkit -c pytorch
     # using pip
     python -m pip install torch
 

docs/user/installation.rst

@@ -90,10 +90,10 @@ instructions for PyTorch, visit the `PyTorch website
 <http://pytorch.org/>`__. skorch officially supports the last four
 minor PyTorch versions, which currently are:
 
-- 1.9.1
 - 1.10.2
 - 1.11.0
-- 1.12.0
+- 1.12.1
+- 1.13.0
 
 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

skorch/tests/test_helper.py

@@ -334,9 +334,9 @@ def test_len_and_shape(self, slds, y):
         [55],
         [-3],
         [0, 10, 3, -8, 3],
-        np.ones(100, dtype=np.bool),
+        np.ones(100, dtype=bool),
         # boolean mask array of length 100
-        np.array([0, 0, 1, 0] * 25, dtype=np.bool),
+        np.array([0, 0, 1, 0] * 25, dtype=bool),
     ])
     def test_len_and_shape_sliced(self, slds, y, sl):
         # torch tensors don't support negative steps, skip test
@@ -374,9 +374,9 @@ def test_slice(self, slds_cls, custom_ds, X, y, sl, n):
         (slice(-1, None), 0),
         ([55], -1),
         ([0, 10, 3, -8, 3], 1),
-        (np.ones(100, dtype=np.bool), 5),
+        (np.ones(100, dtype=bool), 5),
         # boolean mask array of length 100
-        (np.array([0, 0, 1, 0] * 25, dtype=np.bool), 6),
+        (np.array([0, 0, 1, 0] * 25, dtype=bool), 6),
     ])
     def test_slice_twice(self, slds_cls, custom_ds, X, y, sl0, sl1, n):
         data = X if n == 0 else y

skorch/tests/test_hf.py

@@ -569,6 +569,19 @@ def test_mixed_precision(self, net_cls, accelerator_cls, data, mixed_precision):
     def test_mixed_precision_pickling(
             self, net_cls, accelerator_cls, data, mixed_precision
     ):
+        import accelerate
+        from skorch._version import Version
+
+        # https://github.com/huggingface/accelerate/issues/805
+        version_accelerate = Version(accelerate.__version__)
+        version_torch = Version(torch.__version__)
+        if (
+                (version_accelerate <= Version('0.13.2'))
+                and (version_torch >= Version('1.13.0'))
+        ):
+            reason = "skip because of a bug with accelerate <= 0.13.2 and torch >= 1.13"
+            pytest.skip(msg=reason)
+
         # Pickling currently doesn't work because the forward method on modules
         # is overwritten with a modified version of the method using autocast.
         # Pickle doesn't know how to restore those methods.