Fix FutureWarnings from numpy and torch (#921)

- don't use torch.testing.assert_allclose, use assert_close instead
- don't use np.bool, use bool instead

Note that np.dtype('bool') == bool, so replace np.bool by bool should
just work.

- Adjust test that used ragged numpy array

Numpy used to allow creation of ragged arrays, e.g.

np.array([[1], [2, 2], [3, 3, 3]])

This does not work anymore with newer numpy versions (tested on
1.24.0rc2). To make it work, dtype=object has to be passed explicitly.
This does not any skorch code directly, but it was necessary to fix a
test.

skorch/helper.py

@@ -242,7 +242,7 @@ def __getitem__(self, i):
                 raise IndexError("SliceDataset only supports slicing with 1 "
                                  "dimensional arrays, got {} dimensions instead."
                                  "".format(i.ndim))
-            if i.dtype == np.bool:
+            if i.dtype == bool:
                 i = np.flatnonzero(i)
 
         return cls(self.dataset, idx=self.idx, indices=self.indices_[i])

skorch/tests/test_helper.py

@@ -389,7 +389,7 @@ def test_slice_twice(self, slds_cls, custom_ds, X, y, sl0, sl1, n):
     @pytest.mark.parametrize('sl0, sl1, sl2', [
         (slice(0, 50), slice(10, 20), 5),
         ([0, 10, 3, -8, 3], [1, 2, 3], 2),
-        (np.ones(100, dtype=np.bool), np.arange(10, 40), 29),
+        (np.ones(100, dtype=bool), np.arange(10, 40), 29),
     ])
     def test_slice_three_times(self, slds_cls, custom_ds, X, y, sl0, sl1, sl2, n):
         data = y if n else X

skorch/tests/test_hf.py

@@ -1050,7 +1050,7 @@ def test_saved_net_is_same(
         for key, original in net_loaded.module_.state_dict().items():
             original = net.module_.state_dict()[key]
             loaded = net_loaded.module_.state_dict()[key]
-            torch.testing.assert_allclose(loaded, original)
+            torch.testing.assert_close(loaded, original)
 
     @pytest.mark.parametrize('storage', ['memory', 'str', 'path'])
     def test_saved_params_is_same(
@@ -1087,7 +1087,7 @@ def test_saved_params_is_same(
         assert len(state_dict_before) == len(state_dict_after)
         for key, original in state_dict_before.items():
             loaded = state_dict_after[key]
-            torch.testing.assert_allclose(loaded, original)
+            torch.testing.assert_close(loaded, original)
 
     def test_latest_url_attribute(self, net, data, hf_hub_storer_cls):
         # Check that the URL returned by the HF API is stored as latest_url. In

skorch/tests/test_net.py

@@ -2293,11 +2293,13 @@ def loss_fn(y_pred, y_true, **kwargs):
         net.fit(X, y)
 
     def test_net_variable_label_lengths(self, net_cls, sequence_module_cls):
-        # neural net should work fine with varying y_true sequences.
+        # neural net should work fine with variable length y_true sequences.
         X = np.array([1, 5, 3, 6, 2])
-        y = np.array([[1], [1, 0, 1], [1, 1], [1, 1, 0], [1, 0]])
+        y = np.array([[1], [1, 0, 1], [1, 1], [1, 1, 0], [1, 0]], dtype=object)
         X = X[:, np.newaxis].astype('float32')
-        y = np.array([np.array(n, dtype='float32')[:, np.newaxis] for n in y])
+        y = np.array(
+            [np.array(n, dtype='float32')[:, np.newaxis] for n in y], dtype=object
+        )
 
         net = net_cls(
             sequence_module_cls,

skorch/tests/test_probabilistic.py

@@ -10,7 +10,6 @@
 from sklearn.model_selection import GridSearchCV
 from sklearn.pipeline import Pipeline
 import torch
-from torch.testing import assert_allclose
 
 from skorch._version import Version
 from skorch.utils import is_torch_data_type
@@ -272,7 +271,7 @@ def test_save_load_params(self, gp_fit, tmpdir):
                 gp_fit.get_all_learnable_params(), gp2.get_all_learnable_params(),
         ):
             assert n0 == n1
-            assert_allclose(p0, p1)
+            torch.testing.assert_close(p0, p1)
 
     ##############
     # functional #