improve coverage and fix doc

Author: tvayer

ot/backend.py

@@ -2241,6 +2241,8 @@ def randn(self, *size, type_as=None):
             return torch.randn(size=size, generator=self.rng_)
 
     def randperm(self, size, type_as=None):
+        if not isinstance(size, int):
+            raise ValueError("size must be an integer")
         if type_as is not None:
             generator = (
                 self.rng_cuda_ if self.device_type(type_as) == "GPU" else self.rng_

ot/bregman/_empirical.py

@@ -936,7 +936,7 @@ def empirical_sinkhorn_nystroem2(
     >>> X_s = np.reshape(np.arange(n_samples_a, dtype=np.float64), (n_samples_a, 1))
     >>> X_t = np.reshape(np.arange(0, n_samples_b, dtype=np.float64), (n_samples_b, 1))
     >>> empirical_sinkhorn_nystroem2(X_s, X_t, reg, anchors, random_state=42)  # doctest: +ELLIPSIS
-    1.9138489870270898
+    2.5
 
 
     References

test/test_backend.py

@@ -200,6 +200,8 @@ def test_empty_backend():
         nx.seed(42)
     with pytest.raises(NotImplementedError):
         nx.rand()
+    with pytest.raises(NotImplementedError):
+        nx.randperm(12)
     with pytest.raises(NotImplementedError):
         nx.randn()
         nx.coo_matrix(M, M, M)
@@ -215,6 +217,8 @@ def test_empty_backend():
         nx.where(M, M, M)
     with pytest.raises(NotImplementedError):
         nx.copy(M)
+    with pytest.raises(NotImplementedError):
+        nx.pinv(M)
     with pytest.raises(NotImplementedError):
         nx.allclose(M, M)
     with pytest.raises(NotImplementedError):
@@ -759,6 +763,9 @@ def test_random_backends(nx):
     M4 = nx.sort(nx.randperm(5))
     assert np.allclose(M3, M4)
 
+    with pytest.raises(ValueError, match="size must be"):
+        res = nx.randperm(size=[5, 12])
+
 
 def test_gradients_backends():
     rnd = np.random.RandomState(0)

test/test_lowrank.py

@@ -144,6 +144,18 @@ def test_nystroem_sinkhorn2(log):
 
     np.testing.assert_allclose(loss1, loss2, atol=1e-07, rtol=1e-3)
 
+    with pytest.raises(ValueError, match="anchors must"):
+        res = ot.bregman.empirical_sinkhorn_nystroem2(
+            Xs,
+            Xt,
+            anchors=1,
+            reg=reg,
+            numItermax=3000,
+            verbose=True,
+            random_state=random_state,
+            log=log,
+        )
+
 
 def test_compute_lr_sqeuclidean_matrix():
     # test computation of low rank cost matrices M1 and M2