diff --git a/RELEASES.md b/RELEASES.md
index 33be0db8e..9a5ddef50 100644
--- a/RELEASES.md
+++ b/RELEASES.md
@@ -11,6 +11,7 @@
 - Load triton before TensorFlow in `ot.backend` so that building a torch optimizer no longer segfaults the interpreter, and remove the `torch<2.12` pin from the doctest and documentation requirements (PR #839, Issue #816)
 - Fix mean centering in `ot.dr.fda` and `ot.dr.wda`: `np.mean(X)` returned a scalar instead of the per-feature mean, so `proj` did not center the data as documented. In `ot.dr.fda` the same pattern in the class means made the between-class scatter matrix independent of which features separate the classes, and FDA returned a non-discriminant direction (PR #840)
 - `ot.dr.fda` and `ot.dr.wda` no longer modify the input array `X` in place (PR #840)
+- Fix `UnbalancedSinkhornTransport` `transform` failing with `AttributeError: 'NoneType' object has no attribute 'array_equal'` when `fit` was called with missing parameters (PR #837, Issue #650)
 
 ## 0.9.7.post1
 
diff --git a/ot/da.py b/ot/da.py
index 7b4ed7ba2..7e1a7491f 100644
--- a/ot/da.py
+++ b/ot/da.py
@@ -2287,6 +2287,8 @@ class label
             Returns self.
         """
 
+        self._get_backend(Xs, ys, Xt, yt)
+
         # check the necessary inputs parameters are here
         if check_params(Xs=Xs, Xt=Xt):
             super(UnbalancedSinkhornTransport, self).fit(Xs, ys, Xt, yt)
diff --git a/test/test_da.py b/test/test_da.py
index df224a247..0a89ebb79 100644
--- a/test/test_da.py
+++ b/test/test_da.py
@@ -483,6 +483,34 @@ def test_unbalanced_sinkhorn_transport_class(nx):
         assert len(otda.log_.keys()) != 0
 
 
+@pytest.skip_backend("jax")
+@pytest.skip_backend("tf")
+def test_unbalanced_sinkhorn_transport_nx_initialized(nx):
+    """non-regression test for issue #650
+
+    fit must always initialize the backend attribute nx, even when called
+    with missing parameters, so that transform fails on a clear error
+    instead of 'NoneType' object has no attribute 'array_equal'
+    """
+
+    ns = 50
+    Xs, ys = make_data_classif("3gauss", ns)
+    Xt, yt = make_data_classif("3gauss2", ns)
+
+    Xs, ys, Xt, yt = nx.from_numpy(Xs, ys, Xt, yt)
+
+    # incomplete fit (Xt missing) still initializes the backend
+    otda = ot.da.UnbalancedSinkhornTransport()
+    otda.fit(Xs=Xs)
+    assert otda.nx is not None
+
+    # complete fit followed by a separate transform call
+    otda = ot.da.UnbalancedSinkhornTransport()
+    otda.fit(Xs=Xs, Xt=Xt)
+    transp_Xs = otda.transform(Xs)
+    assert_equal(transp_Xs.shape, Xs.shape)
+
+
 @pytest.skip_backend("jax")
 @pytest.skip_backend("tf")
 def test_emd_transport_class(nx):