diff --git a/_unittests/ut_testing/test_einsum_bug.py b/_unittests/ut_testing/test_einsum_bug.py
index 8ffa0ff0e..1f3b7d085 100644
--- a/_unittests/ut_testing/test_einsum_bug.py
+++ b/_unittests/ut_testing/test_einsum_bug.py
@@ -2,8 +2,11 @@
 @brief      test log(time=3s)
 """
 import unittest
+import numpy
 from pyquickhelper.pycode import ExtTestCase
-from mlprodict.testing.einsum import decompose_einsum_equation
+from mlprodict.testing.einsum import (
+    decompose_einsum_equation, optimize_decompose_einsum_equation)
+from mlprodict.onnxrt import OnnxInference
 
 
 class TestEinsumBug(ExtTestCase):
@@ -20,6 +23,39 @@ def test__pprint_forward(self):
         spl = pf.split("<- id")
         self.assertEqual(len(spl), 4)
 
+    def common_test_equation(self, equation, dim1, dim2):
+        seq = decompose_einsum_equation(
+            equation, clean=True, strategy='numpy')
+        onx = seq.to_onnx('Y', 'X1', 'X2')
+        sequ = equation.replace(",", "_").replace("->", "__")
+        with open("temp_%s_A.onnx" % sequ, "wb") as f:
+            f.write(onx.SerializeToString())
+        a = numpy.random.rand(*list((2, ) * dim1))
+        b = numpy.random.rand(*list((2, ) * dim2))
+        oinf = OnnxInference(onx)
+        got = oinf.run({'X1': a, 'X2': b})
+        expected = numpy.einsum(equation, a, b)
+        self.assertEqualArray(expected, got['Y'])
+
+        res = optimize_decompose_einsum_equation(
+            equation, numpy.float64, optimize=True, runtime="python",
+            cache=False, opset=15, decompose=True, strategy='ml',
+            verbose=None)
+        new_eq = res.equation_
+        new_onx = res.onnx_
+        sequ = new_eq.replace(",", "_").replace("->", "__")
+        with open("temp_%s_B.onnx" % sequ, "wb") as f:
+            f.write(new_onx.SerializeToString())
+        oinf = OnnxInference(new_onx)
+        got = oinf.run({'X0': a, 'X1': b})
+        self.assertEqualArray(expected, got['Y'])
+
+    def test_decompose_einsum_abc_cde_abde(self):
+        self.common_test_equation("abc,cde->abde", 3, 3)
+
+    def test_decompose_einsum_abcd_cde_abe(self):
+        self.common_test_equation("abcd,cde->abe", 4, 3)
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/mlprodict/onnxrt/ops_cpu/op_tree_ensemble_regressor.py b/mlprodict/onnxrt/ops_cpu/op_tree_ensemble_regressor.py
index 4b5735e97..9cd057c25 100644
--- a/mlprodict/onnxrt/ops_cpu/op_tree_ensemble_regressor.py
+++ b/mlprodict/onnxrt/ops_cpu/op_tree_ensemble_regressor.py
@@ -157,7 +157,7 @@ class TreeEnsembleRegressorDouble(TreeEnsembleRegressorCommon):
                         onx64 = to_onnx(model, X, rewrite_ops=True, target_opset=15)
                         assert 'TreeEnsembleRegressorDouble' in str(onx64)
                         expected = model.predict(X)
-                        
+
                         oinf = OnnxInference(onx64)
                         got = oinf.run({'X': X})
                         diff = numpy.abs(got['variable'] - expected)
diff --git a/mlprodict/testing/einsum/einsum_fct.py b/mlprodict/testing/einsum/einsum_fct.py
index c46579e81..d9c8b6e34 100644
--- a/mlprodict/testing/einsum/einsum_fct.py
+++ b/mlprodict/testing/einsum/einsum_fct.py
@@ -333,6 +333,8 @@ def _einsum(equation, dtype, optimize=False, runtime="batch_dot",
     if cache:
         key = equation, runtime, opset, optimize, dtype, decompose, strategy
         cached = _einsum_cache.get(key, None)
+    else:
+        key = None
     if cached is None:
         cached = CachedEinsum.build_einsum(
             equation, runtime, opset, optimize,