Moved test out of run_in_graph_and_eager_mode in sparsemax. (#1408)

Author: gabrieldemarmiesse

tensorflow_addons/activations/sparsemax_test.py

@@ -230,28 +230,6 @@ def test_permutation(self, dtype=None):
             )
             self.assertShapeEqual(p_expected, tf_sparsemax_op)
 
-    def test_diffrence(self, dtype=None):
-        """check sparsemax proposition 4."""
-        random = np.random.RandomState(7)
-
-        z = random.uniform(low=-3, high=3, size=(test_obs, 10))
-        _, p = self._tf_sparsemax(z, dtype)
-
-        etol = {"float16": 1e-2, "float32": 1e-6, "float64": 1e-9}[dtype]
-
-        for val in range(0, test_obs):
-            for i in range(0, 10):
-                for j in range(0, 10):
-                    # check condition, the obesite pair will be checked anyway
-                    if z[val, i] > z[val, j]:
-                        continue
-
-                    self.assertTrue(
-                        0 <= p[val, j] - p[val, i] <= z[val, j] - z[val, i] + etol,
-                        "0 <= %.10f <= %.10f"
-                        % (p[val, j] - p[val, i], z[val, j] - z[val, i] + etol),
-                    )
-
 
 @pytest.mark.parametrize("dtype", ["float32", "float64"])
 def test_two_dimentional(dtype):
@@ -270,6 +248,26 @@ def test_two_dimentional(dtype):
     assert z.shape == tf_sparsemax_out.shape
 
 
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_diffrence(dtype):
+    """check sparsemax proposition 4."""
+    random = np.random.RandomState(7)
+
+    z = random.uniform(low=-3, high=3, size=(test_obs, 10))
+    p = sparsemax(z.astype(dtype)).numpy()
+
+    etol = {np.float32: 1e-6, np.float64: 1e-9}[dtype]
+
+    for val in range(0, test_obs):
+        for i in range(0, 10):
+            for j in range(0, 10):
+                # check condition, the obesite pair will be checked anyway
+                if z[val, i] > z[val, j]:
+                    continue
+
+                assert 0 <= p[val, j] - p[val, i] <= z[val, j] - z[val, i] + etol
+
+
 @pytest.mark.parametrize("dtype", ["float32", "float64"])
 def test_gradient_against_estimate(dtype):
     """check sparsemax Rop, against estimated Rop."""