test normalizer all close value

tests/test_normalizers.py

@@ -104,7 +104,7 @@ def test_kernel_svd(kernel_shape):
     )
     # Test if kernel is normalized by sigma
     np.testing.assert_allclose(
-        np.reshape(W_bar, kernel.shape), kernel / sigmas_svd[0], 1e-2, 0
+        np.reshape(W_bar, kernel.shape), kernel / sigmas_svd[0], atol=1e-2
     )
 
 
@@ -211,7 +211,9 @@ def test_kernel_conv_svd(kernel_shape, strides):
         sigma, SVmax, 2, "test failed with kernel_shape " + str(kernel.shape)
     )
     # Test if kernel is normalized by sigma
-    np.testing.assert_allclose(np.reshape(W_bar, kernel.shape), kernel / SVmax, 1e-2, 0)
+    np.testing.assert_allclose(
+        np.reshape(W_bar, kernel.shape), kernel / SVmax, atol=1e-2
+    )
 
 
 @pytest.mark.parametrize(
@@ -244,7 +246,9 @@ def test_bjorck_normalization(kernel_shape):
         compute_uv=False,
     )
     # Test sigma is close to the one computed with svd first run @ 1e-1
-    np.testing.assert_allclose(sigmas_wbar_svd, np.ones(sigmas_wbar_svd.shape), 1e-2, 0)
+    np.testing.assert_allclose(
+        sigmas_wbar_svd, np.ones(sigmas_wbar_svd.shape), atol=1e-2
+    )
     # Test W_bar is reshaped correctly
     np.testing.assert_equal(wbar.shape, (np.prod(kernel.shape[:-1]), kernel.shape[-1]))
 
@@ -256,7 +260,9 @@ def test_bjorck_normalization(kernel_shape):
         full_matrices=False,
         compute_uv=False,
     )
-    np.testing.assert_allclose(sigmas_wbar_svd, np.ones(sigmas_wbar_svd.shape), 1e-4, 0)
+    np.testing.assert_allclose(
+        sigmas_wbar_svd, np.ones(sigmas_wbar_svd.shape), atol=1e-4
+    )
 
 
 @pytest.mark.parametrize(
@@ -306,7 +312,9 @@ def test_reshaped_kernel_orthogonalization(kernel_shape):
         compute_uv=False,
     )
     # Test if SVs of W_bar are close to one
-    np.testing.assert_allclose(sigmas_wbar_svd, np.ones(sigmas_wbar_svd.shape), 1e-2, 0)
+    np.testing.assert_allclose(
+        sigmas_wbar_svd, np.ones(sigmas_wbar_svd.shape), atol=1e-2
+    )
 
 
 @pytest.mark.skipif(
@@ -340,14 +348,18 @@ def test_bjorck_norm():
     x = np.random.rand(2)
     x = uft.to_tensor(x)
     m(x)
-    np.testing.assert_equal(uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)))
+    np.testing.assert_allclose(
+        uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)), atol=1e-5
+    )
 
     # remove the parametrization
     uft.get_instance_framework(remove_bjorck_norm, {"module": m})  # (m)
     uft.check_parametrization(m, is_parametrized=False)
     # assert not hasattr(m, "parametrizations")
     # assert isinstance(m.weight, torch.nn.Parameter)
-    np.testing.assert_equal(uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)))
+    np.testing.assert_allclose(
+        uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)), atol=1e-5
+    )
 
 
 @pytest.mark.skipif(
@@ -379,12 +391,16 @@ def test_frobenius_norm():
     x = np.random.rand(2)
     x = uft.to_tensor(x)
     m(x)
-    np.testing.assert_equal(uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)))
+    np.testing.assert_allclose(
+        uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)), atol=1e-5
+    )
 
     # remove the parametrization
     uft.get_instance_framework(remove_frobenius_norm, {"module": m})
     uft.check_parametrization(m, is_parametrized=False)
-    np.testing.assert_equal(uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)))
+    np.testing.assert_allclose(
+        uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)), atol=1e-5
+    )
 
 
 @pytest.mark.skipif(
@@ -459,4 +475,6 @@ def test_lconv_norm():
     uft.get_instance_framework(remove_lconv_norm, {"module": m})
     uft.check_parametrization(m, is_parametrized=False)
     # assert isinstance(m.weight, torch.nn.Parameter)
-    np.testing.assert_equal(uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)))
+    np.testing.assert_allclose(
+        uft.to_numpy(w1), uft.to_numpy(uft.get_layer_weights(m)), atol=1e-7
+    )

tests/utils_framework.py

@@ -109,6 +109,7 @@
     "Loss",
 ]
 
+
 # not implemented
 def module_Unavailable(**kwargs):
     return None