Add atol arg to all calls to np.testing.assert_allclose (defaults to 0 => bogus errors for values close to 0)

Author: mdbenito

tests/influence/torch/test_influence_model.py

@@ -527,15 +527,19 @@ def test_influences_lissa(
         influence_factors, x_train, y_train, mode=test_case.mode
     )
 
+    atol = 1e-5
+    rtol = 1e-4
     assert torch.allclose(
-        influences_from_factors, approx_influences, atol=1e-5, rtol=1e-4
+        influences_from_factors, approx_influences, atol=atol, rtol=rtol
     )
 
     approx_influences = approx_influences.cpu().numpy()
 
     assert not np.any(np.isnan(approx_influences))
 
-    np.testing.assert_allclose(approx_influences, direct_influences, rtol=1e-1)
+    np.testing.assert_allclose(
+        approx_influences, direct_influences, atol=atol, rtol=rtol
+    )
 
     if test_case.mode == InfluenceMode.Up:
         assert approx_influences.shape == (
@@ -553,7 +557,9 @@ def test_influences_lissa(
     # check that influences are not all constant
     assert not np.all(approx_influences == approx_influences.item(0))
 
-    np.testing.assert_allclose(approx_influences, direct_influences, rtol=1e-1)
+    np.testing.assert_allclose(
+        approx_influences, direct_influences, atol=atol, rtol=rtol
+    )
 
 
 @pytest.mark.parametrize(
@@ -674,6 +680,9 @@ def test_influences_ekfac(
     direct_sym_influences,
     device: torch.device,
 ):
+    atol = 1e-6
+    rtol = 1e-4
+
     model, loss, x_train, y_train, x_test, y_test = model_and_data
 
     train_dataloader = DataLoader(
@@ -731,8 +740,12 @@ def test_influences_ekfac(
             .numpy()
         )
 
-        np.testing.assert_allclose(ekfac_influence_values, influence_from_factors)
-        np.testing.assert_allclose(ekfac_influence_values, accumulated_inf_by_layer)
+        np.testing.assert_allclose(
+            ekfac_influence_values, influence_from_factors, atol=atol, rtol=rtol
+        )
+        np.testing.assert_allclose(
+            ekfac_influence_values, accumulated_inf_by_layer, atol=atol, rtol=rtol
+        )
         check_influence_correlations(
             direct_influences.numpy(), ekfac_influence_values, threshold=0.94
         )
@@ -832,7 +845,7 @@ def test_influences_cg(
             .numpy()
         )
         np.testing.assert_allclose(
-            single_influence, direct_factors[0], atol=1e-6, rtol=1e-4
+            single_influence[0], direct_factors[0], atol=1e-6, rtol=1e-4
         )
 
 

tests/utils/test_numeric.py

@@ -207,7 +207,7 @@ def test_random_matrix_with_condition_number(n, cond, exception):
     else:
         mat = random_matrix_with_condition_number(n, cond)
         (
-            np.testing.assert_allclose(np.linalg.cond(mat), cond),
+            np.testing.assert_allclose(np.linalg.cond(mat), cond, atol=1e-5),
             "Condition number does not match",
         )
         assert np.array_equal(mat, mat.T), "Matrix is not symmetric"
@@ -272,8 +272,8 @@ def test_running_moments():
 
         true_means = [np.mean(vv) for vv in values]
         true_variances = [np.var(vv) for vv in values]
-        np.testing.assert_allclose(means, true_means)
-        np.testing.assert_allclose(variances, true_variances)
+        np.testing.assert_allclose(means, true_means, atol=1e-5)
+        np.testing.assert_allclose(variances, true_variances, atol=1e-5)
 
 
 def test_running_moment_initialization():

tests/valuation/samplers/test_sampler.py

@@ -626,14 +626,14 @@ def test_antithetic_owen_sampler():
     # check that the sample sizes are close to expected sizes
     sizes = np.array([len(sample.subset) for sample in samples])
     avg_sizes = sizes.reshape(n_outer, -1).mean(axis=1)
-    np.testing.assert_allclose(avg_sizes, len(indices) // 2, rtol=0.01)
+    np.testing.assert_allclose(avg_sizes, len(indices) // 2, rtol=0.01, atol=1e-5)
 
 
 def _check_sample_sizes(samples, n_samples_outer, n_indices, probs):
     sizes = np.array([len(sample.subset) for sample in samples])
     avg_sizes = sizes.reshape(n_samples_outer, -1).mean(axis=1)
     expected_sizes = probs * n_indices  # mean of Binomial(n_indices, probs)
-    np.testing.assert_allclose(avg_sizes, expected_sizes, rtol=0.01)
+    np.testing.assert_allclose(avg_sizes, expected_sizes, rtol=0.01, atol=1e-5)
 
 
 @pytest.mark.flaky(reruns=1)

tests/valuation/test_result.py

@@ -265,7 +265,7 @@ def test_from_random_creation(size: int, total: float | None):
     assert result.status == Status.Converged
     assert result.algorithm == "random"
     if total is not None:
-        np.testing.assert_allclose(np.sum(result.values), total)
+        np.testing.assert_allclose(np.sum(result.values), total, atol=1e-5)
 
 
 def test_from_random_creation_errors():
@@ -300,8 +300,10 @@ def test_adding_random():
     true_means = values.mean(axis=1)
     true_variances = values.var(axis=1)
 
-    np.testing.assert_allclose(true_means[result.indices], result.values)
-    np.testing.assert_allclose(true_variances[result.indices], result.variances)
+    np.testing.assert_allclose(true_means[result.indices], result.values, atol=1e-5)
+    np.testing.assert_allclose(
+        true_variances[result.indices], result.variances, atol=1e-5
+    )
 
 
 @pytest.mark.parametrize(
@@ -378,7 +380,7 @@ def test_adding_different_indices(
     v3 = v1 + v2
 
     np.testing.assert_allclose(v3.indices, np.array(expected_indices))
-    np.testing.assert_allclose(v3.values, np.array(expected_values))
+    np.testing.assert_allclose(v3.values, np.array(expected_values), atol=1e-5)
     assert np.all(v3.names == expected_names)
 
 

tests/value/shapley/test_classwise.py

@@ -209,15 +209,19 @@ def test_classwise_scorer_utility(dataset_left_right_margins):
     in_cls_acc_0, out_of_cls_acc_0 = scorer.estimate_in_class_and_out_of_class_score(
         model, x, y
     )
-    np.testing.assert_allclose(in_cls_acc_0, target_in_cls_acc_0)
-    np.testing.assert_allclose(out_of_cls_acc_0, target_out_of_cls_acc_0)
+    np.testing.assert_allclose(in_cls_acc_0, target_in_cls_acc_0, atol=1e-5)
+    np.testing.assert_allclose(out_of_cls_acc_0, target_out_of_cls_acc_0, atol=1e-5)
 
     value = scorer(model, x, y)
-    np.testing.assert_allclose(value, in_cls_acc_0 * np.exp(out_of_cls_acc_0))
+    np.testing.assert_allclose(
+        value, in_cls_acc_0 * np.exp(out_of_cls_acc_0), atol=1e-5
+    )
 
     scorer.label = 1
     value = scorer(model, x, y)
-    np.testing.assert_allclose(value, out_of_cls_acc_0 * np.exp(in_cls_acc_0))
+    np.testing.assert_allclose(
+        value, out_of_cls_acc_0 * np.exp(in_cls_acc_0), atol=1e-5
+    )
 
 
 @pytest.mark.parametrize("n_element, left_margin, right_margin", [(101, 0.3, 0.4)])
@@ -329,8 +333,8 @@ def test_classwise_scorer_accuracies_left_right_margins(dataset_left_right_margi
     in_cls_acc_0, out_of_cls_acc_0 = scorer.estimate_in_class_and_out_of_class_score(
         model, x, y
     )
-    np.testing.assert_allclose(in_cls_acc_0, target_in_cls_acc_0)
-    np.testing.assert_allclose(out_of_cls_acc_0, target_out_of_cls_acc_0)
+    np.testing.assert_allclose(in_cls_acc_0, target_in_cls_acc_0, atol=1e-5)
+    np.testing.assert_allclose(out_of_cls_acc_0, target_out_of_cls_acc_0, atol=1e-5)
 
 
 def test_closed_form_linear_classifier(

tests/value/test_results.py

@@ -296,8 +296,10 @@ def test_adding_random():
     true_means = values.mean(axis=1)
     true_variances = values.var(axis=1)
 
-    np.testing.assert_allclose(true_means[result.indices], result.values)
-    np.testing.assert_allclose(true_variances[result.indices], result.variances)
+    np.testing.assert_allclose(true_means[result.indices], result.values, atol=1e-5)
+    np.testing.assert_allclose(
+        true_variances[result.indices], result.variances, atol=1e-5
+    )
 
 
 @pytest.mark.parametrize(
@@ -374,7 +376,7 @@ def test_adding_different_indices(
     v3 = v1 + v2
 
     np.testing.assert_allclose(v3.indices, np.array(expected_indices))
-    np.testing.assert_allclose(v3.values, np.array(expected_values))
+    np.testing.assert_allclose(v3.values, np.array(expected_values), atol=1e-5)
     assert np.all(v3.names == expected_names)