[python][tests] minor Python tests cleanup (#3860)

* Update test_engine.py

* Update test_sklearn.py

* Update test_engine.py

* Update test_sklearn.py

* Update test_sklearn.py

* Update test_sklearn.py

* Update test_sklearn.py

* Update test_engine.py

* Update .vsts-ci.yml

* Update .vsts-ci.yml

* Update test_engine.py

* Update test_dual.py

* Update test_engine.py

* Update .vsts-ci.yml

* Update .vsts-ci.yml

tests/python_package_test/test_dual.py

@@ -31,5 +31,5 @@ def test_cpu_and_gpu_work():
     gpu_bst = lgb.train(params_gpu, data, num_boost_round=10)
     gpu_score = log_loss(y, gpu_bst.predict(X))
 
-    np.testing.assert_allclose(cpu_score, gpu_score)
+    assert cpu_score == pytest.approx(gpu_score)
     assert gpu_score < 0.242

tests/python_package_test/test_engine.py

@@ -950,7 +950,7 @@ def test_pandas_categorical():
     with pytest.raises(AssertionError):
         np.testing.assert_allclose(pred0, pred7)  # ordered cat features aren't treated as cat features by default
     with pytest.raises(AssertionError):
-        np.testing.assert_allclose(pred0, pred8)  # ordered cat features aren't treated as cat features by default
+        np.testing.assert_allclose(pred0, pred8)
     assert gbm0.pandas_categorical == cat_values
     assert gbm1.pandas_categorical == cat_values
     assert gbm2.pandas_categorical == cat_values
@@ -2500,7 +2500,7 @@ def test_linear_trees(tmp_path):
     est = lgb.train(dict(params, linear_tree=True), lgb_train, num_boost_round=10, evals_result=res,
                     valid_sets=[lgb_train], valid_names=['train'])
     pred2 = est.predict(x)
-    np.testing.assert_allclose(res['train']['l2'][-1], mean_squared_error(y, pred2), atol=10**(-1))
+    assert res['train']['l2'][-1] == pytest.approx(mean_squared_error(y, pred2), abs=1e-1)
     assert mean_squared_error(y, pred2) < mean_squared_error(y, pred1)
     # test again with nans in data
     x[:10] = np.nan
@@ -2512,14 +2512,14 @@ def test_linear_trees(tmp_path):
     est = lgb.train(dict(params, linear_tree=True), lgb_train, num_boost_round=10, evals_result=res,
                     valid_sets=[lgb_train], valid_names=['train'])
     pred2 = est.predict(x)
-    np.testing.assert_allclose(res['train']['l2'][-1], mean_squared_error(y, pred2), atol=10**(-1))
+    assert res['train']['l2'][-1] == pytest.approx(mean_squared_error(y, pred2), abs=1e-1)
     assert mean_squared_error(y, pred2) < mean_squared_error(y, pred1)
     # test again with bagging
     res = {}
     est = lgb.train(dict(params, linear_tree=True, subsample=0.8, bagging_freq=1), lgb_train,
                     num_boost_round=10, evals_result=res, valid_sets=[lgb_train], valid_names=['train'])
     pred = est.predict(x)
-    np.testing.assert_allclose(res['train']['l2'][-1], mean_squared_error(y, pred), atol=10**(-1))
+    assert res['train']['l2'][-1] == pytest.approx(mean_squared_error(y, pred), abs=1e-1)
     # test with a feature that has only one non-nan value
     x = np.concatenate([np.ones([x.shape[0], 1]), x], 1)
     x[500:, 1] = np.nan
@@ -2529,7 +2529,7 @@ def test_linear_trees(tmp_path):
     est = lgb.train(dict(params, linear_tree=True, subsample=0.8, bagging_freq=1), lgb_train,
                     num_boost_round=10, evals_result=res, valid_sets=[lgb_train], valid_names=['train'])
     pred = est.predict(x)
-    np.testing.assert_allclose(res['train']['l2'][-1], mean_squared_error(y, pred), atol=10**(-1))
+    assert res['train']['l2'][-1] == pytest.approx(mean_squared_error(y, pred), abs=1e-1)
     # test with a categorical feature
     x[:250, 0] = 0
     y[:250] += 10

tests/python_package_test/test_sklearn.py

@@ -86,7 +86,7 @@ def test_binary():
     gbm.fit(X_train, y_train, eval_set=[(X_test, y_test)], early_stopping_rounds=5, verbose=False)
     ret = log_loss(y_test, gbm.predict_proba(X_test))
     assert ret < 0.12
-    assert ret == pytest.approx(gbm.evals_result_['valid_0']['binary_logloss'][gbm.best_iteration_ - 1], abs=1e-5)
+    assert gbm.evals_result_['valid_0']['binary_logloss'][gbm.best_iteration_ - 1] == pytest.approx(ret)
 
 
 def test_regression():
@@ -96,7 +96,7 @@ def test_regression():
     gbm.fit(X_train, y_train, eval_set=[(X_test, y_test)], early_stopping_rounds=5, verbose=False)
     ret = mean_squared_error(y_test, gbm.predict(X_test))
     assert ret < 7
-    assert ret == pytest.approx(gbm.evals_result_['valid_0']['l2'][gbm.best_iteration_ - 1], abs=1e-5)
+    assert gbm.evals_result_['valid_0']['l2'][gbm.best_iteration_ - 1] == pytest.approx(ret)
 
 
 def test_multiclass():
@@ -108,7 +108,7 @@ def test_multiclass():
     assert ret < 0.05
     ret = multi_logloss(y_test, gbm.predict_proba(X_test))
     assert ret < 0.16
-    assert ret == pytest.approx(gbm.evals_result_['valid_0']['multi_logloss'][gbm.best_iteration_ - 1], abs=1e-5)
+    assert gbm.evals_result_['valid_0']['multi_logloss'][gbm.best_iteration_ - 1] == pytest.approx(ret)
 
 
 def test_lambdarank():
@@ -152,7 +152,7 @@ def test_regression_with_custom_objective():
     gbm.fit(X_train, y_train, eval_set=[(X_test, y_test)], early_stopping_rounds=5, verbose=False)
     ret = mean_squared_error(y_test, gbm.predict(X_test))
     assert ret < 7.0
-    assert ret == pytest.approx(gbm.evals_result_['valid_0']['l2'][gbm.best_iteration_ - 1], abs=1e-5)
+    assert gbm.evals_result_['valid_0']['l2'][gbm.best_iteration_ - 1] == pytest.approx(ret)
 
 
 def test_binary_classification_with_custom_objective():
@@ -199,10 +199,8 @@ def test_stacking_classifier():
     assert clf.named_estimators_['gbm1'].n_features_in_ == clf.named_estimators_['gbm2'].n_features_in_
     assert clf.final_estimator_.n_features_in_ == 10  # number of concatenated features
     assert len(clf.final_estimator_.feature_importances_) == 10
-    classes = clf.named_estimators_['gbm1'].classes_ == clf.named_estimators_['gbm2'].classes_
-    assert all(classes)
-    classes = clf.classes_ == clf.named_estimators_['gbm1'].classes_
-    assert all(classes)
+    assert all(clf.named_estimators_['gbm1'].classes_ == clf.named_estimators_['gbm2'].classes_)
+    assert all(clf.classes_ == clf.named_estimators_['gbm1'].classes_)
 
 
 # sklearn <0.23 does not have a stacking regressor and n_features_in_ property
@@ -995,6 +993,7 @@ def test_first_metric_only():
                 expected = assumed_iteration + (params_fit['early_stopping_rounds']
                                                 if eval_set_name != 'training'
                                                 and assumed_iteration != gbm.n_estimators else 0)
+                assert expected == actual
                 if eval_set_name != 'training':
                     assert assumed_iteration == gbm.best_iteration_
                 else: