Update workflows for TensorFlow/Keras 2.14

Signed-off-by: Beat Buesser <[email protected]>

tests/attacks/test_adversarial_patch.py

@@ -301,7 +301,7 @@ def test_4_pytorch(self):
 
         self.assertAlmostEqual(patch_adv[0, 8, 8], 0.5, delta=0.05)
         self.assertAlmostEqual(patch_adv[0, 14, 14], 0.5, delta=0.05)
-        self.assertAlmostEqual(float(np.sum(patch_adv)), 371.88014772999827, delta=4.0)
+        self.assertAlmostEqual(float(np.sum(patch_adv)), 367.6218066346819, delta=4.0)
 
         mask = np.ones((1, 28, 28)).astype(bool)
         attack_ap.apply_patch(x=x_train, scale=0.1, mask=mask)

tests/classifiersFrameworks/test_tensorflow.py

@@ -239,7 +239,7 @@ def test_binary_keras_instantiation_and_attack_pgd(art_warning):
             ]
         )
         model.summary()
-        model.compile(optimizer=tf.optimizers.Adam(), loss="binary_crossentropy", metrics=["accuracy"])
+        model.compile(optimizer=tf.optimizers.legacy.Adam(), loss="binary_crossentropy", metrics=["accuracy"])
         classifier = KerasClassifier(model=model)
         classifier.fit(train_x, train_y, nb_epochs=5)
         pred = classifier.predict(test_x)
@@ -268,7 +268,7 @@ def test_binary_keras_instantiation_and_attack_pgd(art_warning):
 #             ]
 #         )
 #         model.summary()
-#         model.compile(optimizer=tf.optimizers.Adam(), loss="binary_crossentropy", metrics=["accuracy"])
+#         model.compile(optimizer=tf.optimizers.legacy.Adam(), loss="binary_crossentropy", metrics=["accuracy"])
 #         classifier = art.estimators.classification.TensorFlowV2Classifier(model=model)
 #         classifier.fit(train_x, train_y, nb_epochs=5)
 #         pred = classifier.predict(test_x)

tests/defences/trainer/test_adversarial_trainer_awp_pytorch.py

@@ -124,7 +124,7 @@ def test_adversarial_trainer_awppgd_pytorch_fit_and_predict(get_adv_trainer_awpp
     else:
         accuracy = np.sum(predictions == y_test_mnist) / x_test_mnist.shape[0]
 
-    trainer.fit(x_train_mnist, y_train_mnist, nb_epochs=20)
+    trainer.fit(x_train_mnist, y_train_mnist, nb_epochs=40)
     predictions_new = np.argmax(trainer.predict(x_test_mnist), axis=1)
 
     if label_format == "one_hot":

tests/estimators/certification/test_macer.py

@@ -57,7 +57,7 @@ def _get_classifier():
             import tensorflow as tf
 
             classifier, _ = get_image_classifier_tf()
-            optimizer = tf.keras.optimizers.SGD(learning_rate=0.01, momentum=0.9, name="SGD", decay=5e-4)
+            optimizer = tf.keras.optimizers.legacy.SGD(learning_rate=0.01, momentum=0.9, name="SGD", decay=5e-4)
             scheduler = tf.keras.optimizers.schedules.PiecewiseConstantDecay([250, 400], [0.01, 0.001, 0.0001])
             rs = TensorFlowV2MACER(
                 model=classifier.model,

tests/estimators/certification/test_smooth_adv.py

@@ -57,7 +57,7 @@ def _get_classifier():
             import tensorflow as tf
 
             classifier, _ = get_image_classifier_tf()
-            optimizer = tf.keras.optimizers.SGD(learning_rate=0.01, momentum=0.9, name="SGD", decay=1e-4)
+            optimizer = tf.keras.optimizers.legacy.SGD(learning_rate=0.01, momentum=0.9, name="SGD", decay=1e-4)
             scheduler = tf.keras.optimizers.schedules.PiecewiseConstantDecay([50, 100], [0.01, 0.001, 0.0001])
             rs = TensorFlowV2SmoothAdv(
                 model=classifier.model,

tests/estimators/classification/test_scikitlearn.py

@@ -300,9 +300,10 @@ def test_class_gradient_none_2(self):
 
     def test_class_gradient_int_1(self):
         grad_predicted = self.classifier.class_gradient(self.x_test_iris[0:1], label=1)
-        grad_expected = [[[-0.56322294, -0.70427608, -0.98874801, -0.67053026]]]
+        grad_expected = [[[-0.56322294, -0.70493763, -0.98874801, -0.67053026]]]
 
         for i_shape in range(4):
+            print(grad_predicted[0, 0, i_shape])
             self.assertAlmostEqual(grad_predicted[0, 0, i_shape], grad_expected[0][0][i_shape], 3)
 
     def test_class_gradient_int_2(self):
@@ -311,13 +312,16 @@ def test_class_gradient_int_2(self):
             [[-0.56322294, -0.70427608, -0.98874801, -0.67053026]],
             [[-0.50528532, -0.71700042, -0.82467848, -0.59614766]],
         ]
+        print("grad_predicted")
+        print(grad_predicted)
         np.testing.assert_array_almost_equal(grad_predicted, grad_expected, decimal=4)
 
     def test_class_gradient_list_1(self):
         grad_predicted = self.classifier.class_gradient(self.x_test_iris[0:1], label=[1])
         grad_expected = [[[-0.56322294, -0.70427608, -0.98874801, -0.67053026]]]
 
         for i_shape in range(4):
+            print(grad_predicted[0, 0, i_shape])
             self.assertAlmostEqual(grad_predicted[0, 0, i_shape], grad_expected[0][0][i_shape], 3)
 
     def test_class_gradient_list_2(self):
@@ -326,6 +330,8 @@ def test_class_gradient_list_2(self):
             [[-0.56322294, -0.70427608, -0.98874801, -0.67053026]],
             [[0.70875132, 0.25104877, 1.70929277, 0.88410652]],
         ]
+        print("grad_predicted")
+        print(grad_predicted)
         np.testing.assert_array_almost_equal(grad_predicted, grad_expected, decimal=4)
 
     def test_class_gradient_label_wrong_type(self):
@@ -340,6 +346,8 @@ def test_class_gradient_label_wrong_type(self):
     def test_loss_gradient(self):
         grad_predicted = self.classifier.loss_gradient(self.x_test_iris[0:1], self.y_test_iris[0:1])
         grad_expected = np.asarray([[-0.21693791, -0.08792436, -0.51507443, -0.26990796]])
+        print("grad_predicted")
+        print(grad_predicted)
         np.testing.assert_array_almost_equal(grad_predicted, grad_expected, decimal=4)
 
     def test_save(self):