Added serializability tests in all optimizers and fixed serializability bug in yogi

tensorflow_addons/optimizers/tests/conditional_gradient_test.py

@@ -1457,3 +1457,15 @@ def test_sparse_nuclear():
             ),
             var1[2].numpy(),
         )
+
+
+def test_serialization():
+    learning_rate = 0.1
+    lambda_ = 0.1
+    ord = "nuclear"
+    optimizer = cg_lib.ConditionalGradient(
+        learning_rate=learning_rate, lambda_=lambda_, ord=ord
+    )
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert optimizer.get_config() == new_optimizer.get_config()

tensorflow_addons/optimizers/tests/lamb_test.py

@@ -345,3 +345,10 @@ def test_exclude_layer_adaptation():
     assert opt._do_layer_adaptation("var0")
     assert not opt._do_layer_adaptation("var1")
     assert not opt._do_layer_adaptation("var1_weight")
+
+
+def test_serialization():
+    optimizer = lamb.LAMB(1e-4)
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/tests/lazy_adam_test.py

@@ -253,3 +253,10 @@ def test_slots_unique_eager():
     # There should be iteration, and two unique slot variables for v1 and v2.
     assert 5 == len(opt.variables())
     assert opt.variables()[0] == opt.iterations
+
+
+def test_serialization():
+    optimizer = lazy_adam.LazyAdam()
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/tests/lookahead_test.py

@@ -149,3 +149,10 @@ def test_get_config():
     config = opt.get_config()
     assert config["sync_period"] == 10
     assert config["slow_step_size"] == 0.4
+
+
+def test_serialization():
+    optimizer = Lookahead("adam", sync_period=10, slow_step_size=0.4)
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/tests/moving_average_test.py

@@ -169,3 +169,13 @@ def test_fit_simple_linear_model():
 
     max_abs_diff = np.max(np.abs(predicted - y))
     assert max_abs_diff < 5e-3
+
+
+def test_serialization():
+    sgd_opt = tf.keras.optimizers.SGD(lr=2.0, nesterov=True, momentum=0.3, decay=0.1)
+    optimizer = MovingAverage(
+        sgd_opt, average_decay=0.5, num_updates=None, sequential_update=False
+    )
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/tests/novograd_test.py

@@ -139,3 +139,10 @@ def test_get_config():
     assert config["learning_rate"] == 1e-4
     assert config["weight_decay"] == 0.0
     assert config["grad_averaging"] is False
+
+
+def test_serialization():
+    optimizer = NovoGrad(lr=1e-4, weight_decay=0.0, grad_averaging=False)
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/tests/stochastic_weight_averaging_test.py

@@ -113,3 +113,13 @@ def test_fit_simple_linear_model():
 
     max_abs_diff = np.max(np.abs(predicted - y))
     assert max_abs_diff < 1e-3
+
+
+def test_serialization():
+    start_averaging = 0
+    average_period = 1
+    sgd = tf.keras.optimizers.SGD(lr=1.0)
+    optimizer = SWA(sgd, start_averaging, average_period)
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/tests/weight_decay_optimizers_test.py

@@ -348,3 +348,10 @@ def test_optimizer_sparse(dtype, optimizer):
     do_test_sparse_repeated_indices(
         dtype, optimizer, learning_rate=0.001, momentum=0.9, weight_decay=WEIGHT_DECAY,
     )
+
+
+def test_serialization():
+    optimizer = weight_decay_optimizers.AdamW(learning_rate=1e-4, weight_decay=1e-4)
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/tests/yogi_test.py

@@ -337,3 +337,10 @@ def test_get_config():
     opt = yogi.Yogi(1e-4)
     config = opt.get_config()
     assert config["learning_rate"] == 1e-4
+
+
+def test_serialization():
+    optimizer = yogi.Yogi(1e-4)
+    config = tf.keras.optimizers.serialize(optimizer)
+    new_optimizer = tf.keras.optimizers.deserialize(config)
+    assert new_optimizer.get_config() == optimizer.get_config()

tensorflow_addons/optimizers/yogi.py

@@ -320,8 +320,12 @@ def get_config(self):
                 "beta1": self._serialize_hyperparameter("beta_1"),
                 "beta2": self._serialize_hyperparameter("beta_2"),
                 "epsilon": self._serialize_hyperparameter("epsilon"),
-                "l1_t": self._serialize_hyperparameter("l1_regularization_strength"),
-                "l2_t": self._serialize_hyperparameter("l2_regularization_strength"),
+                "l1_regularization_strength": self._serialize_hyperparameter(
+                    "l1_regularization_strength"
+                ),
+                "l2_regularization_strength": self._serialize_hyperparameter(
+                    "l2_regularization_strength"
+                ),
                 "activation": self._activation,
                 "initial_accumulator_value": self._initial_accumulator_value,
             }