Do not directly pass an optimizer to CmpNet

An optimizer is a tensorflow object, which (at least in graph mode in
tf1) is not deepcopy-able. Even if we were able to deeocopy it, we
probably wouldn't want to since it contains state. Scikit-learn needs to
be able to deepcopy an estimators arguments so that it can create copies
and derivatives of it.

Instead we pass a function that constructs the optimizer on-demand. This
function is deepcopy-able.

csrank/choicefunction/cmpnet_choice.py

@@ -19,7 +19,7 @@ def __init__(
         kernel_regularizer=l2(1e-4),
         kernel_initializer="lecun_normal",
         activation="relu",
-        optimizer=SGD(lr=1e-4, nesterov=True, momentum=0.9),
+        optimizer=lambda: SGD(lr=1e-4, nesterov=True, momentum=0.9),
         metrics=["binary_accuracy"],
         batch_size=256,
         random_state=None,

csrank/core/cmpnet_core.py

@@ -4,7 +4,6 @@
 from keras import backend as K
 from keras import Input
 from keras import Model
-from keras import optimizers
 from keras.layers import concatenate
 from keras.layers import Dense
 from keras.optimizers import SGD
@@ -29,7 +28,7 @@ def __init__(
         kernel_regularizer=l2(1e-4),
         kernel_initializer="lecun_normal",
         activation="relu",
-        optimizer=SGD(lr=1e-4, nesterov=True, momentum=0.9),
+        optimizer=lambda: SGD(lr=1e-4, nesterov=True, momentum=0.9),
         metrics=["binary_accuracy"],
         batch_size=256,
         random_state=None,
@@ -47,8 +46,7 @@ def __init__(
         self.kernel_initializer = kernel_initializer
         self.loss_function = loss_function
 
-        self.optimizer = optimizers.get(optimizer)
-        self._optimizer_config = self.optimizer.get_config()
+        self.optimizer = optimizer
 
         self.n_hidden = n_hidden
         self.n_units = n_units
@@ -97,6 +95,7 @@ def construct_model(self):
             model: keras :class:`Model`
                 Neural network to learn the CmpNet utility score
         """
+        self.optimizer_ = self.optimizer()  # instantiate the optimizer
         x1x2 = concatenate([self.x1, self.x2])
         x2x1 = concatenate([self.x2, self.x1])
         self.logger.debug("Creating the model")
@@ -110,7 +109,7 @@ def construct_model(self):
         merged_output = concatenate([N_g, N_l])
         model = Model(inputs=[self.x1, self.x2], outputs=merged_output)
         model.compile(
-            loss=self.loss_function, optimizer=self.optimizer, metrics=self.metrics
+            loss=self.loss_function, optimizer=self.optimizer_, metrics=self.metrics
         )
         return model
 
@@ -212,7 +211,7 @@ def clear_memory(self, **kwargs):
             sess = tf.Session()
             K.set_session(sess)
 
-            self.optimizer = self.optimizer.from_config(self._optimizer_config)
+            self.optimizer_ = self.optimizer()
             self._construct_layers(
                 kernel_regularizer=self.kernel_regularizer,
                 kernel_initializer=self.kernel_initializer,
@@ -255,8 +254,8 @@ def set_tunable_parameters(
         self.n_units = n_units
         self.kernel_regularizer = l2(reg_strength)
         self.batch_size = batch_size
-        self.optimizer = self.optimizer.from_config(self._optimizer_config)
-        K.set_value(self.optimizer.lr, learning_rate)
+        self.optimizer_ = self.optimizer()
+        K.set_value(self.optimizer_.lr, learning_rate)
         self._construct_layers(
             kernel_regularizer=self.kernel_regularizer,
             kernel_initializer=self.kernel_initializer,

csrank/discretechoice/cmpnet_discrete_choice.py

@@ -18,7 +18,7 @@ def __init__(
         kernel_regularizer=l2(1e-4),
         kernel_initializer="lecun_normal",
         activation="relu",
-        optimizer=SGD(lr=1e-4, nesterov=True, momentum=0.9),
+        optimizer=lambda: SGD(lr=1e-4, nesterov=True, momentum=0.9),
         metrics=["binary_accuracy"],
         batch_size=256,
         random_state=None,

csrank/objectranking/cmp_net.py

@@ -20,7 +20,7 @@ def __init__(
         kernel_regularizer=l2(1e-4),
         kernel_initializer="lecun_normal",
         activation="relu",
-        optimizer=SGD(lr=1e-4, nesterov=True, momentum=0.9),
+        optimizer=lambda: SGD(lr=1e-4, nesterov=True, momentum=0.9),
         metrics=["binary_accuracy"],
         batch_size=256,
         random_state=None,
@@ -62,8 +62,10 @@ def __init__(
                Regularizer function applied to all the hidden weight matrices.
            activation : function or string
                Type of activation function to use in each hidden layer
-           optimizer : function or string
-               Optimizer to use during stochastic gradient descent
+           optimizer : function
+               Constructor for an optimizer to use during stochastic gradient
+               descent. Must be a function without arguments that returns a
+               Keras optimizer.
            metrics : list
                List of metrics to evaluate during training (can be
                non-differentiable)

csrank/tests/test_choice_functions.py

@@ -68,7 +68,7 @@ def get_vals(values):
     ),
     CMPNET_CHOICE: (
         CmpNetChoiceFunction,
-        {"optimizer": optimizer},
+        {"optimizer": lambda: optimizer},
         get_vals([0.8554, 0.8649, 0.966]),
     ),
     GLM_CHOICE: (GeneralizedLinearModel, {}, get_vals([0.9567, 0.9955, 1.0])),

csrank/tests/test_discrete_choice.py

@@ -50,7 +50,7 @@ def get_vals(values=[1.0, 1.0]):
     ),
     CMPNET_DC: (
         CmpNetDiscreteChoiceFunction,
-        {"optimizer": optimizer},
+        {"optimizer": lambda: optimizer},
         get_vals([0.994, 1.0]),
     ),
     FATE_DC: (

csrank/tests/test_ranking.py

@@ -35,7 +35,7 @@
         (0.0, 1.0),
     ),
     RANKNET: (RankNet, {"optimizer": optimizer}, (0.0, 1.0)),
-    CMPNET: (CmpNet, {"optimizer": optimizer}, (0.0, 1.0)),
+    CMPNET: (CmpNet, {"optimizer": lambda: optimizer}, (0.0, 1.0)),
     LISTNET: (ListNet, {"n_top": 3, "optimizer": optimizer}, (0.0, 1.0)),
     ERR: (ExpectedRankRegression, {}, (0.0, 1.0)),
     RANKSVM: (RankSVM, {}, (0.0, 1.0)),
@@ -79,7 +79,11 @@ def check_params_tunable(tunable_obj, params, rtol=1e-2, atol=1e-4):
                     )
             else:
                 assert value == expected
-        elif key == "learning_rate" and "optimizer" in tunable_obj.__dict__.keys():
+        elif (
+            key == "learning_rate"
+            and "optimizer" in tunable_obj.__dict__.keys()
+            and hasattr(tunable_obj.optimizer, "get_config")
+        ):
             key = (
                 "learning_rate"
                 if "learning_rate" in tunable_obj.optimizer.get_config().keys()