Python 将GridSearchCV与keras函数API一起使用

在线上有很多关于如何将GridSearchCV与顺序API一起使用的教程，但关于函数API的教程很少

运行grid.fit时，以下代码失败

问题是，对于输入，应该向

tensorflow.keras.Model

输入什么

model = tensorflow.keras.Model(inputs=inputs, outputs=core_out)

通常对于keras项目，您可以传递一个输入层列表（在我的代码中由

inputs\u和\u encoded\u features（）

生成），但我认为您不能将其传递到

问题可能来自于尝试使用

tensorflow.keras.layers.experimental.preprocessing

接口来生成传递给模型的编码特征

def build_functional_model(encoded_features, g1dim=10, g2dim=1, c1dim=2500, c2dim=500, c3dim=1, optimizer='ada_grad', learning_rate=0.01, ca='relu' ):
    # Attempt 1: Fails with error: ValueError: A `Concatenate` layer should be called on a list of at least 2 inputs
    # all_features = tf.keras.layers.concatenate(encoded_features)

    # Attempt 2: Fails with error: ValueError: A `Concatenate` layer should be called on a list of at least 2 inputs
    inputs = Input(shape=(len(encoded_features.graph.inputs,)))
    all_features = tf.keras.layers.concatenate(inputs)

    other_norm = Normalization(name='other_norm')(all_features)
    core_layer1 = Dense(c1dim, activation=ca, name='core_layer1')(other_norm)
    core_layer2 = tf.keras.layers.Dropout(0.2)(core_layer1)
    core_layer3 = Dense(c2dim, activation=ca, name='core_layer3')(core_layer2)
    core_out = Dense(1, name='core_out')(core_layer3)

    model = tensorflow.keras.Model(inputs=inputs, outputs=core_out)

    if optimizer == 'ada_grad':
        _opt = Adagrad(learning_rate=lr_schedule)
    elif optimizer == 'SGD':
        _opt = SGD(learning_rate=lr_schedule)
    elif optimizer == 'RMSprop':
        _opt = RMSprop(learning_rate=lr_schedule)
    else: # optimizer == 'adam'
        _opt = Adam(learning_rate=lr_schedule)
    model.compile(loss='mse', optimizer=_opt, metrics=[MeanSquaredError(), MeanAbsolutePercentageError()]) # CosineSimilarity(), MeanAbsoluteError(), 
    tf.keras.utils.plot_model(model, 'model.png', show_shapes=False, rankdir='LR')
    return model

def grid_search(X, Y, encoded_features):
    lrs = [0.3, 0.1, 0.01]
    c1dim = [1000, 500, 100]
    c2dim = [300, 100, 20]
    activations = ['relu', 'tanh', 'softplus' 'elu']
    optimizers = ['ada_grad','SGD','RMSprop','adam']

    model_grid = KerasRegressor(build_fn=build_functional_model, verbose=0) # Only works if build_fn is not within a class
    param_grid = dict(ca=activations,
                    optimizer=optimizers, 
                    encoded_features=encoded_features,
                    epochs=[250], 
                    batch_size=[256], 
                    )
    grid = GridSearchCV(estimator=model_grid, 
                        param_grid=param_grid,
                        cv = 3,
                        verbose=2, n_jobs=1) # n_jobs=1 required as keras currently doesn't support pickle which is needed for multiprocessing
    grid_result = grid.fit(X, Y)

class () ...
    ...
    def inputs_and_encoded_features(self):
        def add_column(header, layer, dtype=None):
            if not dtype:
                column = tf.keras.Input(shape=(1,), name=header)
            else:
                column = tf.keras.Input(shape=(1,), name=header, dtype=dtype)
            encoded_col = layer(column)
            inputs.append(column)
            encoded_features.append(encoded_col) # we might want a seperate gps_features if we want to feed it in seperately

        # build feature columns        
        inputs = []
        encoded_features = []
        for header in self.gps_columns:
            normalization_layer = self.get_normalization_layer(header, self.train_ds)
            add_column(header, normalization_layer)
        for header in self.numeric_columns:
            normalization_layer = self.get_normalization_layer(header, self.train_ds)
            add_column(header, normalization_layer)
        for header in self.word_columns:
            normalization_layer = self.get_normalization_layer(header, self.train_ds)
            add_column(header, normalization_layer) # we might want a seperate gps_features if we want to feed it in seperately
        for header in self.category_columns:
            encoding_layer = self.get_category_encoding_layer(header, self.train_ds, dtype='string',max_tokens=5)
            add_column(header, encoding_layer, dtype='string')
        self.inputs = inputs
        self.encoded_features = encoded_features
        return encoded_features