How to perform KerasClassifier model selection with varying input dimensions

Question

I am trying to perform model selection on a KerasClassifier, for several sliding windows of samples. However, each sliding window has a different input dimension (as a result of feature selection).

The function I have written works for regular scikit-learn classifiers. i.e. it returns a dictionary containing optimal RF models (using random grid search):

# return a dictionary with optimal models for each sliding window
rf_optimal_models = model_selection(RandomForestClassifier(), 
param_distributions = random_grid_rf, n_iter = 10)

However, I am unsure how to set up the KerasClassifier in such a way that it will change the input_dim argument according to the dimensions of the sliding window being passed to it.

The following code sets up the keras scikit-learn wrapper.

def create_model(optimizer='adam', kernel_initializer='normal', dropout_rate=0.0):
    with tf.device("/device:GPU:0"):
        # create model
        model = Sequential()
        model.add(Dense(20, input_dim=X_train.shape[1], activation='relu', kernel_initializer=kernel_initializer))
        model.add(Dropout(dropout_rate))
        model.add(Dense(20, activation='relu'))
        model.add(Dense(1, activation='sigmoid'))
        # Compile model
        model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])
        return model

... and the call to my model_selection() function.

mlp_optimal_models =  model_selection(model = KerasClassifier(build_fn=create_model, verbose=0,), param_distributions = random_grid_mlp, n_iter = 10)

The input_dim argument is static and will throw an error when it receives dimensions of say 49 (the input dim of the next sliding window), but expects 42.

ValueError: Error when checking input: expected dense_1_input to have shape 
(42,) but got array with shape (49,)

The code below is a simplified version of my model_selection() function:

def model_selection(model, param_distributions, n_iter = 100):
    """
    This function performs model selection using random grid search *without cross validation*.
    Inputs:
        model: enter model such as RandomForestClassifier() (which is default)
        param_distributions: pre-defined grid to search over, specific to the input 'model'
        n_iter: Number of parameter settings that are sampled. n_iter trades off runtime vs quality of the solution.
    """

    # dictionary to hold optimal models for each sliding window
    optimal_models = {}

    # 'sets' is a dictionary containing sliding window dataframes e.g. 'X_train_0', 'y_train_0', 'X_test_0', 'y_test_0', 'X_train_1', 'y_train_1', 'X_test_1', 'y_test_1'

    for i in np.arange(0, len(sets), 4): # for each sliding window

            # assign the train and validation sets for the given sliding window
            X_train = list(sets_for_model_selection.values())[i] # THESE HAVE DIFFERENT DIMS FROM WINDOW TO WINDOW
            X_val = list(sets_for_model_selection.values())[i+1] # THESE HAVE DIFFERENT DIMS FROM WINDOW TO WINDOW
            y_train = list(sets_for_model_selection.values())[i+2]
            y_val = list(sets_for_model_selection.values())[i+3]

            # set up the grid search
            mdl_opt = RandomizedSearchCV(estimator = model, param_distributions = param_distributions, 
                                 n_iter = n_iter, cv = ps, verbose=2)

            # Fit the random search model: parameter combinations will be trained, then tested on the validation set
            mdl_opt.fit(np.concatenate((X_train, X_val), axis = 0), 
                np.concatenate((y_train.values.ravel(), y_val.values.ravel()), axis = 0))

            mdl = {'optimal_model_sw'+str(i) : mdl_opt.best_estimator_}

            # update the 'optimal models' dictionary
            optimal_models.update(mdl)

return optimal_models

Answer 1

该解决方案涉及对 KerasClassifier 包装器进行轻微编辑并编辑我的函数model_selection()。

首先我将其更改input_dim为“无”：

def create_model(optimizer='adam', kernel_initializer='normal', dropout_rate=0.0, input_dim=None):
    with tf.device("/device:GPU:0"):
        # create model
        model = Sequential()
        model.add(Dense(20, input_dim=None, activation='relu', kernel_initializer=kernel_initializer))
        model.add(Dropout(dropout_rate))
        model.add(Dense(20, activation='relu'))
        model.add(Dense(1, activation='sigmoid'))
        # Compile model
        model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])
        return model

然后在模型选择函数中，我添加了一个额外的参数“mlp”来断言所讨论的模型是否是神经网络。

如果为 True，则KerasClassifier模型是在model_selection()函数内创建的，模型可以访问相关滑动窗口的维数。这些用作构造函数input_dim中 keyarg的输入KerasClassifier（如链接 Vivek Kumar 指出的那样）：

def model_selection(model, param_distributions, n_iter = 100, mlp=None):
    """
    This function performs model selection using random grid search *without cross validation*.
    Inputs:
        model: enter model such as RandomForestClassifier() (which is default)
        param_distributions: pre-defined grid to search over, specific to the input 'model'
        n_iter: Number of parameter settings that are sampled. n_iter trades off runtime vs quality of the solution.
    """
    # dictionary to hold optimal models for each sliding window
    optimal_models = {}

# 'sets' is a dictionary containing sliding window dataframes e.g. 'X_train_0', 'y_train_0', 'X_test_0', 'y_test_0', 'X_train_1', 'y_train_1', 'X_test_1', 'y_test_1'

    for i in np.arange(0, len(sets), 4): # for each sliding window

        # assign the train and validation sets for the given sliding window
        X_train = list(sets_for_model_selection.values())[i] # THESE HAVE DIFFERENT DIMS FROM WINDOW TO WINDOW
        X_val = list(sets_for_model_selection.values())[i+1] # THESE HAVE DIFFERENT DIMS FROM WINDOW TO WINDOW
        y_train = list(sets_for_model_selection.values())[i+2]
        y_val = list(sets_for_model_selection.values())[i+3]

        if mlp:
            input_dims = list(sets_for_model_selection.values())[i].shape[1]
            model =  KerasClassifier(build_fn=create_model, input_dim=input_dims, verbose=0)

        # set up the grid search
        mdl_opt = RandomizedSearchCV(estimator = model, param_distributions = param_distributions, 
                             n_iter = n_iter, cv = ps, verbose=2)

        # Fit the random search model: parameter combinations will be trained, then tested on the validation set
        mdl_opt.fit(np.concatenate((X_train, X_val), axis = 0), 
            np.concatenate((y_train.values.ravel(), y_val.values.ravel()), axis = 0))

        mdl = {'optimal_model_sw'+str(i) : mdl_opt.best_estimator_}

        # update the 'optimal models' dictionary
        optimal_models.update(mdl)
    return optimal_models