由于目标标签的形状，Keras 不接受目标标签？

Question

我正在尝试让 Keras 处理分类问题，该问题有五个分类目标标签（1、2、3、4、5）。出于某种原因，我在使用 StratifiedKFold 时无法让它工作。 X 和 y 是形状分别为 (500, 20) 和 (500, ) 的 NumPy 数组。

错误信息是"ValueError: Error when checking target: expected dense_35 to have shape (1,) but got array with shape (5,)"，这让我认为错误肯定出在目标变量的格式上。同样值得注意的是，"dense_35" 中的数字似乎随着每次尝试运行代码的尝试而变化。

random_state = 123
n_splits = 10
cv = StratifiedKFold(n_splits=n_splits, 
random_state=random_state, shuffle=False)

def baseline_model():
    nn_model = Sequential()
    nn_model.add(Dense(units=50, input_dim=X.shape[1], init='normal',
                       activation= 'relu' ))
    nn_model.add(Dense(30, init='normal', activation='relu'))
    nn_model.add(Dense(10, init='normal', activation='relu'))
    nn_model.add(Dense(1, init='normal', activation='softmax'))

    nn_model.compile(optimizer='adam', loss='categorical_crossentropy',
                     metrics = ['accuracy'])
    return nn_model

for train, test in cv.split(X, y):   
    X_train, X_test = X[train], X[test]
    y_train, y_test = y[train], y[test]

    np_utils.to_categorical(y_train)
    np_utils.to_categorical(y_test)

    estimator = KerasClassifier(build_fn=baseline_model,
                                epochs=200, batch_size=5,
                                verbose=0)

    estimator.fit(X_train, y_train)
    y_pred = estimator.predict(X_test)

The numpy array (y), that I am trying to split:
[1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2
 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4
 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4
 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 5 5 5 5 5 5 5
 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5
 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5
 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5]

Answer 1

random_state = 123
n_splits = 10
cv = StratifiedKFold(n_splits=n_splits, random_state=random_state, shuffle=False)
def baseline_model():
    nn_model = Sequential(name='model_name')
    nn_model.add(Dense(units=50, input_dim=X.shape[1], init='normal',
                       activation= 'relu', name='dense1'))
    nn_model.add(Dense(30, init='normal', activation='relu', name='dense2'))
    nn_model.add(Dense(10, init='normal', activation='relu', name='dense3'))
    # code changed here
    nn_model.add(Dense(5, init='normal', activation='softmax', name='dense4'))

    nn_model.compile(optimizer='adam', loss='categorical_crossentropy',
                     metrics = ['accuracy'])
    return nn_model

for train, test in cv.split(X, y):   
    X_train, X_test = X[train], X[test]
    y_train, y_test = y[train], y[test]

    # the error is due to this step
    # you have specified only one output in the last dense layer (dense4)
    # but you are giving input of length 5
    np_utils.to_categorical(y_train)
    np_utils.to_categorical(y_test)

    estimator = KerasClassifier(build_fn=baseline_model,
                                epochs=200, batch_size=5,
                                verbose=0)
    estimator.fit(X_train, y_train)
    y_pred = estimator.predict(X_test)

通过在层中指定 name 参数，您可以命名层。这样你每次都会得到明确的图层名称，以防出错。
model.summary() 是另一个有用的函数，您可以使用它检查每一层的输出形状。

Answer 2

由于是分类问题，有五个分类目标标签，所以最后一个密集层（输出层）必须有5个单元：

def baseline_model():
    nn_model = Sequential()
    nn_model.add(Dense(units=50, input_dim=X.shape[1], init='normal',
                       activation= 'relu' ))
    nn_model.add(Dense(30, init='normal', activation='relu'))
    nn_model.add(Dense(10, init='normal', activation='relu'))

    #Output layer
    nn_model.add(Dense(5, init='normal', activation='softmax'))
    nn_model.compile(optimizer='adam', loss='categorical_crossentropy',
                     metrics = ['accuracy'])
    return nn_model

由于目标标签的形状，Keras 不接受目标标签？

Keras does not accept the target label due to its shape?

python

classification

scikit-learn

cross-validation

keras