自动编码器输出没有正确的形状

Question

我创建了一个连续向autoencoder添加图层的类，目的是从数据中学习越来越高的特征表示。然而，在训练我的算法并使用预测函数后，输出形状不是正确的形状。

x: shape = (6000, 23)

alg = SuccessiveAutoencoder()

alg.fit(x, epochs = 10)

x_compressed = alg.predict(x)

x_compressed should have a shape (6000,5) but instead has (6000, 23)

有人能帮我一下吗？问题出在哪里？

请找到下面的代码：

class SuccessiveAutoencoder():
"""The algorithm stacks autoencoders one after others, with the idea to
learn from the data higher level features.

Attributes
-----------------   
- hidden_layers: number of hidden layers
- units_nbr: number of neurons in each hidden layer
- activation: type of activation to be used. Note that the activation of 
the decoder is hardcoded within the class, and is a sigmoid
- fine_tuning: if the algorithm train trains the full network once this one
has been built. It can be 'y' or 'n'
"""

def __init__(self, hidden_layers = 5, units_nbr = 5, activation = 'relu',
             fine_tuning = 'y'):
    """Initialize the algorithm"""

    self.hidden_layers = hidden_layers
    self.units_nbr = units_nbr
    self.activation = activation
    self.fine_tuning = fine_tuning

def create_model(self, X):
    """Create the inital autoencoder structure"""

    # Build simple autoencoder
    input_time_series = Input(shape = (X.shape[1], ), name = 'input')
    layer1 = Dense(units = self.units_nbr, activation = self.activation, 
                   name = 'hidden1')(input_time_series)
    decoded = Dense(units = X.shape[1], activation = 'sigmoid', name = 'output')(layer1)

    return Model(input_time_series, decoded)

def add_layer(self, model, incr):
    """Add a new layer, and make previous layers not trainable"""

    # Set previous layers not trainable
    for layer in model.layers[:-1]:
        layer.trainable = False

    # Output of previous layer
    out = model.layers[-2].output

    # Add the new layer
    layer_new = Dense(units = self.units_nbr, activation = self.activation,
                      name = 'hidden' + str(incr))(out)

    decoded = model.layers[-1](layer_new)

    return Model(model.layers[0].input, decoded)

def fit(self, X, epochs):
    """Run the algorithm, creating the autoencoder and calibrating it"""

    # Create the model and train it
    print('/ Training Hidden Layer 1')
    model = self.create_model(X)
    model.compile(loss = 'mean_squared_error', optimizer = 'adam')

    h = model.fit(X, X, epochs = epochs, verbose = 0)
    print('Last loss: {}'.format(h.history['loss'][-1]))

    # Incrementally add layer, and train these new layers
    for incr in range(2, self.hidden_layers + 1):
        print('/ Training Hidden Layer {}'.format(str(incr)))
        model = self.add_layer(model, incr)
        model.compile(loss = 'mean_squared_error', optimizer = 'adam')

        h = model.fit(X, X, epochs = epochs, verbose = 0)
        print('Last loss: {}'.format(h.history['loss'][-1]))

    # If the user wants to run the calibration again over the complete model
    if self.fine_tuning == 'y':    

        # Final training
        print('/ Final Tuning')
        for layer in model.layers:
            layer.trainable = True

        model.compile(loss = 'mean_squared_error', optimizer = 'adam')

        h = model.fit(X, X, epochs = epochs, verbose = 0)
        print('Last loss: {}'.format(h.history['loss'][-1]))

    # Get rid of last layer, and stored the model
    model.layers.pop()

    model.compile(loss = 'mean_squared_error', optimizer = 'adam')

    self.model = model

def predict(self, X):
    """Predict the compressed information from X"""

    return self.model.predict(X)

Answer 1

代码的第103行，而不是使用

self.model = model

我把

self.model = Model(model.layers[0].input, model.layers[-1].output)

看起来它起作用了。但任何见解都是受欢迎的。