python - 如何在 Keras 中展平不同大小的数据并在下一层中使用它

Question

我想构建一个 CNN，它可以是行数不变但列数可变的馈送矩阵。我从this blog开始但它在某些时候使用了 Flatten 层。根据this GitHub issue , Flatten 层不能与未完全定义形状的数据一起使用。我用了this answer像这样构建一个 CNN:

from keras import Input, Model
from keras.layers import Conv2D, MaxPooling2D, Dense, Dropout, Lambda, Activation
from keras.optimizers import SGD
import keras.backend as K

input = Input(shape=(None, 60, 1))
        x = Conv2D(list_of_neurons[0], (3, 3), padding='same', activation='relu')(input)
        x = MaxPooling2D(pool_size=(2, 2))(x) # None x 30

        x = Conv2D(list_of_neurons[1], (3, 3), padding='same', activation='relu')(x)
        x = MaxPooling2D(pool_size=(2, 2))(x) # None x 15

        x = Conv2D(list_of_neurons[2], (3, 3), padding='same', activation='relu')(x)
        x = MaxPooling2D(pool_size=(3, 3))(x) # None x 5

        x = Lambda(lambda k: K.batch_flatten(k))(x)
        x = Dense(list_of_neurons[3])(x)
        x = Activation("relu")(x)
        x = Dropout(dropout)(x)
        output = Dense(1, activation='sigmoid')(x)

        self._model = Model(inputs=input, outputs=output)
        self._model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
        self._model.summary()

但是，当我尝试运行脚本时出现错误:

Traceback (most recent call last):
  File "/home/kuba/Dropbox/MGR Jakub Kustra/implementacja/first_cnn.py", line 66, in <module>
    mlp = MLP([60, 60, 120, 120], 0.3)
  File "/home/kuba/Dropbox/MGR Jakub Kustra/implementacja/models.py", line 22, in __init__
    x = Dense(list_of_neurons[3])(x)
  File "/home/kuba/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 576, in __call__
    self.build(input_shapes[0])
  File "/home/kuba/anaconda3/lib/python3.6/site-packages/keras/layers/core.py", line 830, in build
    constraint=self.kernel_constraint)
  File "/home/kuba/anaconda3/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 87, in wrapper
    return func(*args, **kwargs)
  File "/home/kuba/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 397, in add_weight
    weight = K.variable(initializer(shape),
  File "/home/kuba/anaconda3/lib/python3.6/site-packages/keras/initializers.py", line 204, in __call__
    scale /= max(1., float(fan_in + fan_out) / 2)
TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'

错误发生在initializers.py中的这一行:

def __call__(self, shape, dtype=None):
    fan_in, fan_out = _compute_fans(shape)
    scale = self.scale
    if self.mode == 'fan_in':
        scale /= max(1., fan_in)
    elif self.mode == 'fan_out':
        scale /= max(1., fan_out)
    else:
        scale /= max(1., float(fan_in + fan_out) / 2) # <<<<<<<< HERE
    if self.distribution == 'normal':
        stddev = np.sqrt(scale)
        return K.truncated_normal(shape, 0., stddev,
                                  dtype=dtype, seed=self.seed)
    else:
        limit = np.sqrt(3. * scale)
        return K.random_uniform(shape, -limit, limit,
                                dtype=dtype, seed=self.seed)

fan_in 变量是None。我认为 Dense 层可以提供不同大小的数据。我该如何克服这个问题？

Answer 1

不幸的是，由于一个简单的原因，您的方案无法实现:您的Flatten 后面跟着一个Dense 层。为了为 Dense 层分配张量，为什么现在应该输入形状 - 因为分配器应该分配形状为 (input_shape, output_shape) 的权重矩阵。这是您的错误来源。

您提供的问题中提出的解决方案解决了一些不同的任务。那里的问题与这样一个事实有关 - 有时需要指定批处理的形状，这就是 batch_flatten 旨在解决的问题。