Question

我有一个Keras模型，尺寸如下：

________________________________________________________________________________
Layer (type)              Output Shape              Param #
================================================================================
stft (InputLayer)         (None, 1, 16384)          0
________________________________________________________________________________
static_stft (Spectrogram) (None, 1, 65, 256)        16640
________________________________________________________________________________
conv2d_1 (Conv2D)         (None, 38, 5, 9)          12882
________________________________________________________________________________
dense_1 (Dense)           (None, 38, 5, 512)        5120
________________________________________________________________________________
predictions (Dense)       (None, 38, 5, 368)        188784
================================================================================

我对最后的密集层的尺寸感到困惑。我希望分别拥有（None，512）和（None，368）。答案如下：https://blog.thoughtram.io/angular/2015/06/29/shadow-dom-strategies-in-angular2.html

最终的致密层创建如下：

x = keras.layers.Dense(512)(x)
outputs = keras.layers.Dense(
        368, activation='sigmoid', name='predictions')(x)

那么为什么它们有超过512个输出？我该如何更改呢？

Answer 1

根据您的应用程序，您可以在Conv2D层之后展平：

input_layer = Input((1, 1710))
x = Reshape((38, 5, 9))(input_layer)
x = Flatten()(x)
x = Dense(512)(x)
x = Dense(368)(x)

Layer (type)                 Output Shape              Param #   
_________________________________________________________________
input_1 (InputLayer)         [(None, 1, 1710)]         0         
_________________________________________________________________
reshape (Reshape)            (None, 38, 5, 9)          0         
_________________________________________________________________
flatten (Flatten)            (None, 1710)              0         
_________________________________________________________________
dense (Dense)                (None, 512)               876032    
_________________________________________________________________
dense_1 (Dense)              (None, 368)               188784

Answer 2

这是Conv2D层。卷积层产生长度为9的38x5输出，然后您的Dense层将38x5长度为9的每个序列作为输入，并将其转换为长度为512的序列作为输出。

要摆脱空间依赖性，您需要使用诸如池层之类的东西，可能是GlobalMaxPool2D。这样会将数据仅合并到通道维中，并生成(None, 9)形状的输出，这将根据Dense层得出您期望的形状。

Keras致密层的尺寸

2 个答案: