为什么在20个全连接层后再dropout，而不是每个全连接层后来一个dropout

Question

为什么在20个全连接层后再dropout，而不是每个全连接层后来一个dropout,就像BatchNormalization那样呢？

---

Layer (type) Output Shape Param #

flatten (Flatten) (None, 784) 0

dense (Dense) (None, 100) 78500

dense_1 (Dense) (None, 100) 10100

dense_2 (Dense) (None, 100) 10100

dense_3 (Dense) (None, 100) 10100

dense_4 (Dense) (None, 100) 10100

dense_5 (Dense) (None, 100) 10100

dense_6 (Dense) (None, 100) 10100

dense_7 (Dense) (None, 100) 10100

dense_8 (Dense) (None, 100) 10100

dense_9 (Dense) (None, 100) 10100

dense_10 (Dense) (None, 100) 10100

dense_11 (Dense) (None, 100) 10100

dense_12 (Dense) (None, 100) 10100

dense_13 (Dense) (None, 100) 10100

dense_14 (Dense) (None, 100) 10100

dense_15 (Dense) (None, 100) 10100

dense_16 (Dense) (None, 100) 10100

dense_17 (Dense) (None, 100) 10100

dense_18 (Dense) (None, 100) 10100

dense_19 (Dense) (None, 100) 10100

alpha_dropout (AlphaDropout (None, 100) 0
)

dense_20 (Dense) (None, 10) 1010

=================================================================
Total params: 271,410
Trainable params: 271,410
Non-trainable params: 0

---

Answer 1

正十七 2022-07-07 21:58:22

可以每层都接dropout，这里只是展示了一种使用的方法。一般来说，如果dropout层较多，有可能会导致模型难训练，所以如果你每层都加dropout，dropout rate不要设的太大。

你可以试试每层都加，看看效果如何。

• 提问者 慕粉2125289011 #1

  非常感谢！