tf.layers.dropout 和 tf.nn.dropout的区别

1. tf.nn.dropout函数

tf.nn.dropout(
    x,
    keep_prob,
    noise_shape=None,
    seed=None,
    name=None
)

x：指输入，输入tensor

keep_prob: float类型，每个元素被保留下来的概率，设置神经元被选中的概率,在初始化时keep_prob是一个占位符, keep_prob = tf.placeholder(tf.float32) 。tensorflow在run时设置keep_prob具体的值，例如keep_prob: 0.5

noise_shape  : 一个1维的int32张量，代表了随机产生“保留/丢弃”标志的shape。

seed : 整形变量，随机数种子。

name：指定该操作的名字

 

2. tf.layers.dropout 函数

tf.layers.dropout(
    inputs,
    rate=0.5,
    noise_shape=None,
    seed=None,
    training=False,
    name=None
)

inputs: Tensor input.
rate: The dropout rate, between 0 and 1. E.g. "rate=0.1" would drop out 10% of input units.
noise_shape: 1D tensor of type int32 representing the shape of the binary dropout mask that will be multiplied with the input. For instance, if your inputs have shape (batch_size, timesteps, features), and you want the dropout mask to be the same for all timesteps, you can use noise_shape=[batch_size, 1, features].
seed: A Python integer. Used to create random seeds. Seetf.set_random_seed for behavior.
training: Either a Python boolean, or a TensorFlow boolean scalar tensor (e.g. a placeholder). Whether to return the output in training mode (apply dropout) or in inference mode (return the input untouched).
name: The name of the layer (string).

 

3. 两者的区别

两者都是用于dropout，但是其参数和使用有些许差别。

（1）tf.nn.dropout的参数是keep_prob,表示元素被保留下来的概率，预测阶段应设置为1；

（2）tf.layers.dropout的参数是rate，表示元素被丢弃的概率，预测阶段应设置为0.

  (3) tf.layers.dropout在预测阶段不生效，应该是默认trainging=False，因此这时传参数不生效。