神经网络权值初始化方法-Xavier
https://blog.csdn.net/u011534057/article/details/51673458
https://blog.csdn.net/qq_34784753/article/details/78668884
https://blog.csdn.net/kangroger/article/details/61414426
https://www.cnblogs.com/lindaxin/p/8027283.html
神经网络中权值初始化的方法
《Understanding the difficulty of training deep feedforward neural networks》
可惜直到近两年,这个方法才逐渐得到更多人的应用和认可。
为了使得网络中信息更好的流动,每一层输出的方差应该尽量相等。
基于这个目标,现在我们就去推导一下:每一层的权重应该满足哪种条件。
文章先假设的是线性激活函数,而且满足0点处导数为1,即
现在我们先来分析一层卷积:
其中ni表示输入个数。
根据概率统计知识我们有下面的方差公式:
特别的,当我们假设输入和权重都是0均值时(目前有了BN之后,这一点也较容易满足),上式可以简化为:
进一步假设输入x和权重w独立同分布,则有:
于是,为了保证输入与输出方差一致,则应该有:
对于一个多层的网络,某一层的方差可以用累积的形式表达:
特别的,反向传播计算梯度时同样具有类似的形式:
综上,为了保证前向传播和反向传播时每一层的方差一致,应满足:
但是,实际当中输入与输出的个数往往不相等,于是为了均衡考量,最终我们的权重方差应满足:
———————————————————————————————————————
———————————————————————————————————————
学过概率统计的都知道 [a,b] 间的均匀分布的方差为:
因此,Xavier初始化的实现就是下面的均匀分布:
——————————————————————————————————————————
caffe的Xavier实现有三种选择
(1) 默认情况,方差只考虑输入个数:
(2) FillerParameter_VarianceNorm_FAN_OUT,方差只考虑输出个数:
(3) FillerParameter_VarianceNorm_AVERAGE,方差同时考虑输入和输出个数:
之所以默认只考虑输入,我个人觉得是因为前向信息的传播更重要一些
———————————————————————————————————————————
Tensorflow 调用接口
https://www.tensorflow.org/api_docs/python/tf/glorot_uniform_initializer
tf.glorot_uniform_initializer
Aliases:
tf.glorot_uniform_initializer
tf.keras.initializers.glorot_uniform
tf.glorot_uniform_initializer(
seed=None,
dtype=tf.float32
)
Defined in tensorflow/python/ops/init_ops.py
.
The Glorot uniform initializer, also called Xavier uniform initializer.
It draws samples from a uniform distribution within [-limit, limit] where limit
is sqrt(6 / (fan_in + fan_out))
where fan_in
is the number of input units in the weight tensor and fan_out
is the number of output units in the weight tensor.
Reference: http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf
Args:
seed
: A Python integer. Used to create random seeds. Seetf.set_random_seed
for behavior.dtype
: The data type. Only floating point types are supported.
Returns:
An initializer.
Mxnet 调用接口
https://mxnet.apache.org/api/python/optimization/optimization.html#mxnet.initializer.Xavier
class mxnet.initializer.
Xavier
(rnd_type='uniform', factor_type='avg', magnitude=3)[source]
Returns an initializer performing “Xavier” initialization for weights.
This initializer is designed to keep the scale of gradients roughly the same in all layers.
By default, rnd_type is 'uniform'
and factor_type is 'avg'
, the initializer fills the weights with random numbers in the range of [−c,c][−c,c], where c=3.0.5∗(nin+nout)−−−−−−−−−√c=3.0.5∗(nin+nout). ninnin is the number of neurons feeding into weights, and noutnout is the number of neurons the result is fed to.
If rnd_type is 'uniform'
and factor_type is 'in'
, the c=3.nin−−−√c=3.nin. Similarly when factor_type is 'out'
, the c=3.nout−−−√c=3.nout.
If rnd_type is 'gaussian'
and factor_type is 'avg'
, the initializer fills the weights with numbers from normal distribution with a standard deviation of 3.0.5∗(nin+nout)−−−−−−−−−√3.0.5∗(nin+nout).
Parameters: |
|
---|