(原)模型的参数初始化

转载请注明出处:

http://www.cnblogs.com/darkknightzh/p/8297793.html

参考网址:

http://pytorch.org/docs/0.3.0/nn.html?highlight=kaiming#torch.nn.init.kaiming_normal

https://github.com/prlz77/ResNeXt.pytorch/blob/master/models/model.py

https://github.com/facebookresearch/ResNeXt/blob/master/models/resnext.lua

https://github.com/bamos/densenet.pytorch/blob/master/densenet.py

https://github.com/szagoruyko/wide-residual-networks/blob/master/models/utils.lua

说明:暂时就这么多吧,错误之处请见谅。前两个初始化的方法见pytorch官方文档

1. xavier初始化

torch.nn.init.xavier_uniform(tensor, gain=1)

对于输入的tensor或者变量,通过论文Understanding the difficulty of training deep feedforward neural networks” - Glorot, X. & Bengio, Y. (2010)的方法初始化数据。初始化服从均匀分布$U(-a,a)$,其中$a=gain\times \sqrt{2/(fan\_in+fan\_out)}\times \sqrt{3}$,该初始化方法也称Glorot initialisation。

参数:

      tensor:n维的 torch.Tensor 或者 autograd.Variable类型的数据

      a:可选择的缩放参数

例如:

w = torch.Tensor(3, 5)
nn.init.xavier_uniform(w, gain=nn.init.calculate_gain('relu'))

torch.nn.init.xavier_normal(tensor, gain=1)

对于输入的tensor或者变量,通过论文Understanding the difficulty of training deep feedforward neural networks” - Glorot, X. & Bengio, Y. (2010)的方法初始化数据。初始化服从高斯分布$N(0,std)$,其中$std=gain\times \sqrt{2/(fan\_in+fan\_out)}$,该初始化方法也称Glorot initialisation。

参数:

      tensor:n维的 torch.Tensor 或者 autograd.Variable类型的数据

      a:可选择的缩放参数

例如:

w = torch.Tensor(3, 5)
nn.init.xavier_normal(w)

2. kaiming初始化

torch.nn.init.kaiming_uniform(tensor, a=0, mode='fan_in')

对于输入的tensor或者变量,通过论文“Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification” - He, K. et al. (2015)的方法初始化数据。初始化服从均匀分布$U(-bound,bound)$,其中$bound=\sqrt{2/((1+{{a}^{2}})\times fan\_in)}\times \sqrt{3}$,该初始化方法也称He initialisation。

参数:

      tensor:n维的 torch.Tensor 或者 autograd.Variable类型的数据

      a:该层后面一层的激活函数中负的斜率(默认为ReLU,此时a=0)

      mode:‘fan_in’ (default) 或者 ‘fan_out’. 使用fan_in保持weights的方差在前向传播中不变;使用fan_out保持weights的方差在反向传播中不变。

例如:

w = torch.Tensor(3, 5)
nn.init.kaiming_uniform(w, mode='fan_in')

torch.nn.init.kaiming_normal(tensor, a=0, mode='fan_in')

对于输入的tensor或者变量,通过论文“Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification” - He, K. et al. (2015)的方法初始化数据。初始化服从高斯分布$N(0,std)$,其中$std=\sqrt{2/((1+{{a}^{2}})\times fan\_in)}$,该初始化方法也称He initialisation。

参数:

      tensor:n维的 torch.Tensor 或者 autograd.Variable类型的数据

      a:该层后面一层的激活函数中负的斜率(默认为ReLU,此时a=0)

      mode:‘fan_in’ (default) 或者 ‘fan_out’. 使用fan_in保持weights的方差在前向传播中不变;使用fan_out保持weights的方差在反向传播中不变。

例如:

w = torch.Tensor(3, 5)
nn.init.kaiming_normal(w, mode='fan_out')

使用的例子(具体参见原始网址):

https://github.com/prlz77/ResNeXt.pytorch/blob/master/models/model.py

from torch.nn import init
self.classifier = nn.Linear(self.stages[3], nlabels)
init.kaiming_normal(self.classifier.weight)
for key in self.state_dict():
    if key.split('.')[-1] == 'weight':
        if 'conv' in key:
            init.kaiming_normal(self.state_dict()[key], mode='fan_out')
        if 'bn' in key:
            self.state_dict()[key][...] = 1
    elif key.split('.')[-1] == 'bias':
        self.state_dict()[key][...] = 0

3. 实际使用中看到的初始化

3.1 ResNeXt,densenet中初始化

https://github.com/facebookresearch/ResNeXt/blob/master/models/resnext.lua

https://github.com/bamos/densenet.pytorch/blob/master/densenet.py

conv

n = kW* kH*nOutputPlane
weight:normal(0,math.sqrt(2/n))
bias:zero()

batchnorm

weight:fill(1)
bias:zero()

linear

bias:zero()

3.2 wide-residual-networks中初始化(MSRinit

https://github.com/szagoruyko/wide-residual-networks/blob/master/models/utils.lua

conv

n = kW* kH*nInputPlane
weight:normal(0,math.sqrt(2/n))
bias:zero()

linear

bias:zero()

 

posted on 2018-01-16 17:33  darkknightzh  阅读(10513)  评论(0编辑  收藏  举报

导航