Class: DNN::Layers::BatchNormalization

Inherits:
HasParamLayer show all
Defined in:
lib/dnn/core/normalizations.rb

Instance Attribute Summary collapse

Attributes inherited from HasParamLayer

#params, #trainable

Attributes inherited from Layer

#input_shape, #learning_phase

Class Method Summary collapse

Instance Method Summary collapse

Methods inherited from Layer

#built?, #output_shape

Constructor Details

#initialize(axis: 0, momentum: 0.9, eps: 1e-7) ⇒ BatchNormalization

Returns a new instance of BatchNormalization.

Parameters:

  • axis (integer) (defaults to: 0) —

    The axis to normalization.

  • momentum (Float) (defaults to: 0.9) —

    Exponential moving average of mean and variance.

  • eps (Float) (defaults to: 1e-7) —

    Value to avoid division by zero.



19
20
21
22
23
24
 
# File 'lib/dnn/core/normalizations.rb', line 19

def initialize(axis: 0, momentum: 0.9, eps: 1e-7)
  super()
  @axis = axis
  @momentum = momentum
  @eps = eps
end

Instance Attribute Details

#axis ⇒ Integer (readonly)

Returns The axis to normalization.

Returns:

  • (Integer) —

    The axis to normalization.



6
7
8
 
# File 'lib/dnn/core/normalizations.rb', line 6

def axis
  @axis
end

#eps ⇒ Float

Returns Value to avoid division by zero.

Returns:

  • (Float) —

    Value to avoid division by zero.



10
11
12
 
# File 'lib/dnn/core/normalizations.rb', line 10

def eps
  @eps
end

#momentum ⇒ Float

Returns Exponential moving average of mean and variance.

Returns:

  • (Float) —

    Exponential moving average of mean and variance.



8
9
10
 
# File 'lib/dnn/core/normalizations.rb', line 8

def momentum
  @momentum
end

Class Method Details

.from_hash(hash) ⇒ Object 



12
13
14
 
# File 'lib/dnn/core/normalizations.rb', line 12

def self.from_hash(hash)
  self.new(axis: hash[:axis], momentum: hash[:momentum])
end

Instance Method Details

#backward(dy) ⇒ Object 



51
52
53
54
55
56
57
58
59
60
61
62
63
64
 
# File 'lib/dnn/core/normalizations.rb', line 51

def backward(dy)
  batch_size = dy.shape[@axis]
  if @trainable
    @beta.grad = dy.sum(axis: @axis, keepdims: true)
    @gamma.grad = (@xn * dy).sum(axis: @axis, keepdims: true)
  end
  dxn = @gamma.data * dy
  dxc = dxn / @std
  dstd = -((dxn * @xc) / (@std**2)).sum(axis: @axis, keepdims: true)
  dvar = 0.5 * dstd / @std
  dxc += (2.0 / batch_size) * @xc * dvar
  dmean = dxc.sum(axis: @axis, keepdims: true)
  dxc - dmean / batch_size
end

#build(input_shape) ⇒ Object 



26
27
28
29
30
31
32
 
# File 'lib/dnn/core/normalizations.rb', line 26

def build(input_shape)
  super
  @params[:gamma] = @gamma = Param.new(Xumo::SFloat.ones(*output_shape), 0)
  @params[:beta] = @beta = Param.new(Xumo::SFloat.zeros(*output_shape), 0)
  @params[:running_mean] = @running_mean = Param.new(Xumo::SFloat.zeros(*output_shape))
  @params[:running_var] = @running_var = Param.new(Xumo::SFloat.zeros(*output_shape))
end

#forward(x) ⇒ Object 



34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
 
# File 'lib/dnn/core/normalizations.rb', line 34

def forward(x)
  if learning_phase
    mean = x.mean(axis: @axis, keepdims: true)
    @xc = x - mean
    var = (@xc**2).mean(axis: @axis, keepdims: true)
    @std = NMath.sqrt(var + @eps)
    xn = @xc / @std
    @xn = xn
    @running_mean.data = @momentum * @running_mean.data + (1 - @momentum) * mean
    @running_var.data = @momentum * @running_var.data + (1 - @momentum) * var
  else
    xc = x - @running_mean.data
    xn = xc / NMath.sqrt(@running_var.data + @eps)
  end
  @gamma.data * xn + @beta.data
end

#to_hash ⇒ Object 



66
67
68
 
# File 'lib/dnn/core/normalizations.rb', line 66

def to_hash
  super({axis: @axis, momentum: @momentum, eps: @eps})
end