ruby-dnn 0.10.1 → 0.10.2
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/README.md +2 -2
- data/examples/cifar100_example.rb +71 -71
- data/examples/cifar10_example.rb +68 -68
- data/examples/iris_example.rb +34 -34
- data/examples/mnist_conv2d_example.rb +50 -50
- data/examples/mnist_example.rb +39 -39
- data/examples/mnist_lstm_example.rb +36 -36
- data/examples/xor_example.rb +24 -24
- data/lib/dnn.rb +27 -26
- data/lib/dnn/cifar10.rb +51 -51
- data/lib/dnn/cifar100.rb +49 -49
- data/lib/dnn/core/activations.rb +148 -148
- data/lib/dnn/core/cnn_layers.rb +464 -464
- data/lib/dnn/core/dataset.rb +34 -34
- data/lib/dnn/core/embedding.rb +56 -0
- data/lib/dnn/core/error.rb +5 -5
- data/lib/dnn/core/initializers.rb +126 -126
- data/lib/dnn/core/layers.rb +307 -307
- data/lib/dnn/core/losses.rb +175 -175
- data/lib/dnn/core/model.rb +461 -461
- data/lib/dnn/core/normalizations.rb +72 -72
- data/lib/dnn/core/optimizers.rb +283 -283
- data/lib/dnn/core/param.rb +9 -9
- data/lib/dnn/core/regularizers.rb +106 -106
- data/lib/dnn/core/rnn_layers.rb +464 -464
- data/lib/dnn/core/utils.rb +34 -34
- data/lib/dnn/downloader.rb +50 -50
- data/lib/dnn/image.rb +41 -41
- data/lib/dnn/iris.rb +60 -60
- data/lib/dnn/mnist.rb +84 -84
- data/lib/dnn/version.rb +3 -3
- metadata +2 -1
data/lib/dnn/core/dataset.rb
CHANGED
@@ -1,34 +1,34 @@
|
|
1
|
-
# This class manages input datas and output datas together.
|
2
|
-
class DNN::Dataset
|
3
|
-
# @param [Numo::SFloat] x_datas input datas.
|
4
|
-
# @param [Numo::SFloat] y_datas output datas.
|
5
|
-
# @param [Bool] random Set true to return batches randomly. Setting false returns batches in order of index.
|
6
|
-
def initialize(x_datas, y_datas, random = true)
|
7
|
-
@x_datas = x_datas
|
8
|
-
@y_datas = y_datas
|
9
|
-
@random = random
|
10
|
-
@num_datas = x_datas.shape[0]
|
11
|
-
reset_indexs
|
12
|
-
end
|
13
|
-
|
14
|
-
# Return the next batch.
|
15
|
-
# If the number of remaining data < batch size, if random = true, shuffle the data again and return a batch.
|
16
|
-
# If random = false, all remaining data will be returned regardless of the batch size.
|
17
|
-
def next_batch(batch_size)
|
18
|
-
if @indexes.length < batch_size
|
19
|
-
batch_indexes = @indexes unless @random
|
20
|
-
reset_indexs
|
21
|
-
batch_indexes = @indexes.shift(batch_size) if @random
|
22
|
-
else
|
23
|
-
batch_indexes = @indexes.shift(batch_size)
|
24
|
-
end
|
25
|
-
x_batch = @x_datas[batch_indexes, false]
|
26
|
-
y_batch = @y_datas[batch_indexes, false]
|
27
|
-
[x_batch, y_batch]
|
28
|
-
end
|
29
|
-
|
30
|
-
private def reset_indexs
|
31
|
-
@indexes = @num_datas.times.to_a
|
32
|
-
@indexes.shuffle! if @random
|
33
|
-
end
|
34
|
-
end
|
1
|
+
# This class manages input datas and output datas together.
|
2
|
+
class DNN::Dataset
|
3
|
+
# @param [Numo::SFloat] x_datas input datas.
|
4
|
+
# @param [Numo::SFloat] y_datas output datas.
|
5
|
+
# @param [Bool] random Set true to return batches randomly. Setting false returns batches in order of index.
|
6
|
+
def initialize(x_datas, y_datas, random = true)
|
7
|
+
@x_datas = x_datas
|
8
|
+
@y_datas = y_datas
|
9
|
+
@random = random
|
10
|
+
@num_datas = x_datas.shape[0]
|
11
|
+
reset_indexs
|
12
|
+
end
|
13
|
+
|
14
|
+
# Return the next batch.
|
15
|
+
# If the number of remaining data < batch size, if random = true, shuffle the data again and return a batch.
|
16
|
+
# If random = false, all remaining data will be returned regardless of the batch size.
|
17
|
+
def next_batch(batch_size)
|
18
|
+
if @indexes.length < batch_size
|
19
|
+
batch_indexes = @indexes unless @random
|
20
|
+
reset_indexs
|
21
|
+
batch_indexes = @indexes.shift(batch_size) if @random
|
22
|
+
else
|
23
|
+
batch_indexes = @indexes.shift(batch_size)
|
24
|
+
end
|
25
|
+
x_batch = @x_datas[batch_indexes, false]
|
26
|
+
y_batch = @y_datas[batch_indexes, false]
|
27
|
+
[x_batch, y_batch]
|
28
|
+
end
|
29
|
+
|
30
|
+
private def reset_indexs
|
31
|
+
@indexes = @num_datas.times.to_a
|
32
|
+
@indexes.shuffle! if @random
|
33
|
+
end
|
34
|
+
end
|
@@ -0,0 +1,56 @@
|
|
1
|
+
module DNN
|
2
|
+
module Layers
|
3
|
+
|
4
|
+
class Embedding < HasParamLayer
|
5
|
+
# @return [Integer] Return the input length.
|
6
|
+
attr_reader :input_length
|
7
|
+
# @return [Initializers::Initializer] Return the weight initializer.
|
8
|
+
attr_reader :weight_initializer
|
9
|
+
|
10
|
+
def self.from_hash(hash)
|
11
|
+
self.new(hash[:input_shape], hash[:input_length],
|
12
|
+
weight_initializer: DNN::Utils.from_hash(hash[:weight_initializer]))
|
13
|
+
end
|
14
|
+
|
15
|
+
# @param [Integer | Array] input_dim_or_shape Set input data dimension or shape.
|
16
|
+
# @param [Integer] input_length input Set the time series length of input data.
|
17
|
+
def initialize(input_dim_or_shape, input_length, weight_initializer: Initializers::RandomUniform.new)
|
18
|
+
super()
|
19
|
+
@input_shape = input_dim_or_shape.is_a?(Array) ? input_dim_or_shape : [input_dim_or_shape]
|
20
|
+
@input_length = input_length
|
21
|
+
@weight_initializer = weight_initializer
|
22
|
+
end
|
23
|
+
|
24
|
+
def build
|
25
|
+
@built = true
|
26
|
+
@params[:weight] = @weight = Param.new(Xumo::SFloat.new(@input_length), 0)
|
27
|
+
@weight_initializer.init_param(self, @weight)
|
28
|
+
@input_shape
|
29
|
+
end
|
30
|
+
|
31
|
+
def forward(x)
|
32
|
+
@x = x
|
33
|
+
y = Xumo::SFloat.zeros(*x.shape)
|
34
|
+
x.shape[0].times do |i|
|
35
|
+
y[i, false] = @weight.data[x[i, false]]
|
36
|
+
end
|
37
|
+
y
|
38
|
+
end
|
39
|
+
|
40
|
+
def backward(dy)
|
41
|
+
@weight.grad += Xumo::SFloat.zeros(*@weight.data.shape)
|
42
|
+
@x.shape[0].times do |i|
|
43
|
+
@x.shape[1].times do |j|
|
44
|
+
@weight.grad[@x[i, j]] += dy[i, j]
|
45
|
+
end
|
46
|
+
end
|
47
|
+
nil
|
48
|
+
end
|
49
|
+
|
50
|
+
def to_hash
|
51
|
+
super(input_shape: @input_shape, input_length: @input_length, weight_initializer: @weight_initializer.to_hash)
|
52
|
+
end
|
53
|
+
end
|
54
|
+
|
55
|
+
end
|
56
|
+
end
|
data/lib/dnn/core/error.rb
CHANGED
@@ -1,5 +1,5 @@
|
|
1
|
-
module DNN
|
2
|
-
class DNN_Error < StandardError; end
|
3
|
-
|
4
|
-
class DNN_ShapeError < DNN_Error; end
|
5
|
-
end
|
1
|
+
module DNN
|
2
|
+
class DNN_Error < StandardError; end
|
3
|
+
|
4
|
+
class DNN_ShapeError < DNN_Error; end
|
5
|
+
end
|
@@ -1,126 +1,126 @@
|
|
1
|
-
module DNN
|
2
|
-
module Initializers
|
3
|
-
|
4
|
-
class Initializer
|
5
|
-
def initialize(seed = false)
|
6
|
-
@seed = seed == true ? rand(1 << 31) : seed
|
7
|
-
end
|
8
|
-
|
9
|
-
def init_param(layer, param)
|
10
|
-
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'init_params'")
|
11
|
-
end
|
12
|
-
|
13
|
-
def to_hash(merge_hash = nil)
|
14
|
-
hash = {class: self.class.name, seed: @seed}
|
15
|
-
hash.merge!(merge_hash) if merge_hash
|
16
|
-
hash
|
17
|
-
end
|
18
|
-
end
|
19
|
-
|
20
|
-
|
21
|
-
class Zeros < Initializer
|
22
|
-
def init_param(layer, param)
|
23
|
-
param.data = param.data.fill(0)
|
24
|
-
end
|
25
|
-
end
|
26
|
-
|
27
|
-
|
28
|
-
class Const < Initializer
|
29
|
-
attr_reader :const
|
30
|
-
|
31
|
-
def self.from_hash(hash)
|
32
|
-
self.new(hash[:const])
|
33
|
-
end
|
34
|
-
|
35
|
-
def initialize(const)
|
36
|
-
super()
|
37
|
-
@const = const
|
38
|
-
end
|
39
|
-
|
40
|
-
def init_param(layer, param)
|
41
|
-
param.data = param.data.fill(@const)
|
42
|
-
end
|
43
|
-
|
44
|
-
def to_hash
|
45
|
-
super({const: @const})
|
46
|
-
end
|
47
|
-
end
|
48
|
-
|
49
|
-
|
50
|
-
class RandomNormal < Initializer
|
51
|
-
attr_reader :mean
|
52
|
-
attr_reader :std
|
53
|
-
|
54
|
-
def self.from_hash(hash)
|
55
|
-
self.new(hash[:mean], hash[:std], hash[:seed])
|
56
|
-
end
|
57
|
-
|
58
|
-
def initialize(mean = 0, std = 0.05, seed = true)
|
59
|
-
super(seed)
|
60
|
-
@mean = mean
|
61
|
-
@std = std
|
62
|
-
end
|
63
|
-
|
64
|
-
def init_param(layer, param)
|
65
|
-
Xumo::SFloat.srand(@seed)
|
66
|
-
param.data = param.data.rand_norm(@mean, @std)
|
67
|
-
end
|
68
|
-
|
69
|
-
def to_hash
|
70
|
-
super({mean: @mean, std: @std})
|
71
|
-
end
|
72
|
-
end
|
73
|
-
|
74
|
-
|
75
|
-
class RandomUniform < Initializer
|
76
|
-
attr_reader :min
|
77
|
-
attr_reader :max
|
78
|
-
|
79
|
-
def self.from_hash(hash)
|
80
|
-
self.new(hash[:min], hash[:max], hash[:seed])
|
81
|
-
end
|
82
|
-
|
83
|
-
def initialize(min = -0.05, max = 0.05, seed = true)
|
84
|
-
super(seed)
|
85
|
-
@min = min
|
86
|
-
@max = max
|
87
|
-
end
|
88
|
-
|
89
|
-
def init_param(layer, param)
|
90
|
-
Xumo::SFloat.srand(@seed)
|
91
|
-
param.data = param.data.rand(@min, @max)
|
92
|
-
end
|
93
|
-
|
94
|
-
def to_hash
|
95
|
-
super({min: @min, max: @max})
|
96
|
-
end
|
97
|
-
end
|
98
|
-
|
99
|
-
|
100
|
-
class Xavier < Initializer
|
101
|
-
def initialize(seed = true)
|
102
|
-
super
|
103
|
-
end
|
104
|
-
|
105
|
-
def init_param(layer, param)
|
106
|
-
Xumo::SFloat.srand(@seed)
|
107
|
-
num_prev_nodes = layer.input_shape.reduce(:*)
|
108
|
-
param.data = param.data.rand_norm / Math.sqrt(num_prev_nodes)
|
109
|
-
end
|
110
|
-
end
|
111
|
-
|
112
|
-
|
113
|
-
class He < Initializer
|
114
|
-
def initialize(seed = true)
|
115
|
-
super
|
116
|
-
end
|
117
|
-
|
118
|
-
def init_param(layer, param)
|
119
|
-
Xumo::SFloat.srand(@seed)
|
120
|
-
num_prev_nodes = layer.input_shape.reduce(:*)
|
121
|
-
param.data = param.data.rand_norm / Math.sqrt(num_prev_nodes) * Math.sqrt(2)
|
122
|
-
end
|
123
|
-
end
|
124
|
-
|
125
|
-
end
|
126
|
-
end
|
1
|
+
module DNN
|
2
|
+
module Initializers
|
3
|
+
|
4
|
+
class Initializer
|
5
|
+
def initialize(seed = false)
|
6
|
+
@seed = seed == true ? rand(1 << 31) : seed
|
7
|
+
end
|
8
|
+
|
9
|
+
def init_param(layer, param)
|
10
|
+
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'init_params'")
|
11
|
+
end
|
12
|
+
|
13
|
+
def to_hash(merge_hash = nil)
|
14
|
+
hash = {class: self.class.name, seed: @seed}
|
15
|
+
hash.merge!(merge_hash) if merge_hash
|
16
|
+
hash
|
17
|
+
end
|
18
|
+
end
|
19
|
+
|
20
|
+
|
21
|
+
class Zeros < Initializer
|
22
|
+
def init_param(layer, param)
|
23
|
+
param.data = param.data.fill(0)
|
24
|
+
end
|
25
|
+
end
|
26
|
+
|
27
|
+
|
28
|
+
class Const < Initializer
|
29
|
+
attr_reader :const
|
30
|
+
|
31
|
+
def self.from_hash(hash)
|
32
|
+
self.new(hash[:const])
|
33
|
+
end
|
34
|
+
|
35
|
+
def initialize(const)
|
36
|
+
super()
|
37
|
+
@const = const
|
38
|
+
end
|
39
|
+
|
40
|
+
def init_param(layer, param)
|
41
|
+
param.data = param.data.fill(@const)
|
42
|
+
end
|
43
|
+
|
44
|
+
def to_hash
|
45
|
+
super({const: @const})
|
46
|
+
end
|
47
|
+
end
|
48
|
+
|
49
|
+
|
50
|
+
class RandomNormal < Initializer
|
51
|
+
attr_reader :mean
|
52
|
+
attr_reader :std
|
53
|
+
|
54
|
+
def self.from_hash(hash)
|
55
|
+
self.new(hash[:mean], hash[:std], hash[:seed])
|
56
|
+
end
|
57
|
+
|
58
|
+
def initialize(mean = 0, std = 0.05, seed = true)
|
59
|
+
super(seed)
|
60
|
+
@mean = mean
|
61
|
+
@std = std
|
62
|
+
end
|
63
|
+
|
64
|
+
def init_param(layer, param)
|
65
|
+
Xumo::SFloat.srand(@seed)
|
66
|
+
param.data = param.data.rand_norm(@mean, @std)
|
67
|
+
end
|
68
|
+
|
69
|
+
def to_hash
|
70
|
+
super({mean: @mean, std: @std})
|
71
|
+
end
|
72
|
+
end
|
73
|
+
|
74
|
+
|
75
|
+
class RandomUniform < Initializer
|
76
|
+
attr_reader :min
|
77
|
+
attr_reader :max
|
78
|
+
|
79
|
+
def self.from_hash(hash)
|
80
|
+
self.new(hash[:min], hash[:max], hash[:seed])
|
81
|
+
end
|
82
|
+
|
83
|
+
def initialize(min = -0.05, max = 0.05, seed = true)
|
84
|
+
super(seed)
|
85
|
+
@min = min
|
86
|
+
@max = max
|
87
|
+
end
|
88
|
+
|
89
|
+
def init_param(layer, param)
|
90
|
+
Xumo::SFloat.srand(@seed)
|
91
|
+
param.data = param.data.rand(@min, @max)
|
92
|
+
end
|
93
|
+
|
94
|
+
def to_hash
|
95
|
+
super({min: @min, max: @max})
|
96
|
+
end
|
97
|
+
end
|
98
|
+
|
99
|
+
|
100
|
+
class Xavier < Initializer
|
101
|
+
def initialize(seed = true)
|
102
|
+
super
|
103
|
+
end
|
104
|
+
|
105
|
+
def init_param(layer, param)
|
106
|
+
Xumo::SFloat.srand(@seed)
|
107
|
+
num_prev_nodes = layer.input_shape.reduce(:*)
|
108
|
+
param.data = param.data.rand_norm / Math.sqrt(num_prev_nodes)
|
109
|
+
end
|
110
|
+
end
|
111
|
+
|
112
|
+
|
113
|
+
class He < Initializer
|
114
|
+
def initialize(seed = true)
|
115
|
+
super
|
116
|
+
end
|
117
|
+
|
118
|
+
def init_param(layer, param)
|
119
|
+
Xumo::SFloat.srand(@seed)
|
120
|
+
num_prev_nodes = layer.input_shape.reduce(:*)
|
121
|
+
param.data = param.data.rand_norm / Math.sqrt(num_prev_nodes) * Math.sqrt(2)
|
122
|
+
end
|
123
|
+
end
|
124
|
+
|
125
|
+
end
|
126
|
+
end
|
data/lib/dnn/core/layers.rb
CHANGED
@@ -1,307 +1,307 @@
|
|
1
|
-
module DNN
|
2
|
-
module Layers
|
3
|
-
|
4
|
-
# Super class of all optimizer classes.
|
5
|
-
class Layer
|
6
|
-
# @return [Bool] learning_phase Return the true if learning.
|
7
|
-
attr_accessor :learning_phase
|
8
|
-
# @return [Array] Return the shape of the input data.
|
9
|
-
attr_reader :input_shape
|
10
|
-
|
11
|
-
def initialize
|
12
|
-
@built = false
|
13
|
-
end
|
14
|
-
|
15
|
-
# Build the layer.
|
16
|
-
# @param [Array] input_shape Setting the shape of the input data.
|
17
|
-
def build(input_shape)
|
18
|
-
@input_shape = input_shape
|
19
|
-
@learning_phase = true
|
20
|
-
@built = true
|
21
|
-
end
|
22
|
-
|
23
|
-
# Does the layer have already been built?
|
24
|
-
# @return [Bool] If layer have already been built then return true.
|
25
|
-
def built?
|
26
|
-
@built
|
27
|
-
end
|
28
|
-
|
29
|
-
# Forward propagation.
|
30
|
-
def forward(x)
|
31
|
-
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'forward'")
|
32
|
-
end
|
33
|
-
|
34
|
-
# Backward propagation.
|
35
|
-
def backward(dy)
|
36
|
-
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'backward'")
|
37
|
-
end
|
38
|
-
|
39
|
-
# Please reimplement this method as needed.
|
40
|
-
# The default implementation return input_shape.
|
41
|
-
# @return [Array] Return the shape of the output data.
|
42
|
-
def output_shape
|
43
|
-
@input_shape
|
44
|
-
end
|
45
|
-
|
46
|
-
# Layer to a hash.
|
47
|
-
def to_hash(merge_hash = nil)
|
48
|
-
hash = {class: self.class.name}
|
49
|
-
hash.merge!(merge_hash) if merge_hash
|
50
|
-
hash
|
51
|
-
end
|
52
|
-
end
|
53
|
-
|
54
|
-
|
55
|
-
# This class is a superclass of all classes with learning parameters.
|
56
|
-
class HasParamLayer < Layer
|
57
|
-
# @return [Bool] trainable Setting false prevents learning of parameters.
|
58
|
-
attr_accessor :trainable
|
59
|
-
# @return [Array] The parameters of the layer.
|
60
|
-
attr_reader :params
|
61
|
-
|
62
|
-
def initialize
|
63
|
-
super()
|
64
|
-
@params = {}
|
65
|
-
@trainable = true
|
66
|
-
end
|
67
|
-
end
|
68
|
-
|
69
|
-
|
70
|
-
class InputLayer < Layer
|
71
|
-
def self.from_hash(hash)
|
72
|
-
self.new(hash[:input_shape])
|
73
|
-
end
|
74
|
-
|
75
|
-
def initialize(input_dim_or_shape)
|
76
|
-
super()
|
77
|
-
@input_shape = input_dim_or_shape.is_a?(Array) ? input_dim_or_shape : [input_dim_or_shape]
|
78
|
-
end
|
79
|
-
|
80
|
-
def build
|
81
|
-
@built = true
|
82
|
-
@input_shape
|
83
|
-
end
|
84
|
-
|
85
|
-
def forward(x)
|
86
|
-
x
|
87
|
-
end
|
88
|
-
|
89
|
-
def backward(dy)
|
90
|
-
dy
|
91
|
-
end
|
92
|
-
|
93
|
-
def to_hash
|
94
|
-
super({input_shape: @input_shape})
|
95
|
-
end
|
96
|
-
end
|
97
|
-
|
98
|
-
|
99
|
-
# It is a superclass of all connection layers.
|
100
|
-
class Connection < HasParamLayer
|
101
|
-
# @return [DNN::Initializers::Initializer] Weight initializer.
|
102
|
-
attr_reader :weight_initializer
|
103
|
-
# @return [DNN::Initializers::Initializer] Bias initializer.
|
104
|
-
attr_reader :bias_initializer
|
105
|
-
# @return [DNN::Regularizers::Regularizer] Weight regularization.
|
106
|
-
attr_reader :weight_regularizer
|
107
|
-
# @return [DNN::Regularizers::Regularizer] Bias regularization.
|
108
|
-
attr_reader :bias_regularizer
|
109
|
-
|
110
|
-
# @param [DNN::Initializers::Initializer] weight_initializer Weight initializer.
|
111
|
-
# @param [DNN::Initializers::Initializer] bias_initializer Bias initializer.
|
112
|
-
# @param [DNN::Regularizers::Regularizer] weight_regularizer Weight regularization.
|
113
|
-
# @param [DNN::Regularizers::Regularizer] bias_regularizer Bias regularization.
|
114
|
-
# @param [Bool] use_bias whether to use bias.
|
115
|
-
def initialize(weight_initializer: Initializers::RandomNormal.new,
|
116
|
-
bias_initializer: Initializers::Zeros.new,
|
117
|
-
weight_regularizer: nil,
|
118
|
-
bias_regularizer: nil,
|
119
|
-
use_bias: true)
|
120
|
-
super()
|
121
|
-
@weight_initializer = weight_initializer
|
122
|
-
@bias_initializer = bias_initializer
|
123
|
-
@weight_regularizer = weight_regularizer
|
124
|
-
@bias_regularizer = bias_regularizer
|
125
|
-
@params[:weight] = @weight = Param.new(nil, 0)
|
126
|
-
if use_bias
|
127
|
-
@params[:bias] = @bias = Param.new(nil, 0)
|
128
|
-
else
|
129
|
-
@bias = nil
|
130
|
-
end
|
131
|
-
end
|
132
|
-
|
133
|
-
def regularizers
|
134
|
-
regularizers = []
|
135
|
-
regularizers << @weight_regularizer if @weight_regularizer
|
136
|
-
regularizers << @bias_regularizer if @bias_regularizer
|
137
|
-
regularizers
|
138
|
-
end
|
139
|
-
|
140
|
-
# @return [Bool] Return whether to use bias.
|
141
|
-
def use_bias
|
142
|
-
@bias ? true : false
|
143
|
-
end
|
144
|
-
|
145
|
-
def to_hash(merge_hash)
|
146
|
-
super({weight_initializer: @weight_initializer.to_hash,
|
147
|
-
bias_initializer: @bias_initializer.to_hash,
|
148
|
-
weight_regularizer: @weight_regularizer&.to_hash,
|
149
|
-
bias_regularizer: @bias_regularizer&.to_hash,
|
150
|
-
use_bias: use_bias}.merge(merge_hash))
|
151
|
-
end
|
152
|
-
|
153
|
-
private def init_weight_and_bias
|
154
|
-
@weight_initializer.init_param(self, @weight)
|
155
|
-
@weight_regularizer.param = @weight if @weight_regularizer
|
156
|
-
if @bias
|
157
|
-
@bias_initializer.init_param(self, @bias)
|
158
|
-
@bias_regularizer.param = @bias if @bias_regularizer
|
159
|
-
end
|
160
|
-
end
|
161
|
-
end
|
162
|
-
|
163
|
-
|
164
|
-
# Full connnection layer.
|
165
|
-
class Dense < Connection
|
166
|
-
# @return [Integer] number of nodes.
|
167
|
-
attr_reader :num_nodes
|
168
|
-
|
169
|
-
def self.from_hash(hash)
|
170
|
-
self.new(hash[:num_nodes],
|
171
|
-
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
172
|
-
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
173
|
-
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
174
|
-
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
175
|
-
use_bias: hash[:use_bias])
|
176
|
-
end
|
177
|
-
|
178
|
-
# @param [Integer] num_nodes number of nodes.
|
179
|
-
def initialize(num_nodes,
|
180
|
-
weight_initializer: Initializers::RandomNormal.new,
|
181
|
-
bias_initializer: Initializers::Zeros.new,
|
182
|
-
weight_regularizer: nil,
|
183
|
-
bias_regularizer: nil,
|
184
|
-
use_bias: true)
|
185
|
-
super(weight_initializer: weight_initializer, bias_initializer: bias_initializer,
|
186
|
-
weight_regularizer: weight_regularizer, bias_regularizer: bias_regularizer, use_bias: use_bias)
|
187
|
-
@num_nodes = num_nodes
|
188
|
-
end
|
189
|
-
|
190
|
-
def build(input_shape)
|
191
|
-
super
|
192
|
-
num_prev_nodes = input_shape[0]
|
193
|
-
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes)
|
194
|
-
@bias.data = Xumo::SFloat.new(@num_nodes) if @bias
|
195
|
-
init_weight_and_bias
|
196
|
-
end
|
197
|
-
|
198
|
-
def forward(x)
|
199
|
-
@x = x
|
200
|
-
y = x.dot(@weight.data)
|
201
|
-
y += @bias.data if @bias
|
202
|
-
y
|
203
|
-
end
|
204
|
-
|
205
|
-
def backward(dy)
|
206
|
-
if @trainable
|
207
|
-
@weight.grad += @x.transpose.dot(dy)
|
208
|
-
@bias.grad += dy.sum(0) if @bias
|
209
|
-
end
|
210
|
-
dy.dot(@weight.data.transpose)
|
211
|
-
end
|
212
|
-
|
213
|
-
def output_shape
|
214
|
-
[@num_nodes]
|
215
|
-
end
|
216
|
-
|
217
|
-
def to_hash
|
218
|
-
super({num_nodes: @num_nodes})
|
219
|
-
end
|
220
|
-
end
|
221
|
-
|
222
|
-
|
223
|
-
class Flatten < Layer
|
224
|
-
def forward(x)
|
225
|
-
x.reshape(x.shape[0], *output_shape)
|
226
|
-
end
|
227
|
-
|
228
|
-
def backward(dy)
|
229
|
-
dy.reshape(dy.shape[0], *@input_shape)
|
230
|
-
end
|
231
|
-
|
232
|
-
def output_shape
|
233
|
-
[@input_shape.reduce(:*)]
|
234
|
-
end
|
235
|
-
end
|
236
|
-
|
237
|
-
|
238
|
-
class Reshape < Layer
|
239
|
-
def self.from_hash(hash)
|
240
|
-
self.new(hash[:output_shape])
|
241
|
-
end
|
242
|
-
|
243
|
-
def initialize(output_shape)
|
244
|
-
super()
|
245
|
-
@output_shape = output_shape
|
246
|
-
end
|
247
|
-
|
248
|
-
def forward(x)
|
249
|
-
x.reshape(x.shape[0], *@output_shape)
|
250
|
-
end
|
251
|
-
|
252
|
-
def backward(dy)
|
253
|
-
dy.reshape(dy.shape[0], *@input_shape)
|
254
|
-
end
|
255
|
-
|
256
|
-
def output_shape
|
257
|
-
@output_shape
|
258
|
-
end
|
259
|
-
|
260
|
-
def to_hash
|
261
|
-
super({output_shape: @output_shape})
|
262
|
-
end
|
263
|
-
end
|
264
|
-
|
265
|
-
|
266
|
-
class Dropout < Layer
|
267
|
-
# @return [Float] dropout ratio.
|
268
|
-
attr_accessor :dropout_ratio
|
269
|
-
# @return [Float] Use 'weight scaling inference rule'.
|
270
|
-
attr_reader :use_scale
|
271
|
-
|
272
|
-
def self.from_hash(hash)
|
273
|
-
self.new(hash[:dropout_ratio], seed: hash[:seed], use_scale: hash[:use_scale])
|
274
|
-
end
|
275
|
-
|
276
|
-
def initialize(dropout_ratio = 0.5, seed: rand(1 << 31), use_scale: true)
|
277
|
-
super()
|
278
|
-
@dropout_ratio = dropout_ratio
|
279
|
-
@seed = seed
|
280
|
-
@use_scale = use_scale
|
281
|
-
@mask = nil
|
282
|
-
end
|
283
|
-
|
284
|
-
def forward(x)
|
285
|
-
if learning_phase
|
286
|
-
Xumo::SFloat.srand(@seed)
|
287
|
-
@mask = Xumo::SFloat.ones(*x.shape).rand < @dropout_ratio
|
288
|
-
x[@mask] = 0
|
289
|
-
else
|
290
|
-
x *= (1 - @dropout_ratio) if @use_scale
|
291
|
-
end
|
292
|
-
x
|
293
|
-
end
|
294
|
-
|
295
|
-
def backward(dy)
|
296
|
-
dy[@mask] = 0
|
297
|
-
dy
|
298
|
-
end
|
299
|
-
|
300
|
-
def to_hash
|
301
|
-
super({dropout_ratio: @dropout_ratio, seed: @seed, use_scale: @use_scale})
|
302
|
-
end
|
303
|
-
end
|
304
|
-
|
305
|
-
end
|
306
|
-
|
307
|
-
end
|
1
|
+
module DNN
|
2
|
+
module Layers
|
3
|
+
|
4
|
+
# Super class of all optimizer classes.
|
5
|
+
class Layer
|
6
|
+
# @return [Bool] learning_phase Return the true if learning.
|
7
|
+
attr_accessor :learning_phase
|
8
|
+
# @return [Array] Return the shape of the input data.
|
9
|
+
attr_reader :input_shape
|
10
|
+
|
11
|
+
def initialize
|
12
|
+
@built = false
|
13
|
+
end
|
14
|
+
|
15
|
+
# Build the layer.
|
16
|
+
# @param [Array] input_shape Setting the shape of the input data.
|
17
|
+
def build(input_shape)
|
18
|
+
@input_shape = input_shape
|
19
|
+
@learning_phase = true
|
20
|
+
@built = true
|
21
|
+
end
|
22
|
+
|
23
|
+
# Does the layer have already been built?
|
24
|
+
# @return [Bool] If layer have already been built then return true.
|
25
|
+
def built?
|
26
|
+
@built
|
27
|
+
end
|
28
|
+
|
29
|
+
# Forward propagation.
|
30
|
+
def forward(x)
|
31
|
+
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'forward'")
|
32
|
+
end
|
33
|
+
|
34
|
+
# Backward propagation.
|
35
|
+
def backward(dy)
|
36
|
+
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'backward'")
|
37
|
+
end
|
38
|
+
|
39
|
+
# Please reimplement this method as needed.
|
40
|
+
# The default implementation return input_shape.
|
41
|
+
# @return [Array] Return the shape of the output data.
|
42
|
+
def output_shape
|
43
|
+
@input_shape
|
44
|
+
end
|
45
|
+
|
46
|
+
# Layer to a hash.
|
47
|
+
def to_hash(merge_hash = nil)
|
48
|
+
hash = {class: self.class.name}
|
49
|
+
hash.merge!(merge_hash) if merge_hash
|
50
|
+
hash
|
51
|
+
end
|
52
|
+
end
|
53
|
+
|
54
|
+
|
55
|
+
# This class is a superclass of all classes with learning parameters.
|
56
|
+
class HasParamLayer < Layer
|
57
|
+
# @return [Bool] trainable Setting false prevents learning of parameters.
|
58
|
+
attr_accessor :trainable
|
59
|
+
# @return [Array] The parameters of the layer.
|
60
|
+
attr_reader :params
|
61
|
+
|
62
|
+
def initialize
|
63
|
+
super()
|
64
|
+
@params = {}
|
65
|
+
@trainable = true
|
66
|
+
end
|
67
|
+
end
|
68
|
+
|
69
|
+
|
70
|
+
class InputLayer < Layer
|
71
|
+
def self.from_hash(hash)
|
72
|
+
self.new(hash[:input_shape])
|
73
|
+
end
|
74
|
+
|
75
|
+
def initialize(input_dim_or_shape)
|
76
|
+
super()
|
77
|
+
@input_shape = input_dim_or_shape.is_a?(Array) ? input_dim_or_shape : [input_dim_or_shape]
|
78
|
+
end
|
79
|
+
|
80
|
+
def build
|
81
|
+
@built = true
|
82
|
+
@input_shape
|
83
|
+
end
|
84
|
+
|
85
|
+
def forward(x)
|
86
|
+
x
|
87
|
+
end
|
88
|
+
|
89
|
+
def backward(dy)
|
90
|
+
dy
|
91
|
+
end
|
92
|
+
|
93
|
+
def to_hash
|
94
|
+
super({input_shape: @input_shape})
|
95
|
+
end
|
96
|
+
end
|
97
|
+
|
98
|
+
|
99
|
+
# It is a superclass of all connection layers.
|
100
|
+
class Connection < HasParamLayer
|
101
|
+
# @return [DNN::Initializers::Initializer] Weight initializer.
|
102
|
+
attr_reader :weight_initializer
|
103
|
+
# @return [DNN::Initializers::Initializer] Bias initializer.
|
104
|
+
attr_reader :bias_initializer
|
105
|
+
# @return [DNN::Regularizers::Regularizer] Weight regularization.
|
106
|
+
attr_reader :weight_regularizer
|
107
|
+
# @return [DNN::Regularizers::Regularizer] Bias regularization.
|
108
|
+
attr_reader :bias_regularizer
|
109
|
+
|
110
|
+
# @param [DNN::Initializers::Initializer] weight_initializer Weight initializer.
|
111
|
+
# @param [DNN::Initializers::Initializer] bias_initializer Bias initializer.
|
112
|
+
# @param [DNN::Regularizers::Regularizer] weight_regularizer Weight regularization.
|
113
|
+
# @param [DNN::Regularizers::Regularizer] bias_regularizer Bias regularization.
|
114
|
+
# @param [Bool] use_bias whether to use bias.
|
115
|
+
def initialize(weight_initializer: Initializers::RandomNormal.new,
|
116
|
+
bias_initializer: Initializers::Zeros.new,
|
117
|
+
weight_regularizer: nil,
|
118
|
+
bias_regularizer: nil,
|
119
|
+
use_bias: true)
|
120
|
+
super()
|
121
|
+
@weight_initializer = weight_initializer
|
122
|
+
@bias_initializer = bias_initializer
|
123
|
+
@weight_regularizer = weight_regularizer
|
124
|
+
@bias_regularizer = bias_regularizer
|
125
|
+
@params[:weight] = @weight = Param.new(nil, 0)
|
126
|
+
if use_bias
|
127
|
+
@params[:bias] = @bias = Param.new(nil, 0)
|
128
|
+
else
|
129
|
+
@bias = nil
|
130
|
+
end
|
131
|
+
end
|
132
|
+
|
133
|
+
def regularizers
|
134
|
+
regularizers = []
|
135
|
+
regularizers << @weight_regularizer if @weight_regularizer
|
136
|
+
regularizers << @bias_regularizer if @bias_regularizer
|
137
|
+
regularizers
|
138
|
+
end
|
139
|
+
|
140
|
+
# @return [Bool] Return whether to use bias.
|
141
|
+
def use_bias
|
142
|
+
@bias ? true : false
|
143
|
+
end
|
144
|
+
|
145
|
+
def to_hash(merge_hash)
|
146
|
+
super({weight_initializer: @weight_initializer.to_hash,
|
147
|
+
bias_initializer: @bias_initializer.to_hash,
|
148
|
+
weight_regularizer: @weight_regularizer&.to_hash,
|
149
|
+
bias_regularizer: @bias_regularizer&.to_hash,
|
150
|
+
use_bias: use_bias}.merge(merge_hash))
|
151
|
+
end
|
152
|
+
|
153
|
+
private def init_weight_and_bias
|
154
|
+
@weight_initializer.init_param(self, @weight)
|
155
|
+
@weight_regularizer.param = @weight if @weight_regularizer
|
156
|
+
if @bias
|
157
|
+
@bias_initializer.init_param(self, @bias)
|
158
|
+
@bias_regularizer.param = @bias if @bias_regularizer
|
159
|
+
end
|
160
|
+
end
|
161
|
+
end
|
162
|
+
|
163
|
+
|
164
|
+
# Full connnection layer.
|
165
|
+
class Dense < Connection
|
166
|
+
# @return [Integer] number of nodes.
|
167
|
+
attr_reader :num_nodes
|
168
|
+
|
169
|
+
def self.from_hash(hash)
|
170
|
+
self.new(hash[:num_nodes],
|
171
|
+
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
172
|
+
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
173
|
+
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
174
|
+
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
175
|
+
use_bias: hash[:use_bias])
|
176
|
+
end
|
177
|
+
|
178
|
+
# @param [Integer] num_nodes number of nodes.
|
179
|
+
def initialize(num_nodes,
|
180
|
+
weight_initializer: Initializers::RandomNormal.new,
|
181
|
+
bias_initializer: Initializers::Zeros.new,
|
182
|
+
weight_regularizer: nil,
|
183
|
+
bias_regularizer: nil,
|
184
|
+
use_bias: true)
|
185
|
+
super(weight_initializer: weight_initializer, bias_initializer: bias_initializer,
|
186
|
+
weight_regularizer: weight_regularizer, bias_regularizer: bias_regularizer, use_bias: use_bias)
|
187
|
+
@num_nodes = num_nodes
|
188
|
+
end
|
189
|
+
|
190
|
+
def build(input_shape)
|
191
|
+
super
|
192
|
+
num_prev_nodes = input_shape[0]
|
193
|
+
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes)
|
194
|
+
@bias.data = Xumo::SFloat.new(@num_nodes) if @bias
|
195
|
+
init_weight_and_bias
|
196
|
+
end
|
197
|
+
|
198
|
+
def forward(x)
|
199
|
+
@x = x
|
200
|
+
y = x.dot(@weight.data)
|
201
|
+
y += @bias.data if @bias
|
202
|
+
y
|
203
|
+
end
|
204
|
+
|
205
|
+
def backward(dy)
|
206
|
+
if @trainable
|
207
|
+
@weight.grad += @x.transpose.dot(dy)
|
208
|
+
@bias.grad += dy.sum(0) if @bias
|
209
|
+
end
|
210
|
+
dy.dot(@weight.data.transpose)
|
211
|
+
end
|
212
|
+
|
213
|
+
def output_shape
|
214
|
+
[@num_nodes]
|
215
|
+
end
|
216
|
+
|
217
|
+
def to_hash
|
218
|
+
super({num_nodes: @num_nodes})
|
219
|
+
end
|
220
|
+
end
|
221
|
+
|
222
|
+
|
223
|
+
class Flatten < Layer
|
224
|
+
def forward(x)
|
225
|
+
x.reshape(x.shape[0], *output_shape)
|
226
|
+
end
|
227
|
+
|
228
|
+
def backward(dy)
|
229
|
+
dy.reshape(dy.shape[0], *@input_shape)
|
230
|
+
end
|
231
|
+
|
232
|
+
def output_shape
|
233
|
+
[@input_shape.reduce(:*)]
|
234
|
+
end
|
235
|
+
end
|
236
|
+
|
237
|
+
|
238
|
+
class Reshape < Layer
|
239
|
+
def self.from_hash(hash)
|
240
|
+
self.new(hash[:output_shape])
|
241
|
+
end
|
242
|
+
|
243
|
+
def initialize(output_shape)
|
244
|
+
super()
|
245
|
+
@output_shape = output_shape
|
246
|
+
end
|
247
|
+
|
248
|
+
def forward(x)
|
249
|
+
x.reshape(x.shape[0], *@output_shape)
|
250
|
+
end
|
251
|
+
|
252
|
+
def backward(dy)
|
253
|
+
dy.reshape(dy.shape[0], *@input_shape)
|
254
|
+
end
|
255
|
+
|
256
|
+
def output_shape
|
257
|
+
@output_shape
|
258
|
+
end
|
259
|
+
|
260
|
+
def to_hash
|
261
|
+
super({output_shape: @output_shape})
|
262
|
+
end
|
263
|
+
end
|
264
|
+
|
265
|
+
|
266
|
+
class Dropout < Layer
|
267
|
+
# @return [Float] dropout ratio.
|
268
|
+
attr_accessor :dropout_ratio
|
269
|
+
# @return [Float] Use 'weight scaling inference rule'.
|
270
|
+
attr_reader :use_scale
|
271
|
+
|
272
|
+
def self.from_hash(hash)
|
273
|
+
self.new(hash[:dropout_ratio], seed: hash[:seed], use_scale: hash[:use_scale])
|
274
|
+
end
|
275
|
+
|
276
|
+
def initialize(dropout_ratio = 0.5, seed: rand(1 << 31), use_scale: true)
|
277
|
+
super()
|
278
|
+
@dropout_ratio = dropout_ratio
|
279
|
+
@seed = seed
|
280
|
+
@use_scale = use_scale
|
281
|
+
@mask = nil
|
282
|
+
end
|
283
|
+
|
284
|
+
def forward(x)
|
285
|
+
if learning_phase
|
286
|
+
Xumo::SFloat.srand(@seed)
|
287
|
+
@mask = Xumo::SFloat.ones(*x.shape).rand < @dropout_ratio
|
288
|
+
x[@mask] = 0
|
289
|
+
else
|
290
|
+
x *= (1 - @dropout_ratio) if @use_scale
|
291
|
+
end
|
292
|
+
x
|
293
|
+
end
|
294
|
+
|
295
|
+
def backward(dy)
|
296
|
+
dy[@mask] = 0
|
297
|
+
dy
|
298
|
+
end
|
299
|
+
|
300
|
+
def to_hash
|
301
|
+
super({dropout_ratio: @dropout_ratio, seed: @seed, use_scale: @use_scale})
|
302
|
+
end
|
303
|
+
end
|
304
|
+
|
305
|
+
end
|
306
|
+
|
307
|
+
end
|