ruby-dnn 0.10.1 → 0.10.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- 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/param.rb
CHANGED
|
@@ -1,9 +1,9 @@
|
|
|
1
|
-
class DNN::Param
|
|
2
|
-
attr_accessor :data
|
|
3
|
-
attr_accessor :grad
|
|
4
|
-
|
|
5
|
-
def initialize(data = nil, grad = nil)
|
|
6
|
-
@data = data
|
|
7
|
-
@grad = grad
|
|
8
|
-
end
|
|
9
|
-
end
|
|
1
|
+
class DNN::Param
|
|
2
|
+
attr_accessor :data
|
|
3
|
+
attr_accessor :grad
|
|
4
|
+
|
|
5
|
+
def initialize(data = nil, grad = nil)
|
|
6
|
+
@data = data
|
|
7
|
+
@grad = grad
|
|
8
|
+
end
|
|
9
|
+
end
|
|
@@ -1,106 +1,106 @@
|
|
|
1
|
-
module DNN
|
|
2
|
-
module Regularizers
|
|
3
|
-
|
|
4
|
-
class Regularizer
|
|
5
|
-
attr_accessor :param
|
|
6
|
-
|
|
7
|
-
def forward(x)
|
|
8
|
-
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'forward'")
|
|
9
|
-
end
|
|
10
|
-
|
|
11
|
-
def backward
|
|
12
|
-
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'backward'")
|
|
13
|
-
end
|
|
14
|
-
|
|
15
|
-
def to_hash(merge_hash)
|
|
16
|
-
hash = {class: self.class.name}
|
|
17
|
-
hash.merge!(merge_hash)
|
|
18
|
-
hash
|
|
19
|
-
end
|
|
20
|
-
end
|
|
21
|
-
|
|
22
|
-
class L1 < Regularizer
|
|
23
|
-
attr_accessor :l1_lambda
|
|
24
|
-
|
|
25
|
-
def self.from_hash(hash)
|
|
26
|
-
L1.new(hash[:l1_lambda])
|
|
27
|
-
end
|
|
28
|
-
|
|
29
|
-
def initialize(l1_lambda = 0.01)
|
|
30
|
-
@l1_lambda = l1_lambda
|
|
31
|
-
end
|
|
32
|
-
|
|
33
|
-
def forward(x)
|
|
34
|
-
x + @l1_lambda * @param.data.abs.sum
|
|
35
|
-
end
|
|
36
|
-
|
|
37
|
-
def backward
|
|
38
|
-
dparam = Xumo::SFloat.ones(*@param.data.shape)
|
|
39
|
-
dparam[@param.data < 0] = -1
|
|
40
|
-
@param.grad += @l1_lambda * dparam
|
|
41
|
-
end
|
|
42
|
-
|
|
43
|
-
def to_hash
|
|
44
|
-
super(l1_lambda: @l1_lambda)
|
|
45
|
-
end
|
|
46
|
-
end
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
class L2 < Regularizer
|
|
50
|
-
attr_accessor :l2_lambda
|
|
51
|
-
|
|
52
|
-
def self.from_hash(hash)
|
|
53
|
-
L2.new(hash[:l2_lambda])
|
|
54
|
-
end
|
|
55
|
-
|
|
56
|
-
def initialize(l2_lambda = 0.01)
|
|
57
|
-
@l2_lambda = l2_lambda
|
|
58
|
-
end
|
|
59
|
-
|
|
60
|
-
def forward(x)
|
|
61
|
-
x + 0.5 * @l2_lambda * (@param.data**2).sum
|
|
62
|
-
end
|
|
63
|
-
|
|
64
|
-
def backward
|
|
65
|
-
@param.grad += @l2_lambda * @param.data
|
|
66
|
-
end
|
|
67
|
-
|
|
68
|
-
def to_hash
|
|
69
|
-
super(l2_lambda: @l2_lambda)
|
|
70
|
-
end
|
|
71
|
-
end
|
|
72
|
-
|
|
73
|
-
class L1L2 < Regularizer
|
|
74
|
-
attr_accessor :l1_lambda
|
|
75
|
-
attr_accessor :l2_lambda
|
|
76
|
-
|
|
77
|
-
def self.from_hash(hash)
|
|
78
|
-
L1L2.new(hash[:l1_lambda], hash[:l2_lambda])
|
|
79
|
-
end
|
|
80
|
-
|
|
81
|
-
def initialize(l1_lambda = 0.01, l2_lambda = 0.01)
|
|
82
|
-
@l1_lambda = l1_lambda
|
|
83
|
-
@l2_lambda = l2_lambda
|
|
84
|
-
end
|
|
85
|
-
|
|
86
|
-
def forward(x)
|
|
87
|
-
l1 = @l1_lambda * @param.data.abs.sum
|
|
88
|
-
l2 = 0.5 * @l2_lambda * (@param.data**2).sum
|
|
89
|
-
x + l1 + l2
|
|
90
|
-
end
|
|
91
|
-
|
|
92
|
-
def backward
|
|
93
|
-
dparam = Xumo::SFloat.ones(*@param.data.shape)
|
|
94
|
-
dparam[@param.data < 0] = -1
|
|
95
|
-
@param.grad += @l1_lambda * dparam
|
|
96
|
-
@param.grad += @l2_lambda * @param.data
|
|
97
|
-
end
|
|
98
|
-
|
|
99
|
-
def to_hash
|
|
100
|
-
super(l1_lambda: l1_lambda, l2_lambda: l2_lambda)
|
|
101
|
-
end
|
|
102
|
-
|
|
103
|
-
end
|
|
104
|
-
|
|
105
|
-
end
|
|
106
|
-
end
|
|
1
|
+
module DNN
|
|
2
|
+
module Regularizers
|
|
3
|
+
|
|
4
|
+
class Regularizer
|
|
5
|
+
attr_accessor :param
|
|
6
|
+
|
|
7
|
+
def forward(x)
|
|
8
|
+
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'forward'")
|
|
9
|
+
end
|
|
10
|
+
|
|
11
|
+
def backward
|
|
12
|
+
raise NotImplementedError.new("Class '#{self.class.name}' has implement method 'backward'")
|
|
13
|
+
end
|
|
14
|
+
|
|
15
|
+
def to_hash(merge_hash)
|
|
16
|
+
hash = {class: self.class.name}
|
|
17
|
+
hash.merge!(merge_hash)
|
|
18
|
+
hash
|
|
19
|
+
end
|
|
20
|
+
end
|
|
21
|
+
|
|
22
|
+
class L1 < Regularizer
|
|
23
|
+
attr_accessor :l1_lambda
|
|
24
|
+
|
|
25
|
+
def self.from_hash(hash)
|
|
26
|
+
L1.new(hash[:l1_lambda])
|
|
27
|
+
end
|
|
28
|
+
|
|
29
|
+
def initialize(l1_lambda = 0.01)
|
|
30
|
+
@l1_lambda = l1_lambda
|
|
31
|
+
end
|
|
32
|
+
|
|
33
|
+
def forward(x)
|
|
34
|
+
x + @l1_lambda * @param.data.abs.sum
|
|
35
|
+
end
|
|
36
|
+
|
|
37
|
+
def backward
|
|
38
|
+
dparam = Xumo::SFloat.ones(*@param.data.shape)
|
|
39
|
+
dparam[@param.data < 0] = -1
|
|
40
|
+
@param.grad += @l1_lambda * dparam
|
|
41
|
+
end
|
|
42
|
+
|
|
43
|
+
def to_hash
|
|
44
|
+
super(l1_lambda: @l1_lambda)
|
|
45
|
+
end
|
|
46
|
+
end
|
|
47
|
+
|
|
48
|
+
|
|
49
|
+
class L2 < Regularizer
|
|
50
|
+
attr_accessor :l2_lambda
|
|
51
|
+
|
|
52
|
+
def self.from_hash(hash)
|
|
53
|
+
L2.new(hash[:l2_lambda])
|
|
54
|
+
end
|
|
55
|
+
|
|
56
|
+
def initialize(l2_lambda = 0.01)
|
|
57
|
+
@l2_lambda = l2_lambda
|
|
58
|
+
end
|
|
59
|
+
|
|
60
|
+
def forward(x)
|
|
61
|
+
x + 0.5 * @l2_lambda * (@param.data**2).sum
|
|
62
|
+
end
|
|
63
|
+
|
|
64
|
+
def backward
|
|
65
|
+
@param.grad += @l2_lambda * @param.data
|
|
66
|
+
end
|
|
67
|
+
|
|
68
|
+
def to_hash
|
|
69
|
+
super(l2_lambda: @l2_lambda)
|
|
70
|
+
end
|
|
71
|
+
end
|
|
72
|
+
|
|
73
|
+
class L1L2 < Regularizer
|
|
74
|
+
attr_accessor :l1_lambda
|
|
75
|
+
attr_accessor :l2_lambda
|
|
76
|
+
|
|
77
|
+
def self.from_hash(hash)
|
|
78
|
+
L1L2.new(hash[:l1_lambda], hash[:l2_lambda])
|
|
79
|
+
end
|
|
80
|
+
|
|
81
|
+
def initialize(l1_lambda = 0.01, l2_lambda = 0.01)
|
|
82
|
+
@l1_lambda = l1_lambda
|
|
83
|
+
@l2_lambda = l2_lambda
|
|
84
|
+
end
|
|
85
|
+
|
|
86
|
+
def forward(x)
|
|
87
|
+
l1 = @l1_lambda * @param.data.abs.sum
|
|
88
|
+
l2 = 0.5 * @l2_lambda * (@param.data**2).sum
|
|
89
|
+
x + l1 + l2
|
|
90
|
+
end
|
|
91
|
+
|
|
92
|
+
def backward
|
|
93
|
+
dparam = Xumo::SFloat.ones(*@param.data.shape)
|
|
94
|
+
dparam[@param.data < 0] = -1
|
|
95
|
+
@param.grad += @l1_lambda * dparam
|
|
96
|
+
@param.grad += @l2_lambda * @param.data
|
|
97
|
+
end
|
|
98
|
+
|
|
99
|
+
def to_hash
|
|
100
|
+
super(l1_lambda: l1_lambda, l2_lambda: l2_lambda)
|
|
101
|
+
end
|
|
102
|
+
|
|
103
|
+
end
|
|
104
|
+
|
|
105
|
+
end
|
|
106
|
+
end
|
data/lib/dnn/core/rnn_layers.rb
CHANGED
|
@@ -1,464 +1,464 @@
|
|
|
1
|
-
module DNN
|
|
2
|
-
module Layers
|
|
3
|
-
|
|
4
|
-
# Super class of all RNN classes.
|
|
5
|
-
class RNN < Connection
|
|
6
|
-
include Initializers
|
|
7
|
-
|
|
8
|
-
# @return [Integer] number of nodes.
|
|
9
|
-
attr_reader :num_nodes
|
|
10
|
-
# @return [Bool] Maintain state between batches.
|
|
11
|
-
attr_reader :stateful
|
|
12
|
-
# @return [Bool] Set the false, only the last of each cell of RNN is left.
|
|
13
|
-
attr_reader :return_sequences
|
|
14
|
-
# @return [DNN::Initializers::Initializer] Recurrent weight initializer.
|
|
15
|
-
attr_reader :recurrent_weight_initializer
|
|
16
|
-
# @return [DNN::Regularizers::Regularizer] Recurrent weight regularization.
|
|
17
|
-
attr_reader :recurrent_weight_regularizer
|
|
18
|
-
|
|
19
|
-
def initialize(num_nodes,
|
|
20
|
-
stateful: false,
|
|
21
|
-
return_sequences: true,
|
|
22
|
-
weight_initializer: RandomNormal.new,
|
|
23
|
-
recurrent_weight_initializer: RandomNormal.new,
|
|
24
|
-
bias_initializer: Zeros.new,
|
|
25
|
-
weight_regularizer: nil,
|
|
26
|
-
recurrent_weight_regularizer: nil,
|
|
27
|
-
bias_regularizer: nil,
|
|
28
|
-
use_bias: true)
|
|
29
|
-
super(weight_initializer: weight_initializer, bias_initializer: bias_initializer,
|
|
30
|
-
weight_regularizer: weight_regularizer, bias_regularizer: bias_regularizer, use_bias: use_bias)
|
|
31
|
-
@num_nodes = num_nodes
|
|
32
|
-
@stateful = stateful
|
|
33
|
-
@return_sequences = return_sequences
|
|
34
|
-
@layers = []
|
|
35
|
-
@hidden = @params[:hidden] = Param.new
|
|
36
|
-
@params[:recurrent_weight] = @recurrent_weight = Param.new(nil, 0)
|
|
37
|
-
@recurrent_weight_initializer = recurrent_weight_initializer
|
|
38
|
-
@recurrent_weight_regularizer = recurrent_weight_regularizer
|
|
39
|
-
end
|
|
40
|
-
|
|
41
|
-
def build(input_shape)
|
|
42
|
-
super
|
|
43
|
-
@time_length = @input_shape[0]
|
|
44
|
-
end
|
|
45
|
-
|
|
46
|
-
def forward(xs)
|
|
47
|
-
@xs_shape = xs.shape
|
|
48
|
-
hs = Xumo::SFloat.zeros(xs.shape[0], @time_length, @num_nodes)
|
|
49
|
-
h = (@stateful && @hidden.data) ? @hidden.data : Xumo::SFloat.zeros(xs.shape[0], @num_nodes)
|
|
50
|
-
xs.shape[1].times do |t|
|
|
51
|
-
x = xs[true, t, false]
|
|
52
|
-
@layers[t].trainable = @trainable
|
|
53
|
-
h = @layers[t].forward(x, h)
|
|
54
|
-
hs[true, t, false] = h
|
|
55
|
-
end
|
|
56
|
-
@hidden.data = h
|
|
57
|
-
@return_sequences ? hs : h
|
|
58
|
-
end
|
|
59
|
-
|
|
60
|
-
def backward(dh2s)
|
|
61
|
-
unless @return_sequences
|
|
62
|
-
dh = dh2s
|
|
63
|
-
dh2s = Xumo::SFloat.zeros(dh.shape[0], @time_length, dh.shape[1])
|
|
64
|
-
dh2s[true, -1, false] = dh
|
|
65
|
-
end
|
|
66
|
-
dxs = Xumo::SFloat.zeros(@xs_shape)
|
|
67
|
-
dh = 0
|
|
68
|
-
(0...dh2s.shape[1]).to_a.reverse.each do |t|
|
|
69
|
-
dh2 = dh2s[true, t, false]
|
|
70
|
-
dx, dh = @layers[t].backward(dh2 + dh)
|
|
71
|
-
dxs[true, t, false] = dx
|
|
72
|
-
end
|
|
73
|
-
dxs
|
|
74
|
-
end
|
|
75
|
-
|
|
76
|
-
def output_shape
|
|
77
|
-
@return_sequences ? [@time_length, @num_nodes] : [@num_nodes]
|
|
78
|
-
end
|
|
79
|
-
|
|
80
|
-
def to_hash(merge_hash = nil)
|
|
81
|
-
hash = {
|
|
82
|
-
num_nodes: @num_nodes,
|
|
83
|
-
stateful: @stateful,
|
|
84
|
-
return_sequences: @return_sequences,
|
|
85
|
-
recurrent_weight_initializer: @recurrent_weight_initializer.to_hash,
|
|
86
|
-
recurrent_weight_regularizer: @recurrent_weight_regularizer&.to_hash,
|
|
87
|
-
}
|
|
88
|
-
hash.merge!(merge_hash) if merge_hash
|
|
89
|
-
super(hash)
|
|
90
|
-
end
|
|
91
|
-
|
|
92
|
-
# Reset the state of RNN.
|
|
93
|
-
def reset_state
|
|
94
|
-
@hidden.data = @hidden.data.fill(0) if @hidden.data
|
|
95
|
-
end
|
|
96
|
-
|
|
97
|
-
def regularizers
|
|
98
|
-
regularizers = []
|
|
99
|
-
regularizers << @weight_regularizer if @weight_regularizer
|
|
100
|
-
regularizers << @recurrent_weight_regularizer if @recurrent_weight_regularizer
|
|
101
|
-
regularizers << @bias_regularizer if @bias_regularizer
|
|
102
|
-
regularizers
|
|
103
|
-
end
|
|
104
|
-
|
|
105
|
-
private def init_weight_and_bias
|
|
106
|
-
super
|
|
107
|
-
@recurrent_weight_initializer.init_param(self, @recurrent_weight)
|
|
108
|
-
@recurrent_weight_regularizer.param = @recurrent_weight if @recurrent_weight_regularizer
|
|
109
|
-
end
|
|
110
|
-
end
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
class SimpleRNN_Dense
|
|
114
|
-
attr_accessor :trainable
|
|
115
|
-
|
|
116
|
-
def initialize(weight, recurrent_weight, bias, activation)
|
|
117
|
-
@weight = weight
|
|
118
|
-
@recurrent_weight = recurrent_weight
|
|
119
|
-
@bias = bias
|
|
120
|
-
@activation = activation.clone
|
|
121
|
-
@trainable = true
|
|
122
|
-
end
|
|
123
|
-
|
|
124
|
-
def forward(x, h)
|
|
125
|
-
@x = x
|
|
126
|
-
@h = h
|
|
127
|
-
h2 = x.dot(@weight.data) + h.dot(@recurrent_weight.data)
|
|
128
|
-
h2 += @bias.data if @bias
|
|
129
|
-
@activation.forward(h2)
|
|
130
|
-
end
|
|
131
|
-
|
|
132
|
-
def backward(dh2)
|
|
133
|
-
dh2 = @activation.backward(dh2)
|
|
134
|
-
if @trainable
|
|
135
|
-
@weight.grad += @x.transpose.dot(dh2)
|
|
136
|
-
@recurrent_weight.grad += @h.transpose.dot(dh2)
|
|
137
|
-
@bias.grad += dh2.sum(0) if @bias
|
|
138
|
-
end
|
|
139
|
-
dx = dh2.dot(@weight.data.transpose)
|
|
140
|
-
dh = dh2.dot(@recurrent_weight.data.transpose)
|
|
141
|
-
[dx, dh]
|
|
142
|
-
end
|
|
143
|
-
end
|
|
144
|
-
|
|
145
|
-
|
|
146
|
-
class SimpleRNN < RNN
|
|
147
|
-
include Activations
|
|
148
|
-
|
|
149
|
-
attr_reader :activation
|
|
150
|
-
|
|
151
|
-
def self.from_hash(hash)
|
|
152
|
-
simple_rnn = self.new(hash[:num_nodes],
|
|
153
|
-
stateful: hash[:stateful],
|
|
154
|
-
return_sequences: hash[:return_sequences],
|
|
155
|
-
activation: Utils.from_hash(hash[:activation]),
|
|
156
|
-
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
|
157
|
-
recurrent_weight_initializer: Utils.from_hash(hash[:recurrent_weight_initializer]),
|
|
158
|
-
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
|
159
|
-
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
|
160
|
-
recurrent_weight_regularizer: Utils.from_hash(hash[:recurrent_weight_regularizer]),
|
|
161
|
-
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
|
162
|
-
use_bias: hash[:use_bias])
|
|
163
|
-
simple_rnn
|
|
164
|
-
end
|
|
165
|
-
|
|
166
|
-
def initialize(num_nodes,
|
|
167
|
-
stateful: false,
|
|
168
|
-
return_sequences: true,
|
|
169
|
-
activation: Tanh.new,
|
|
170
|
-
weight_initializer: RandomNormal.new,
|
|
171
|
-
recurrent_weight_initializer: RandomNormal.new,
|
|
172
|
-
bias_initializer: Zeros.new,
|
|
173
|
-
weight_regularizer: nil,
|
|
174
|
-
recurrent_weight_regularizer: nil,
|
|
175
|
-
bias_regularizer: nil,
|
|
176
|
-
use_bias: true)
|
|
177
|
-
super(num_nodes,
|
|
178
|
-
stateful: stateful,
|
|
179
|
-
return_sequences: return_sequences,
|
|
180
|
-
weight_initializer: weight_initializer,
|
|
181
|
-
recurrent_weight_initializer: recurrent_weight_initializer,
|
|
182
|
-
bias_initializer: bias_initializer,
|
|
183
|
-
weight_regularizer: weight_regularizer,
|
|
184
|
-
recurrent_weight_regularizer: recurrent_weight_regularizer,
|
|
185
|
-
bias_regularizer: bias_regularizer,
|
|
186
|
-
use_bias: use_bias)
|
|
187
|
-
@activation = activation
|
|
188
|
-
end
|
|
189
|
-
|
|
190
|
-
def build(input_shape)
|
|
191
|
-
super
|
|
192
|
-
num_prev_nodes = input_shape[1]
|
|
193
|
-
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes)
|
|
194
|
-
@recurrent_weight.data = Xumo::SFloat.new(@num_nodes, @num_nodes)
|
|
195
|
-
@bias.data = Xumo::SFloat.new(@num_nodes) if @bias
|
|
196
|
-
init_weight_and_bias
|
|
197
|
-
@time_length.times do |t|
|
|
198
|
-
@layers << SimpleRNN_Dense.new(@weight, @recurrent_weight, @bias, @activation)
|
|
199
|
-
end
|
|
200
|
-
end
|
|
201
|
-
|
|
202
|
-
def to_hash
|
|
203
|
-
super({activation: @activation.to_hash})
|
|
204
|
-
end
|
|
205
|
-
end
|
|
206
|
-
|
|
207
|
-
|
|
208
|
-
class LSTM_Dense
|
|
209
|
-
attr_accessor :trainable
|
|
210
|
-
|
|
211
|
-
def initialize(weight, recurrent_weight, bias)
|
|
212
|
-
@weight = weight
|
|
213
|
-
@recurrent_weight = recurrent_weight
|
|
214
|
-
@bias = bias
|
|
215
|
-
@tanh = Tanh.new
|
|
216
|
-
@g_tanh = Tanh.new
|
|
217
|
-
@forget_sigmoid = Sigmoid.new
|
|
218
|
-
@in_sigmoid = Sigmoid.new
|
|
219
|
-
@out_sigmoid = Sigmoid.new
|
|
220
|
-
@trainable = true
|
|
221
|
-
end
|
|
222
|
-
|
|
223
|
-
def forward(x, h, c)
|
|
224
|
-
@x = x
|
|
225
|
-
@h = h
|
|
226
|
-
@c = c
|
|
227
|
-
num_nodes = h.shape[1]
|
|
228
|
-
a = x.dot(@weight.data) + h.dot(@recurrent_weight.data)
|
|
229
|
-
a += @bias.data if @bias
|
|
230
|
-
|
|
231
|
-
@forget = @forget_sigmoid.forward(a[true, 0...num_nodes])
|
|
232
|
-
@g = @g_tanh.forward(a[true, num_nodes...(num_nodes * 2)])
|
|
233
|
-
@in = @in_sigmoid.forward(a[true, (num_nodes * 2)...(num_nodes * 3)])
|
|
234
|
-
@out = @out_sigmoid.forward(a[true, (num_nodes * 3)..-1])
|
|
235
|
-
|
|
236
|
-
c2 = @forget * c + @g * @in
|
|
237
|
-
@tanh_c2 = @tanh.forward(c2)
|
|
238
|
-
h2 = @out * @tanh_c2
|
|
239
|
-
[h2, c2]
|
|
240
|
-
end
|
|
241
|
-
|
|
242
|
-
def backward(dh2, dc2)
|
|
243
|
-
dh2_tmp = @tanh_c2 * dh2
|
|
244
|
-
dc2_tmp = @tanh.backward(@out * dh2) + dc2
|
|
245
|
-
|
|
246
|
-
dout = @out_sigmoid.backward(dh2_tmp)
|
|
247
|
-
din = @in_sigmoid.backward(dc2_tmp * @g)
|
|
248
|
-
dg = @g_tanh.backward(dc2_tmp * @in)
|
|
249
|
-
dforget = @forget_sigmoid.backward(dc2_tmp * @c)
|
|
250
|
-
|
|
251
|
-
da = Xumo::SFloat.hstack([dforget, dg, din, dout])
|
|
252
|
-
|
|
253
|
-
if @trainable
|
|
254
|
-
@weight.grad += @x.transpose.dot(da)
|
|
255
|
-
@recurrent_weight.grad += @h.transpose.dot(da)
|
|
256
|
-
@bias.grad += da.sum(0) if @bias
|
|
257
|
-
end
|
|
258
|
-
dx = da.dot(@weight.data.transpose)
|
|
259
|
-
dh = da.dot(@recurrent_weight.data.transpose)
|
|
260
|
-
dc = dc2_tmp * @forget
|
|
261
|
-
[dx, dh, dc]
|
|
262
|
-
end
|
|
263
|
-
end
|
|
264
|
-
|
|
265
|
-
|
|
266
|
-
class LSTM < RNN
|
|
267
|
-
def self.from_hash(hash)
|
|
268
|
-
lstm = self.new(hash[:num_nodes],
|
|
269
|
-
stateful: hash[:stateful],
|
|
270
|
-
return_sequences: hash[:return_sequences],
|
|
271
|
-
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
|
272
|
-
recurrent_weight_initializer: Utils.from_hash(hash[:recurrent_weight_initializer]),
|
|
273
|
-
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
|
274
|
-
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
|
275
|
-
recurrent_weight_regularizer: Utils.from_hash(hash[:recurrent_weight_regularizer]),
|
|
276
|
-
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
|
277
|
-
use_bias: hash[:use_bias])
|
|
278
|
-
lstm
|
|
279
|
-
end
|
|
280
|
-
|
|
281
|
-
def initialize(num_nodes,
|
|
282
|
-
stateful: false,
|
|
283
|
-
return_sequences: true,
|
|
284
|
-
weight_initializer: RandomNormal.new,
|
|
285
|
-
recurrent_weight_initializer: RandomNormal.new,
|
|
286
|
-
bias_initializer: Zeros.new,
|
|
287
|
-
weight_regularizer: nil,
|
|
288
|
-
recurrent_weight_regularizer: nil,
|
|
289
|
-
bias_regularizer: nil,
|
|
290
|
-
use_bias: true)
|
|
291
|
-
super
|
|
292
|
-
@cell = @params[:cell] = Param.new
|
|
293
|
-
end
|
|
294
|
-
|
|
295
|
-
def build(input_shape)
|
|
296
|
-
super
|
|
297
|
-
num_prev_nodes = input_shape[1]
|
|
298
|
-
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes * 4)
|
|
299
|
-
@recurrent_weight.data = Xumo::SFloat.new(@num_nodes, @num_nodes * 4)
|
|
300
|
-
@bias.data = Xumo::SFloat.new(@num_nodes * 4) if @bias
|
|
301
|
-
init_weight_and_bias
|
|
302
|
-
@time_length.times do |t|
|
|
303
|
-
@layers << LSTM_Dense.new(@weight, @recurrent_weight, @bias)
|
|
304
|
-
end
|
|
305
|
-
end
|
|
306
|
-
|
|
307
|
-
def forward(xs)
|
|
308
|
-
@xs_shape = xs.shape
|
|
309
|
-
hs = Xumo::SFloat.zeros(xs.shape[0], @time_length, @num_nodes)
|
|
310
|
-
h = nil
|
|
311
|
-
c = nil
|
|
312
|
-
if @stateful
|
|
313
|
-
h = @hidden.data if @hidden.data
|
|
314
|
-
c = @cell.data if @cell.data
|
|
315
|
-
end
|
|
316
|
-
h ||= Xumo::SFloat.zeros(xs.shape[0], @num_nodes)
|
|
317
|
-
c ||= Xumo::SFloat.zeros(xs.shape[0], @num_nodes)
|
|
318
|
-
xs.shape[1].times do |t|
|
|
319
|
-
x = xs[true, t, false]
|
|
320
|
-
@layers[t].trainable = @trainable
|
|
321
|
-
h, c = @layers[t].forward(x, h, c)
|
|
322
|
-
hs[true, t, false] = h
|
|
323
|
-
end
|
|
324
|
-
@hidden.data = h
|
|
325
|
-
@cell.data = c
|
|
326
|
-
@return_sequences ? hs : h
|
|
327
|
-
end
|
|
328
|
-
|
|
329
|
-
def backward(dh2s)
|
|
330
|
-
unless @return_sequences
|
|
331
|
-
dh = dh2s
|
|
332
|
-
dh2s = Xumo::SFloat.zeros(dh.shape[0], @time_length, dh.shape[1])
|
|
333
|
-
dh2s[true, -1, false] = dh
|
|
334
|
-
end
|
|
335
|
-
dxs = Xumo::SFloat.zeros(@xs_shape)
|
|
336
|
-
dh = 0
|
|
337
|
-
dc = 0
|
|
338
|
-
(0...dh2s.shape[1]).to_a.reverse.each do |t|
|
|
339
|
-
dh2 = dh2s[true, t, false]
|
|
340
|
-
dx, dh, dc = @layers[t].backward(dh2 + dh, dc)
|
|
341
|
-
dxs[true, t, false] = dx
|
|
342
|
-
end
|
|
343
|
-
dxs
|
|
344
|
-
end
|
|
345
|
-
|
|
346
|
-
def reset_state
|
|
347
|
-
super()
|
|
348
|
-
@cell.data = @cell.data.fill(0) if @cell.data
|
|
349
|
-
end
|
|
350
|
-
end
|
|
351
|
-
|
|
352
|
-
|
|
353
|
-
class GRU_Dense
|
|
354
|
-
attr_accessor :trainable
|
|
355
|
-
|
|
356
|
-
def initialize(weight, recurrent_weight, bias)
|
|
357
|
-
@weight = weight
|
|
358
|
-
@recurrent_weight = recurrent_weight
|
|
359
|
-
@bias = bias
|
|
360
|
-
@update_sigmoid = Sigmoid.new
|
|
361
|
-
@reset_sigmoid = Sigmoid.new
|
|
362
|
-
@tanh = Tanh.new
|
|
363
|
-
@trainable = true
|
|
364
|
-
end
|
|
365
|
-
|
|
366
|
-
def forward(x, h)
|
|
367
|
-
@x = x
|
|
368
|
-
@h = h
|
|
369
|
-
num_nodes = h.shape[1]
|
|
370
|
-
@weight_a = @weight.data[true, 0...(num_nodes * 2)]
|
|
371
|
-
@weight2_a = @recurrent_weight.data[true, 0...(num_nodes * 2)]
|
|
372
|
-
a = x.dot(@weight_a) + h.dot(@weight2_a)
|
|
373
|
-
a += @bias.data[0...(num_nodes * 2)] if @bias
|
|
374
|
-
@update = @update_sigmoid.forward(a[true, 0...num_nodes])
|
|
375
|
-
@reset = @reset_sigmoid.forward(a[true, num_nodes..-1])
|
|
376
|
-
|
|
377
|
-
@weight_h = @weight.data[true, (num_nodes * 2)..-1]
|
|
378
|
-
@weight2_h = @recurrent_weight.data[true, (num_nodes * 2)..-1]
|
|
379
|
-
@tanh_h = if @bias
|
|
380
|
-
bias_h = @bias.data[(num_nodes * 2)..-1]
|
|
381
|
-
@tanh.forward(x.dot(@weight_h) + (h * @reset).dot(@weight2_h) + bias_h)
|
|
382
|
-
else
|
|
383
|
-
@tanh.forward(x.dot(@weight_h) + (h * @reset).dot(@weight2_h))
|
|
384
|
-
end
|
|
385
|
-
h2 = (1 - @update) * @tanh_h + @update * h
|
|
386
|
-
h2
|
|
387
|
-
end
|
|
388
|
-
|
|
389
|
-
def backward(dh2)
|
|
390
|
-
dtanh_h = @tanh.backward(dh2 * (1 - @update))
|
|
391
|
-
dh = dh2 * @update
|
|
392
|
-
|
|
393
|
-
if @trainable
|
|
394
|
-
dweight_h = @x.transpose.dot(dtanh_h)
|
|
395
|
-
dweight2_h = (@h * @reset).transpose.dot(dtanh_h)
|
|
396
|
-
dbias_h = dtanh_h.sum(0) if @bias
|
|
397
|
-
end
|
|
398
|
-
dx = dtanh_h.dot(@weight_h.transpose)
|
|
399
|
-
dh += dtanh_h.dot(@weight2_h.transpose) * @reset
|
|
400
|
-
|
|
401
|
-
dreset = @reset_sigmoid.backward(dtanh_h.dot(@weight2_h.transpose) * @h)
|
|
402
|
-
dupdate = @update_sigmoid.backward(dh2 * @h - dh2 * @tanh_h)
|
|
403
|
-
da = Xumo::SFloat.hstack([dupdate, dreset])
|
|
404
|
-
if @trainable
|
|
405
|
-
dweight_a = @x.transpose.dot(da)
|
|
406
|
-
dweight2_a = @h.transpose.dot(da)
|
|
407
|
-
dbias_a = da.sum(0) if @bias
|
|
408
|
-
end
|
|
409
|
-
dx += da.dot(@weight_a.transpose)
|
|
410
|
-
dh += da.dot(@weight2_a.transpose)
|
|
411
|
-
|
|
412
|
-
if @trainable
|
|
413
|
-
@weight.grad += Xumo::SFloat.hstack([dweight_a, dweight_h])
|
|
414
|
-
@recurrent_weight.grad += Xumo::SFloat.hstack([dweight2_a, dweight2_h])
|
|
415
|
-
@bias.grad += Xumo::SFloat.hstack([dbias_a, dbias_h]) if @bias
|
|
416
|
-
end
|
|
417
|
-
[dx, dh]
|
|
418
|
-
end
|
|
419
|
-
end
|
|
420
|
-
|
|
421
|
-
|
|
422
|
-
class GRU < RNN
|
|
423
|
-
def self.from_hash(hash)
|
|
424
|
-
gru = self.new(hash[:num_nodes],
|
|
425
|
-
stateful: hash[:stateful],
|
|
426
|
-
return_sequences: hash[:return_sequences],
|
|
427
|
-
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
|
428
|
-
recurrent_weight_initializer: Utils.from_hash(hash[:recurrent_weight_initializer]),
|
|
429
|
-
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
|
430
|
-
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
|
431
|
-
recurrent_weight_regularizer: Utils.from_hash(hash[:recurrent_weight_regularizer]),
|
|
432
|
-
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
|
433
|
-
use_bias: hash[:use_bias])
|
|
434
|
-
gru
|
|
435
|
-
end
|
|
436
|
-
|
|
437
|
-
def initialize(num_nodes,
|
|
438
|
-
stateful: false,
|
|
439
|
-
return_sequences: true,
|
|
440
|
-
weight_initializer: RandomNormal.new,
|
|
441
|
-
recurrent_weight_initializer: RandomNormal.new,
|
|
442
|
-
bias_initializer: Zeros.new,
|
|
443
|
-
weight_regularizer: nil,
|
|
444
|
-
recurrent_weight_regularizer: nil,
|
|
445
|
-
bias_regularizer: nil,
|
|
446
|
-
use_bias: true)
|
|
447
|
-
super
|
|
448
|
-
end
|
|
449
|
-
|
|
450
|
-
def build(input_shape)
|
|
451
|
-
super
|
|
452
|
-
num_prev_nodes = @input_shape[1]
|
|
453
|
-
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes * 3)
|
|
454
|
-
@recurrent_weight.data = Xumo::SFloat.new(@num_nodes, @num_nodes * 3)
|
|
455
|
-
@bias.data = Xumo::SFloat.new(@num_nodes * 3) if @bias
|
|
456
|
-
init_weight_and_bias
|
|
457
|
-
@time_length.times do |t|
|
|
458
|
-
@layers << GRU_Dense.new(@weight, @recurrent_weight, @bias)
|
|
459
|
-
end
|
|
460
|
-
end
|
|
461
|
-
end
|
|
462
|
-
|
|
463
|
-
end
|
|
464
|
-
end
|
|
1
|
+
module DNN
|
|
2
|
+
module Layers
|
|
3
|
+
|
|
4
|
+
# Super class of all RNN classes.
|
|
5
|
+
class RNN < Connection
|
|
6
|
+
include Initializers
|
|
7
|
+
|
|
8
|
+
# @return [Integer] number of nodes.
|
|
9
|
+
attr_reader :num_nodes
|
|
10
|
+
# @return [Bool] Maintain state between batches.
|
|
11
|
+
attr_reader :stateful
|
|
12
|
+
# @return [Bool] Set the false, only the last of each cell of RNN is left.
|
|
13
|
+
attr_reader :return_sequences
|
|
14
|
+
# @return [DNN::Initializers::Initializer] Recurrent weight initializer.
|
|
15
|
+
attr_reader :recurrent_weight_initializer
|
|
16
|
+
# @return [DNN::Regularizers::Regularizer] Recurrent weight regularization.
|
|
17
|
+
attr_reader :recurrent_weight_regularizer
|
|
18
|
+
|
|
19
|
+
def initialize(num_nodes,
|
|
20
|
+
stateful: false,
|
|
21
|
+
return_sequences: true,
|
|
22
|
+
weight_initializer: RandomNormal.new,
|
|
23
|
+
recurrent_weight_initializer: RandomNormal.new,
|
|
24
|
+
bias_initializer: Zeros.new,
|
|
25
|
+
weight_regularizer: nil,
|
|
26
|
+
recurrent_weight_regularizer: nil,
|
|
27
|
+
bias_regularizer: nil,
|
|
28
|
+
use_bias: true)
|
|
29
|
+
super(weight_initializer: weight_initializer, bias_initializer: bias_initializer,
|
|
30
|
+
weight_regularizer: weight_regularizer, bias_regularizer: bias_regularizer, use_bias: use_bias)
|
|
31
|
+
@num_nodes = num_nodes
|
|
32
|
+
@stateful = stateful
|
|
33
|
+
@return_sequences = return_sequences
|
|
34
|
+
@layers = []
|
|
35
|
+
@hidden = @params[:hidden] = Param.new
|
|
36
|
+
@params[:recurrent_weight] = @recurrent_weight = Param.new(nil, 0)
|
|
37
|
+
@recurrent_weight_initializer = recurrent_weight_initializer
|
|
38
|
+
@recurrent_weight_regularizer = recurrent_weight_regularizer
|
|
39
|
+
end
|
|
40
|
+
|
|
41
|
+
def build(input_shape)
|
|
42
|
+
super
|
|
43
|
+
@time_length = @input_shape[0]
|
|
44
|
+
end
|
|
45
|
+
|
|
46
|
+
def forward(xs)
|
|
47
|
+
@xs_shape = xs.shape
|
|
48
|
+
hs = Xumo::SFloat.zeros(xs.shape[0], @time_length, @num_nodes)
|
|
49
|
+
h = (@stateful && @hidden.data) ? @hidden.data : Xumo::SFloat.zeros(xs.shape[0], @num_nodes)
|
|
50
|
+
xs.shape[1].times do |t|
|
|
51
|
+
x = xs[true, t, false]
|
|
52
|
+
@layers[t].trainable = @trainable
|
|
53
|
+
h = @layers[t].forward(x, h)
|
|
54
|
+
hs[true, t, false] = h
|
|
55
|
+
end
|
|
56
|
+
@hidden.data = h
|
|
57
|
+
@return_sequences ? hs : h
|
|
58
|
+
end
|
|
59
|
+
|
|
60
|
+
def backward(dh2s)
|
|
61
|
+
unless @return_sequences
|
|
62
|
+
dh = dh2s
|
|
63
|
+
dh2s = Xumo::SFloat.zeros(dh.shape[0], @time_length, dh.shape[1])
|
|
64
|
+
dh2s[true, -1, false] = dh
|
|
65
|
+
end
|
|
66
|
+
dxs = Xumo::SFloat.zeros(@xs_shape)
|
|
67
|
+
dh = 0
|
|
68
|
+
(0...dh2s.shape[1]).to_a.reverse.each do |t|
|
|
69
|
+
dh2 = dh2s[true, t, false]
|
|
70
|
+
dx, dh = @layers[t].backward(dh2 + dh)
|
|
71
|
+
dxs[true, t, false] = dx
|
|
72
|
+
end
|
|
73
|
+
dxs
|
|
74
|
+
end
|
|
75
|
+
|
|
76
|
+
def output_shape
|
|
77
|
+
@return_sequences ? [@time_length, @num_nodes] : [@num_nodes]
|
|
78
|
+
end
|
|
79
|
+
|
|
80
|
+
def to_hash(merge_hash = nil)
|
|
81
|
+
hash = {
|
|
82
|
+
num_nodes: @num_nodes,
|
|
83
|
+
stateful: @stateful,
|
|
84
|
+
return_sequences: @return_sequences,
|
|
85
|
+
recurrent_weight_initializer: @recurrent_weight_initializer.to_hash,
|
|
86
|
+
recurrent_weight_regularizer: @recurrent_weight_regularizer&.to_hash,
|
|
87
|
+
}
|
|
88
|
+
hash.merge!(merge_hash) if merge_hash
|
|
89
|
+
super(hash)
|
|
90
|
+
end
|
|
91
|
+
|
|
92
|
+
# Reset the state of RNN.
|
|
93
|
+
def reset_state
|
|
94
|
+
@hidden.data = @hidden.data.fill(0) if @hidden.data
|
|
95
|
+
end
|
|
96
|
+
|
|
97
|
+
def regularizers
|
|
98
|
+
regularizers = []
|
|
99
|
+
regularizers << @weight_regularizer if @weight_regularizer
|
|
100
|
+
regularizers << @recurrent_weight_regularizer if @recurrent_weight_regularizer
|
|
101
|
+
regularizers << @bias_regularizer if @bias_regularizer
|
|
102
|
+
regularizers
|
|
103
|
+
end
|
|
104
|
+
|
|
105
|
+
private def init_weight_and_bias
|
|
106
|
+
super
|
|
107
|
+
@recurrent_weight_initializer.init_param(self, @recurrent_weight)
|
|
108
|
+
@recurrent_weight_regularizer.param = @recurrent_weight if @recurrent_weight_regularizer
|
|
109
|
+
end
|
|
110
|
+
end
|
|
111
|
+
|
|
112
|
+
|
|
113
|
+
class SimpleRNN_Dense
|
|
114
|
+
attr_accessor :trainable
|
|
115
|
+
|
|
116
|
+
def initialize(weight, recurrent_weight, bias, activation)
|
|
117
|
+
@weight = weight
|
|
118
|
+
@recurrent_weight = recurrent_weight
|
|
119
|
+
@bias = bias
|
|
120
|
+
@activation = activation.clone
|
|
121
|
+
@trainable = true
|
|
122
|
+
end
|
|
123
|
+
|
|
124
|
+
def forward(x, h)
|
|
125
|
+
@x = x
|
|
126
|
+
@h = h
|
|
127
|
+
h2 = x.dot(@weight.data) + h.dot(@recurrent_weight.data)
|
|
128
|
+
h2 += @bias.data if @bias
|
|
129
|
+
@activation.forward(h2)
|
|
130
|
+
end
|
|
131
|
+
|
|
132
|
+
def backward(dh2)
|
|
133
|
+
dh2 = @activation.backward(dh2)
|
|
134
|
+
if @trainable
|
|
135
|
+
@weight.grad += @x.transpose.dot(dh2)
|
|
136
|
+
@recurrent_weight.grad += @h.transpose.dot(dh2)
|
|
137
|
+
@bias.grad += dh2.sum(0) if @bias
|
|
138
|
+
end
|
|
139
|
+
dx = dh2.dot(@weight.data.transpose)
|
|
140
|
+
dh = dh2.dot(@recurrent_weight.data.transpose)
|
|
141
|
+
[dx, dh]
|
|
142
|
+
end
|
|
143
|
+
end
|
|
144
|
+
|
|
145
|
+
|
|
146
|
+
class SimpleRNN < RNN
|
|
147
|
+
include Activations
|
|
148
|
+
|
|
149
|
+
attr_reader :activation
|
|
150
|
+
|
|
151
|
+
def self.from_hash(hash)
|
|
152
|
+
simple_rnn = self.new(hash[:num_nodes],
|
|
153
|
+
stateful: hash[:stateful],
|
|
154
|
+
return_sequences: hash[:return_sequences],
|
|
155
|
+
activation: Utils.from_hash(hash[:activation]),
|
|
156
|
+
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
|
157
|
+
recurrent_weight_initializer: Utils.from_hash(hash[:recurrent_weight_initializer]),
|
|
158
|
+
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
|
159
|
+
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
|
160
|
+
recurrent_weight_regularizer: Utils.from_hash(hash[:recurrent_weight_regularizer]),
|
|
161
|
+
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
|
162
|
+
use_bias: hash[:use_bias])
|
|
163
|
+
simple_rnn
|
|
164
|
+
end
|
|
165
|
+
|
|
166
|
+
def initialize(num_nodes,
|
|
167
|
+
stateful: false,
|
|
168
|
+
return_sequences: true,
|
|
169
|
+
activation: Tanh.new,
|
|
170
|
+
weight_initializer: RandomNormal.new,
|
|
171
|
+
recurrent_weight_initializer: RandomNormal.new,
|
|
172
|
+
bias_initializer: Zeros.new,
|
|
173
|
+
weight_regularizer: nil,
|
|
174
|
+
recurrent_weight_regularizer: nil,
|
|
175
|
+
bias_regularizer: nil,
|
|
176
|
+
use_bias: true)
|
|
177
|
+
super(num_nodes,
|
|
178
|
+
stateful: stateful,
|
|
179
|
+
return_sequences: return_sequences,
|
|
180
|
+
weight_initializer: weight_initializer,
|
|
181
|
+
recurrent_weight_initializer: recurrent_weight_initializer,
|
|
182
|
+
bias_initializer: bias_initializer,
|
|
183
|
+
weight_regularizer: weight_regularizer,
|
|
184
|
+
recurrent_weight_regularizer: recurrent_weight_regularizer,
|
|
185
|
+
bias_regularizer: bias_regularizer,
|
|
186
|
+
use_bias: use_bias)
|
|
187
|
+
@activation = activation
|
|
188
|
+
end
|
|
189
|
+
|
|
190
|
+
def build(input_shape)
|
|
191
|
+
super
|
|
192
|
+
num_prev_nodes = input_shape[1]
|
|
193
|
+
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes)
|
|
194
|
+
@recurrent_weight.data = Xumo::SFloat.new(@num_nodes, @num_nodes)
|
|
195
|
+
@bias.data = Xumo::SFloat.new(@num_nodes) if @bias
|
|
196
|
+
init_weight_and_bias
|
|
197
|
+
@time_length.times do |t|
|
|
198
|
+
@layers << SimpleRNN_Dense.new(@weight, @recurrent_weight, @bias, @activation)
|
|
199
|
+
end
|
|
200
|
+
end
|
|
201
|
+
|
|
202
|
+
def to_hash
|
|
203
|
+
super({activation: @activation.to_hash})
|
|
204
|
+
end
|
|
205
|
+
end
|
|
206
|
+
|
|
207
|
+
|
|
208
|
+
class LSTM_Dense
|
|
209
|
+
attr_accessor :trainable
|
|
210
|
+
|
|
211
|
+
def initialize(weight, recurrent_weight, bias)
|
|
212
|
+
@weight = weight
|
|
213
|
+
@recurrent_weight = recurrent_weight
|
|
214
|
+
@bias = bias
|
|
215
|
+
@tanh = Tanh.new
|
|
216
|
+
@g_tanh = Tanh.new
|
|
217
|
+
@forget_sigmoid = Sigmoid.new
|
|
218
|
+
@in_sigmoid = Sigmoid.new
|
|
219
|
+
@out_sigmoid = Sigmoid.new
|
|
220
|
+
@trainable = true
|
|
221
|
+
end
|
|
222
|
+
|
|
223
|
+
def forward(x, h, c)
|
|
224
|
+
@x = x
|
|
225
|
+
@h = h
|
|
226
|
+
@c = c
|
|
227
|
+
num_nodes = h.shape[1]
|
|
228
|
+
a = x.dot(@weight.data) + h.dot(@recurrent_weight.data)
|
|
229
|
+
a += @bias.data if @bias
|
|
230
|
+
|
|
231
|
+
@forget = @forget_sigmoid.forward(a[true, 0...num_nodes])
|
|
232
|
+
@g = @g_tanh.forward(a[true, num_nodes...(num_nodes * 2)])
|
|
233
|
+
@in = @in_sigmoid.forward(a[true, (num_nodes * 2)...(num_nodes * 3)])
|
|
234
|
+
@out = @out_sigmoid.forward(a[true, (num_nodes * 3)..-1])
|
|
235
|
+
|
|
236
|
+
c2 = @forget * c + @g * @in
|
|
237
|
+
@tanh_c2 = @tanh.forward(c2)
|
|
238
|
+
h2 = @out * @tanh_c2
|
|
239
|
+
[h2, c2]
|
|
240
|
+
end
|
|
241
|
+
|
|
242
|
+
def backward(dh2, dc2)
|
|
243
|
+
dh2_tmp = @tanh_c2 * dh2
|
|
244
|
+
dc2_tmp = @tanh.backward(@out * dh2) + dc2
|
|
245
|
+
|
|
246
|
+
dout = @out_sigmoid.backward(dh2_tmp)
|
|
247
|
+
din = @in_sigmoid.backward(dc2_tmp * @g)
|
|
248
|
+
dg = @g_tanh.backward(dc2_tmp * @in)
|
|
249
|
+
dforget = @forget_sigmoid.backward(dc2_tmp * @c)
|
|
250
|
+
|
|
251
|
+
da = Xumo::SFloat.hstack([dforget, dg, din, dout])
|
|
252
|
+
|
|
253
|
+
if @trainable
|
|
254
|
+
@weight.grad += @x.transpose.dot(da)
|
|
255
|
+
@recurrent_weight.grad += @h.transpose.dot(da)
|
|
256
|
+
@bias.grad += da.sum(0) if @bias
|
|
257
|
+
end
|
|
258
|
+
dx = da.dot(@weight.data.transpose)
|
|
259
|
+
dh = da.dot(@recurrent_weight.data.transpose)
|
|
260
|
+
dc = dc2_tmp * @forget
|
|
261
|
+
[dx, dh, dc]
|
|
262
|
+
end
|
|
263
|
+
end
|
|
264
|
+
|
|
265
|
+
|
|
266
|
+
class LSTM < RNN
|
|
267
|
+
def self.from_hash(hash)
|
|
268
|
+
lstm = self.new(hash[:num_nodes],
|
|
269
|
+
stateful: hash[:stateful],
|
|
270
|
+
return_sequences: hash[:return_sequences],
|
|
271
|
+
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
|
272
|
+
recurrent_weight_initializer: Utils.from_hash(hash[:recurrent_weight_initializer]),
|
|
273
|
+
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
|
274
|
+
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
|
275
|
+
recurrent_weight_regularizer: Utils.from_hash(hash[:recurrent_weight_regularizer]),
|
|
276
|
+
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
|
277
|
+
use_bias: hash[:use_bias])
|
|
278
|
+
lstm
|
|
279
|
+
end
|
|
280
|
+
|
|
281
|
+
def initialize(num_nodes,
|
|
282
|
+
stateful: false,
|
|
283
|
+
return_sequences: true,
|
|
284
|
+
weight_initializer: RandomNormal.new,
|
|
285
|
+
recurrent_weight_initializer: RandomNormal.new,
|
|
286
|
+
bias_initializer: Zeros.new,
|
|
287
|
+
weight_regularizer: nil,
|
|
288
|
+
recurrent_weight_regularizer: nil,
|
|
289
|
+
bias_regularizer: nil,
|
|
290
|
+
use_bias: true)
|
|
291
|
+
super
|
|
292
|
+
@cell = @params[:cell] = Param.new
|
|
293
|
+
end
|
|
294
|
+
|
|
295
|
+
def build(input_shape)
|
|
296
|
+
super
|
|
297
|
+
num_prev_nodes = input_shape[1]
|
|
298
|
+
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes * 4)
|
|
299
|
+
@recurrent_weight.data = Xumo::SFloat.new(@num_nodes, @num_nodes * 4)
|
|
300
|
+
@bias.data = Xumo::SFloat.new(@num_nodes * 4) if @bias
|
|
301
|
+
init_weight_and_bias
|
|
302
|
+
@time_length.times do |t|
|
|
303
|
+
@layers << LSTM_Dense.new(@weight, @recurrent_weight, @bias)
|
|
304
|
+
end
|
|
305
|
+
end
|
|
306
|
+
|
|
307
|
+
def forward(xs)
|
|
308
|
+
@xs_shape = xs.shape
|
|
309
|
+
hs = Xumo::SFloat.zeros(xs.shape[0], @time_length, @num_nodes)
|
|
310
|
+
h = nil
|
|
311
|
+
c = nil
|
|
312
|
+
if @stateful
|
|
313
|
+
h = @hidden.data if @hidden.data
|
|
314
|
+
c = @cell.data if @cell.data
|
|
315
|
+
end
|
|
316
|
+
h ||= Xumo::SFloat.zeros(xs.shape[0], @num_nodes)
|
|
317
|
+
c ||= Xumo::SFloat.zeros(xs.shape[0], @num_nodes)
|
|
318
|
+
xs.shape[1].times do |t|
|
|
319
|
+
x = xs[true, t, false]
|
|
320
|
+
@layers[t].trainable = @trainable
|
|
321
|
+
h, c = @layers[t].forward(x, h, c)
|
|
322
|
+
hs[true, t, false] = h
|
|
323
|
+
end
|
|
324
|
+
@hidden.data = h
|
|
325
|
+
@cell.data = c
|
|
326
|
+
@return_sequences ? hs : h
|
|
327
|
+
end
|
|
328
|
+
|
|
329
|
+
def backward(dh2s)
|
|
330
|
+
unless @return_sequences
|
|
331
|
+
dh = dh2s
|
|
332
|
+
dh2s = Xumo::SFloat.zeros(dh.shape[0], @time_length, dh.shape[1])
|
|
333
|
+
dh2s[true, -1, false] = dh
|
|
334
|
+
end
|
|
335
|
+
dxs = Xumo::SFloat.zeros(@xs_shape)
|
|
336
|
+
dh = 0
|
|
337
|
+
dc = 0
|
|
338
|
+
(0...dh2s.shape[1]).to_a.reverse.each do |t|
|
|
339
|
+
dh2 = dh2s[true, t, false]
|
|
340
|
+
dx, dh, dc = @layers[t].backward(dh2 + dh, dc)
|
|
341
|
+
dxs[true, t, false] = dx
|
|
342
|
+
end
|
|
343
|
+
dxs
|
|
344
|
+
end
|
|
345
|
+
|
|
346
|
+
def reset_state
|
|
347
|
+
super()
|
|
348
|
+
@cell.data = @cell.data.fill(0) if @cell.data
|
|
349
|
+
end
|
|
350
|
+
end
|
|
351
|
+
|
|
352
|
+
|
|
353
|
+
class GRU_Dense
|
|
354
|
+
attr_accessor :trainable
|
|
355
|
+
|
|
356
|
+
def initialize(weight, recurrent_weight, bias)
|
|
357
|
+
@weight = weight
|
|
358
|
+
@recurrent_weight = recurrent_weight
|
|
359
|
+
@bias = bias
|
|
360
|
+
@update_sigmoid = Sigmoid.new
|
|
361
|
+
@reset_sigmoid = Sigmoid.new
|
|
362
|
+
@tanh = Tanh.new
|
|
363
|
+
@trainable = true
|
|
364
|
+
end
|
|
365
|
+
|
|
366
|
+
def forward(x, h)
|
|
367
|
+
@x = x
|
|
368
|
+
@h = h
|
|
369
|
+
num_nodes = h.shape[1]
|
|
370
|
+
@weight_a = @weight.data[true, 0...(num_nodes * 2)]
|
|
371
|
+
@weight2_a = @recurrent_weight.data[true, 0...(num_nodes * 2)]
|
|
372
|
+
a = x.dot(@weight_a) + h.dot(@weight2_a)
|
|
373
|
+
a += @bias.data[0...(num_nodes * 2)] if @bias
|
|
374
|
+
@update = @update_sigmoid.forward(a[true, 0...num_nodes])
|
|
375
|
+
@reset = @reset_sigmoid.forward(a[true, num_nodes..-1])
|
|
376
|
+
|
|
377
|
+
@weight_h = @weight.data[true, (num_nodes * 2)..-1]
|
|
378
|
+
@weight2_h = @recurrent_weight.data[true, (num_nodes * 2)..-1]
|
|
379
|
+
@tanh_h = if @bias
|
|
380
|
+
bias_h = @bias.data[(num_nodes * 2)..-1]
|
|
381
|
+
@tanh.forward(x.dot(@weight_h) + (h * @reset).dot(@weight2_h) + bias_h)
|
|
382
|
+
else
|
|
383
|
+
@tanh.forward(x.dot(@weight_h) + (h * @reset).dot(@weight2_h))
|
|
384
|
+
end
|
|
385
|
+
h2 = (1 - @update) * @tanh_h + @update * h
|
|
386
|
+
h2
|
|
387
|
+
end
|
|
388
|
+
|
|
389
|
+
def backward(dh2)
|
|
390
|
+
dtanh_h = @tanh.backward(dh2 * (1 - @update))
|
|
391
|
+
dh = dh2 * @update
|
|
392
|
+
|
|
393
|
+
if @trainable
|
|
394
|
+
dweight_h = @x.transpose.dot(dtanh_h)
|
|
395
|
+
dweight2_h = (@h * @reset).transpose.dot(dtanh_h)
|
|
396
|
+
dbias_h = dtanh_h.sum(0) if @bias
|
|
397
|
+
end
|
|
398
|
+
dx = dtanh_h.dot(@weight_h.transpose)
|
|
399
|
+
dh += dtanh_h.dot(@weight2_h.transpose) * @reset
|
|
400
|
+
|
|
401
|
+
dreset = @reset_sigmoid.backward(dtanh_h.dot(@weight2_h.transpose) * @h)
|
|
402
|
+
dupdate = @update_sigmoid.backward(dh2 * @h - dh2 * @tanh_h)
|
|
403
|
+
da = Xumo::SFloat.hstack([dupdate, dreset])
|
|
404
|
+
if @trainable
|
|
405
|
+
dweight_a = @x.transpose.dot(da)
|
|
406
|
+
dweight2_a = @h.transpose.dot(da)
|
|
407
|
+
dbias_a = da.sum(0) if @bias
|
|
408
|
+
end
|
|
409
|
+
dx += da.dot(@weight_a.transpose)
|
|
410
|
+
dh += da.dot(@weight2_a.transpose)
|
|
411
|
+
|
|
412
|
+
if @trainable
|
|
413
|
+
@weight.grad += Xumo::SFloat.hstack([dweight_a, dweight_h])
|
|
414
|
+
@recurrent_weight.grad += Xumo::SFloat.hstack([dweight2_a, dweight2_h])
|
|
415
|
+
@bias.grad += Xumo::SFloat.hstack([dbias_a, dbias_h]) if @bias
|
|
416
|
+
end
|
|
417
|
+
[dx, dh]
|
|
418
|
+
end
|
|
419
|
+
end
|
|
420
|
+
|
|
421
|
+
|
|
422
|
+
class GRU < RNN
|
|
423
|
+
def self.from_hash(hash)
|
|
424
|
+
gru = self.new(hash[:num_nodes],
|
|
425
|
+
stateful: hash[:stateful],
|
|
426
|
+
return_sequences: hash[:return_sequences],
|
|
427
|
+
weight_initializer: Utils.from_hash(hash[:weight_initializer]),
|
|
428
|
+
recurrent_weight_initializer: Utils.from_hash(hash[:recurrent_weight_initializer]),
|
|
429
|
+
bias_initializer: Utils.from_hash(hash[:bias_initializer]),
|
|
430
|
+
weight_regularizer: Utils.from_hash(hash[:weight_regularizer]),
|
|
431
|
+
recurrent_weight_regularizer: Utils.from_hash(hash[:recurrent_weight_regularizer]),
|
|
432
|
+
bias_regularizer: Utils.from_hash(hash[:bias_regularizer]),
|
|
433
|
+
use_bias: hash[:use_bias])
|
|
434
|
+
gru
|
|
435
|
+
end
|
|
436
|
+
|
|
437
|
+
def initialize(num_nodes,
|
|
438
|
+
stateful: false,
|
|
439
|
+
return_sequences: true,
|
|
440
|
+
weight_initializer: RandomNormal.new,
|
|
441
|
+
recurrent_weight_initializer: RandomNormal.new,
|
|
442
|
+
bias_initializer: Zeros.new,
|
|
443
|
+
weight_regularizer: nil,
|
|
444
|
+
recurrent_weight_regularizer: nil,
|
|
445
|
+
bias_regularizer: nil,
|
|
446
|
+
use_bias: true)
|
|
447
|
+
super
|
|
448
|
+
end
|
|
449
|
+
|
|
450
|
+
def build(input_shape)
|
|
451
|
+
super
|
|
452
|
+
num_prev_nodes = @input_shape[1]
|
|
453
|
+
@weight.data = Xumo::SFloat.new(num_prev_nodes, @num_nodes * 3)
|
|
454
|
+
@recurrent_weight.data = Xumo::SFloat.new(@num_nodes, @num_nodes * 3)
|
|
455
|
+
@bias.data = Xumo::SFloat.new(@num_nodes * 3) if @bias
|
|
456
|
+
init_weight_and_bias
|
|
457
|
+
@time_length.times do |t|
|
|
458
|
+
@layers << GRU_Dense.new(@weight, @recurrent_weight, @bias)
|
|
459
|
+
end
|
|
460
|
+
end
|
|
461
|
+
end
|
|
462
|
+
|
|
463
|
+
end
|
|
464
|
+
end
|