Class: Daimond::Tensor

Inherits:
Object
  • Object
show all
Defined in:
lib/daimond/tensor.rb

Instance Attribute Summary collapse

Class Method Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(data, prev: [], op: nil, label: nil) ⇒ Tensor

Returns a new instance of Tensor.



8
9
10
11
12
13
14
15
 
# File 'lib/daimond/tensor.rb', line 8

def initialize(data, prev: [], op: nil, label: nil)
  @data = data.is_a?(Numo::DFloat) ? data : Numo::DFloat[*data]
  @grad = Numo::DFloat.zeros(*@data.shape)
  @prev = prev
  @op = op
  @label = label
  @_backward = lambda {}  # По умолчанию пустая функция
end

Instance Attribute Details

#_backward ⇒ Object

Returns the value of attribute _backward.



6
7
8
 
# File 'lib/daimond/tensor.rb', line 6

def _backward
  @_backward
end

#data ⇒ Object

Returns the value of attribute data.



6
7
8
 
# File 'lib/daimond/tensor.rb', line 6

def data
  @data
end

#grad ⇒ Object

Returns the value of attribute grad.



6
7
8
 
# File 'lib/daimond/tensor.rb', line 6

def grad
  @grad
end

#label ⇒ Object

Returns the value of attribute label.



6
7
8
 
# File 'lib/daimond/tensor.rb', line 6

def label
  @label
end

#op ⇒ Object

Returns the value of attribute op.



6
7
8
 
# File 'lib/daimond/tensor.rb', line 6

def op
  @op
end

#prev ⇒ Object

Returns the value of attribute prev.



6
7
8
 
# File 'lib/daimond/tensor.rb', line 6

def prev
  @prev
end

Class Method Details

.randn(*shape) ⇒ Object 



214
215
216
217
 
# File 'lib/daimond/tensor.rb', line 214

def self.randn(*shape)
  data = Numo::DFloat.new(*shape).rand_norm
  Tensor.new(data)
end

.zeros(*shape) ⇒ Object 



219
220
221
 
# File 'lib/daimond/tensor.rb', line 219

def self.zeros(*shape)
  Tensor.new(Numo::DFloat.zeros(*shape))
end

Instance Method Details

#*(other) ⇒ Object

Поэлементное



68
69
70
71
72
73
74
75
76
77
78
79
80
81
 
# File 'lib/daimond/tensor.rb', line 68

def *(other)  # Поэлементное
  other = other.is_a?(Tensor) ? other : Tensor.new(other)
  left = self
  right = other
  out = Tensor.new(@data * other.data, prev: [self, other], op: '*')

  out._backward = lambda do
    grad = out.grad
    left.grad += right.data * grad
    right.grad += left.data * grad
  end

  out
end

#+(other) ⇒ Object 



21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
 
# File 'lib/daimond/tensor.rb', line 21

def +(other)
  other = other.is_a?(Tensor) ? other : Tensor.new(other)
  left = self
  right = other
  out = Tensor.new(@data + other.data, prev: [self, other], op: '+')

  out._backward = lambda do
    grad = out.grad

    # Для left (может быть broadcasted, но здесь обычно нет)
    if grad.shape.length > left.shape.length
      left.grad += grad.sum(axis: 0)
    else
      left.grad += grad
    end

    # Для right (bias) — суммируем по batch
    if grad.shape.length > right.shape.length
      right.grad += grad.sum(axis: 0)
    else
      right.grad += grad
    end
  end

  out
end

#-(other) ⇒ Object 



48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
 
# File 'lib/daimond/tensor.rb', line 48

def -(other)
  other = other.is_a?(Tensor) ? other : Tensor.new(other)
  left = self
  right = other
  out = Tensor.new(@data - other.data, prev: [self, other], op: '-')

  out._backward = lambda do
    grad = out.grad
    left.grad += grad

    if grad.shape.length > right.shape.length
      right.grad -= grad.sum(axis: 0)
    else
      right.grad -= grad
    end
  end

  out
end

#backward! ⇒ Object 



186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
 
# File 'lib/daimond/tensor.rb', line 186

def backward!
  # Топологическая сортировка
  topo = []
  visited = []

  build_topo = lambda do |v|
    return if visited.include?(v)
    visited << v
    v.prev.each { |child| build_topo.call(child) }
    topo << v
  end

  build_topo.call(self)

  self.grad = Numo::DFloat[1.0]  # seed gradient

  # Идём в обратном порядке (от loss к входам)
  topo.reverse.each do |node|
    node._backward.call
  end
end

#dot(other) ⇒ Object 



83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
 
# File 'lib/daimond/tensor.rb', line 83

def dot(other)
  other = other.is_a?(Tensor) ? other : Tensor.new(other)

  inner_dim = @data.shape[1]
  out_dim = other.data.shape[1]

  # Глобальный счетчик для отладки
  $rust_count ||= 0
  $ruby_count ||= 0

  rust_available = defined?(Daimond::Rust)
  condition = (inner_dim > 100 || out_dim > 50)

  if rust_available && condition
    begin
      rust_a = Daimond::Rust::Tensor.from_array(@data.to_a)
      rust_b = Daimond::Rust::Tensor.from_array(other.data.to_a)
      rust_result = rust_a.matmul(rust_b)

      out = Tensor.new(Numo::DFloat[*rust_result.to_a], prev: [self, other], op: 'dot')
      out._backward = lambda do
        grad = out.grad
        self.grad += grad.dot(other.data.transpose)
        other.grad += self.data.transpose.dot(grad)
      end

      $rust_count += 1
      out
    rescue => e
      $ruby_count += 1

      # Fallback
      out = Tensor.new(@data.dot(other.data), prev: [self, other], op: 'dot')
      out._backward = lambda do
        grad = out.grad
        self.grad += grad.dot(other.data.transpose)
        other.grad += self.data.transpose.dot(grad)
      end
      out
    end
  else
    $ruby_count += 1
    # Ruby version
    out = Tensor.new(@data.dot(other.data), prev: [self, other], op: 'dot')
    out._backward = lambda do
      grad = out.grad
      self.grad += grad.dot(other.data.transpose)
      other.grad += self.data.transpose.dot(grad)
    end
    out
  end
end

#mean ⇒ Object 



174
175
176
177
178
179
180
181
182
183
184
 
# File 'lib/daimond/tensor.rb', line 174

def mean
  input = self
  out = Tensor.new(Numo::DFloat[@data.mean], prev: [self], op: 'mean')
  n = @data.size

  out._backward = lambda do
    input.grad += Numo::DFloat.ones(*input.shape) * (out.grad[0] / n)
  end

  out
end

#relu ⇒ Object 



136
137
138
139
140
141
142
143
144
145
146
147
148
 
# File 'lib/daimond/tensor.rb', line 136

def relu
  out_data = @data.map { |x| x > 0 ? x : 0.0 }
  out = Tensor.new(out_data, prev: [self], op: 'relu')
  input = self

  out._backward = lambda do
    grad = out.grad
    mask = input.data.map { |x| x > 0 ? 1.0 : 0.0 }
    input.grad += mask * grad
  end

  out
end

#shape ⇒ Object 



17
18
19
 
# File 'lib/daimond/tensor.rb', line 17

def shape
  @data.shape
end

#sigmoid ⇒ Object 



150
151
152
153
154
155
156
157
158
159
160
161
 
# File 'lib/daimond/tensor.rb', line 150

def sigmoid
  input = self
  s = @data.map { |x| 1.0 / (1.0 + Math.exp(-x)) }
  out = Tensor.new(s, prev: [self], op: 'sigmoid')

  out._backward = lambda do
    grad = out.grad
    input.grad += (out.data * (1.0 - out.data)) * grad
  end

  out
end

#softmax ⇒ Object 



223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
 
# File 'lib/daimond/tensor.rb', line 223

def softmax
  input = self
  # Численная стабильность: вычитаем max по каждой строке
  max_val = @data.max(axis: 1).reshape(@data.shape[0], 1)
  exp_data = Numo::NMath.exp(@data - max_val)
  sum_exp = exp_data.sum(axis: 1).reshape(@data.shape[0], 1)
  out_data = exp_data / sum_exp

  out = Tensor.new(out_data, prev: [self], op: 'softmax')

  # Backward упрощенный (для связки с CrossEntropy)
  out._backward = lambda do
    input.grad += out.grad
  end

  out
end

#sum ⇒ Object 



163
164
165
166
167
168
169
170
171
172
 
# File 'lib/daimond/tensor.rb', line 163

def sum
  input = self
  out = Tensor.new(Numo::DFloat[@data.sum], prev: [self], op: 'sum')

  out._backward = lambda do
    input.grad += Numo::DFloat.ones(*input.shape) * out.grad[0]
  end

  out
end

#to_s ⇒ Object Also known as: inspect 



208
209
210
 
# File 'lib/daimond/tensor.rb', line 208

def to_s
  "Tensor(shape=#{shape}, mean=#{@data.mean.round(4)})"
end