Method: TensorStream::MathGradients._compute_derivative

Defined in:
lib/tensor_stream/math_gradients.rb

._compute_derivative(node, grad) ⇒ Object

TODO: refactor and implement registerGradient



52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
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
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
 
# File 'lib/tensor_stream/math_gradients.rb', line 52

def self._compute_derivative(node, grad)
  node.graph.name_scope("#{node.name}_grad") do
    x = node.inputs[0] if node.inputs[0]
    y = node.inputs[1] if node.inputs[1]
    z = node.inputs[2] if node.inputs[2]

    case node.operation
    when :add_n
      return [grad] * node.inputs.size
    when :asin
      ts.control_dependencies([grad]) do
        x2 = ts.square(x)
        one = ts.constant(1, dtype: grad.data_type)
        den = ts.sqrt(ts.subtract(one, x2))
        inv = ts.reciprocal(den)
        grad * inv
      end
    when :acos
      ts.control_dependencies([grad]) do
        x2 = ts.square(x)
        one = ts.constant(1, dtype: grad.data_type)
        den = ts.sqrt(ts.subtract(one, x2))
        inv = ts.reciprocal(den)
        -grad * inv
      end
    when :atan
      ts.control_dependencies([grad]) do
        x2 = ts.square(x)
        one = ts.constant(1, dtype: grad.data_type)
        inv = ts.reciprocal(ts.add(one, x2))
        grad * inv
      end
    when :squared_difference
      sx = i_op(:shape, x)
      sy = i_op(:shape, y)
      rx, ry = _broadcast_gradient_args(sx, sy)

      x_grad = ts.mul(2.0, grad) * (x - y)

      [ts.reshape(ts.reduce_sum(x_grad, rx), sx),
       ts.reshape(-ts.reduce_sum(x_grad, ry), sy),]
    when :abs
      grad * ts.sign(x)
    when :exp
      grad * node
    when :identity, :print
      grad
    when :tile
      input_shape = ts.shape(x)
      split_shape = ts.reshape(ts.transpose(ts.stack([y, input_shape])), [-1])
      axes = ts.range(0, ts.size(split_shape), 2)
      input_grad = ts.reduce_sum(ts.reshape(grad, split_shape), axes)

      [input_grad, nil]
    when :reciprocal
      -grad * (ts.constant(1, dtype: x.dtype) / x**2)
    when :sqrt
      ts.constant(1, dtype: x.dtype) / (ts.constant(2, dtype: x.dtype) * ts.sqrt(x)) * grad
    when :stop_gradient
      ts.zeros_like(grad)
    when :square
      y = ts.constant(2.0, dtype: x.dtype)
      ts.multiply(grad, ts.multiply(x, y))
    when :where
      x_mask = i_op(:where, x, i_op(:ones_like, y), i_op(:zeros_like, z))
      y_mask = i_op(:where, x, i_op(:zeros_like, y), i_op(:ones_like, z))
      [nil, x_mask * grad, y_mask * grad]
    when :mean
      sum_grad = _sum_grad(x, y, grad)[0]
      input_shape = ts.shape(x)
      output_shape = ts.shape(node)
      factor = _safe_shape_div(ts.reduce_prod(input_shape), ts.reduce_prod(output_shape))
      [ts.div(sum_grad, ts.cast(factor, sum_grad.data_type)), nil]
    when :log1p
      grad * ts.reciprocal(i_cons(1, dtype: grad.data_type) + x)
    when :sigmoid_grad
      gb = grad * y
      [gb - 2.0 * gb * x, i_op(:sigmoid_grad, x, grad)]
    when :softmax
      i_op(:softmax_grad, x, grad)
    when :softmax_cross_entropy_with_logits_v2
      output = node
      logits = node.inputs[0]
      [_broadcast_mul(grad, output[1]), -ts.nn.log_softmax(logits)]
    when :sparse_softmax_cross_entropy_with_logits
      output = node
      [_broadcast_mul(grad, output[1]), nil]
    when :zeros_like
      # non differentiable
      nil
    when :transpose
      return [ts.transpose(grad, ts.invert_permutation(y)), nil]
    when :index
      # hack!! not sure how to fix this yet
      return grad if %i[softmax_cross_entropy_with_logits_v2 sparse_softmax_cross_entropy_with_logits].include?(node.inputs[0].operation)

      if node.inputs[0].shape.known? && node.inputs[1].const_value
        multiplier = node.inputs[0].shape.shape[0]
        filler = ts.zeros_like(grad)

        res = Array.new(multiplier) { |index|
          index == node.inputs[1].const_value ? grad : filler
        }
        [res]
      end
    when :squeeze
      _reshape_to_input(node, grad)
    when :concat
      _concat_grad_helper(node, grad, 1, node.inputs.size, 0)
    when :stack
      res = ts.unstack(grad, num: node.inputs.size, axis: node.options[:axis])
      Array.new(node.inputs.size) { |i| res[i] }
    when :unstack
      ts.stack(grad, axis: node.options[:axis])
    when :conv2d
      _Conv2DGrad(node, grad)
    when :flow_dynamic_stitch
      num_values = node.inputs.size / 2
      indices_grad = [nil] * num_values

      inputs = (0...num_values).map { |i| _int32(node, node.inputs[i]) }

      values_grad = inputs.map { |inp| TensorStream.gather(grad, inp) }
      indices_grad + values_grad
    when :gather
      [_op(:gather_grad, grad, node.inputs[1], TensorStream.shape(node.inputs[0])), nil]
    else
      TensorStream::OpMaker.gradient_op(self, node, grad)
    end
  end
end