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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
|
# 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 :add
return [grad, grad] if shapes_fully_specified_and_equal(x, y)
sx = ts.shape(x, name: 'add/shape_x')
sy = ts.shape(y, name: 'add/shape_y')
rx, ry = _broadcast_gradient_args(sx, sy)
[ts.reshape(ts.reduce_sum(grad, rx, name: 'add/reduce_sum_x'), sx),
ts.reshape(ts.reduce_sum(grad, ry, name: 'add/reduce_sum_y'), sy)]
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 :fill
[nil, ts.reduce_sum(grad)]
when :sub
return [grad, -grad] if shapes_fully_specified_and_equal(x, y)
sx = ts.shape(x, name: 'sub/shape_x')
sy = ts.shape(y, name: 'sub/shape_y')
rx, ry = _broadcast_gradient_args(sx, sy)
[ts.reshape(ts.reduce_sum(grad, rx, name: 'add/reduce_sub_x'), sx),
-ts.reshape(ts.reduce_sum(grad, ry, name: 'add/reduce_sub_y'), sy)]
when :mul
sx = ts.shape(x)
sy = ts.shape(y)
rx, ry = _broadcast_gradient_args(sx, sy)
[ts.reshape(ts.reduce_sum(ts.mul(grad, y), rx), sx),
ts.reshape(ts.reduce_sum(ts.mul(x, grad), ry), sy)]
when :div
sx = i_op(:shape, x)
sy = i_op(:shape, y)
rx, ry = _broadcast_gradient_args(sx, sy)
[ts.reshape(ts.reduce_sum(ts.div(grad, y), rx), sx),
ts.reshape(ts.reduce_sum(grad * ts.div(ts.div(-x, y), y), ry), sy)]
when :mod
sx = ts.shape(x)
sy = ts.shape(y)
rx, ry = _broadcast_gradient_args(sx, sy)
floor_xy = ts.floor_div(x, y)
gx = ts.reshape(ts.reduce_sum(grad, rx), sx)
gy = ts.reshape(ts.reduce_sum(grad * ts.negative(floor_xy), ry), sy)
[gx, gy]
when :prod
input_shape = ts.shape(x)
y = ts.range(0, ts.rank(x)) if y.nil?
reduction_indices = ts.reshape(y, [-1])
output_shape_kept_dims = ts.reduced_shape(input_shape, y)
tile_scaling = _safe_shape_div(input_shape, output_shape_kept_dims)
grad = ts.reshape(grad, output_shape_kept_dims)
grad = ts.tile(grad, tile_scaling)
perm, reduced_num, other_num = ts.device("/cpu:0") do
rank = ts.rank(x)
reduction_indices = (reduction_indices + rank) % rank
reduced = ts.cast(reduction_indices, :int32)
idx = ts.range(0, rank)
other, = ts.setdiff1d(idx, reduced)
[ts.concat([reduced, other], 0),
ts.reduce_prod(ts.gather(input_shape, reduced)),
ts.reduce_prod(ts.gather(input_shape, other))]
end
permuted = ts.transpose(x, perm)
permuted_shape = ts.shape(permuted)
reshaped = ts.reshape(permuted, [reduced_num, other_num])
left = ts.cumprod(reshaped, axis: 0, exclusive: true)
right = ts.cumprod(reshaped, axis: 0, exclusive: true, reverse: true)
y = ts.reshape(left * right, permuted_shape)
out = grad * ts.transpose(y, ts.invert_permutation(perm))
[ts.reshape(out, input_shape, name: 'prod'), nil]
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 :mat_mul
t_a = node.options[:transpose_a]
t_b = node.options[:transpose_b]
if !t_a && !t_b
grad_a = ts.matmul(grad, y, transpose_b: true)
grad_b = ts.matmul(x, grad, transpose_a: true)
elsif !ta && tb
grad_a = ts.matmul(grad, y)
grad_b = ts.matmul(grad, x, transpose_a: true)
elsif t_a && !t_b
grad_a = ts.matmul(y, grad, transpose_b: true)
grad_b = ts.matmul(x, grad)
elsif t_a && t_b
grad_a = ts.matmul(y, grad, transpose_a: true, transpose_b: true)
grad_b = ts.matmul(grad, x, transpose_a: true, transpose_b: true)
end
[grad_a, grad_b]
when :sin
grad * ts.cos(x)
when :tanh
grad * i_op(:tanh_grad, x)
when :pow
z = node
sx = ts.shape(x)
sy = ts.shape(y)
rx, ry = _broadcast_gradient_args(sx, sy)
gx = ts.reduce_sum(grad * y * ts.pow(x, y - 1), rx)
log_x = ts.where(x > 0, ts.log(x), ts.zeros_like(x))
gy = ts.reduce_sum(grad * z * log_x, ry)
[gx, gy]
when :abs
grad * ts.sign(x)
when :log
grad * ts.reciprocal(x)
when :cos
-grad * ts.sin(x)
when :max
_min_or_max_grad(node.inputs, grad, ->(a, b) { ts.greater_equal(a, b) })
when :min
_min_or_max_grad(node.inputs, grad, ->(a, b) { ts.less_equal(a, b) })
when :tan
secx = ts.reciprocal(ts.cos(x))
secx2 = ts.square(secx)
grad * secx2
when :negate
-grad
when :exp
grad * node
when :identity, :print
grad
when :sign
ts.zeros(ts.shape(x), dtype: x.data_type)
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 :sum
_sum_grad(x, y, grad)
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 :case
n_preds = node.inputs.size - 2
case_grads = Array.new(n_preds) do |index|
i_op(:case_grad, index, node.inputs[0], node.inputs[2 + index], grad)
end
[nil, i_op(:case_grad, -1, node.inputs[0], node.inputs[1], grad)] + case_grads
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
i_op(:sigmoid_grad, x, grad)
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 :floor, :ceil, :round
nil
when :zeros_like
nil
when :argmin, :argmax, :floor_div
[nil, nil]
when :transpose
return [ts.transpose(grad, ts.invert_permutation(y)), nil]
when :index
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) do |index|
index == node.inputs[1].const_value ? grad : filler
end
[res]
end
when :squeeze
_reshape_to_input(node, grad)
when :expand_dims
[_reshape_to_input(node, grad), nil]
when :concat
_concat_grad_helper(node, grad, 1, node.inputs.size, 0)
when :reshape
[ts.reshape(grad, ts.shape(node.inputs[0])), nil]
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 :cast
t = %i[float16 float32 float64]
src_type = node.inputs[0].data_type
dst_type = grad.data_type
if t.key?(src_type) && t.key?(dst_type)
ts.cast(grad, src_type)
end
nil
else
raise "no derivative op for #{node.operation}"
end
end
end
|