legacy_backward.yaml 55.9 KB
Newer Older
1
- backward_op : abs_double_grad
Z
zyfncg 已提交
2 3 4 5 6 7 8 9 10
  forward : abs_grad (Tensor x, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_x_grad)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : abs_double_grad

11
- backward_op : abs_grad
Z
zyfncg 已提交
12 13 14 15 16 17 18 19
  forward : abs (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : abs_grad
20
  composite : abs_grad(x, out_grad, x_grad)
Z
zyfncg 已提交
21 22
  backward : abs_double_grad

23
- backward_op : add_double_grad
Z
zyfncg 已提交
24 25 26 27 28 29 30 31 32 33 34 35
  forward : add_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : add_double_grad
  optional : grad_x_grad, grad_y_grad
  backward : add_triple_grad
  inplace : (grad_x_grad -> grad_out_grad)

36
- backward_op : add_grad
Z
zyfncg 已提交
37 38 39 40 41 42 43 44 45
  forward : add (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : add_grad
  no_need_buffer : x, y
46
  composite : add_grad(x, y, out_grad, axis)
Z
zyfncg 已提交
47 48 49
  backward : add_double_grad
  inplace : (out_grad -> x_grad)

50
- backward_op : add_triple_grad
Z
zyfncg 已提交
51 52 53 54 55 56 57 58 59 60
  forward : add_double_grad (Tensor y, Tensor grad_out, Tensor grad_grad_x, Tensor grad_grad_y, int axis = -1) -> Tensor(grad_grad_out)
  args : (Tensor grad_grad_x, Tensor grad_grad_y, Tensor grad_grad_out_grad, int axis = -1)
  output : Tensor(grad_grad_x_grad), Tensor(grad_grad_y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [grad_grad_x, grad_grad_y]
  kernel :
    func : add_triple_grad
  inplace : (grad_grad_out_grad -> grad_grad_x_grad)

61
- backward_op : affine_grid_grad
62 63
  forward : affine_grid (Tensor input, IntArray outputShape, bool align_corners=true) -> Tensor(output)
  args : (Tensor input, Tensor output_grad, IntArray outputShape, bool align_corners=true)
64 65 66 67 68 69 70
  output : Tensor(input_grad)
  infer_meta :
    func : AffineGridGradInferMeta
    param : [output_grad, outputShape, align_corners]
  kernel :
    func : affine_grid_grad
    param : [output_grad, outputShape, align_corners]
71
  no_need_buffer : input
72

73
- backward_op : amax_grad
74 75
  forward: amax (Tensor x,  int64_t[] axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis={},  bool keepdim=false, bool reduce_all=false)
76 77 78 79 80 81 82
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : amax_grad

83
- backward_op : amin_grad
84 85
  forward: amin (Tensor x,  int64_t[] axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis={},  bool keepdim=false, bool reduce_all=false)
86 87 88 89 90 91 92
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : amin_grad

93
- backward_op : assign_grad
Z
zyfncg 已提交
94 95 96
  forward : assign (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
97
  invoke : assign(out_grad)
Z
zyfncg 已提交
98

99
- backward_op : assign_out__grad
Z
zyfncg 已提交
100 101 102 103 104 105 106 107 108
  forward : assign_out_ (Tensor x, Tensor output) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
  kernel :
    func : assign
  inplace : (out_grad -> x_grad)

109
- backward_op : batch_norm_double_grad
110 111
  forward : batch_norm_grad (Tensor x, Tensor scale, Tensor bias, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor grad_out, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics) -> Tensor(grad_x), Tensor(grad_scale), Tensor(grad_bias)
  args : (Tensor x, Tensor scale, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor grad_out,  Tensor grad_x_grad, Tensor grad_scale_grad, Tensor grad_bias_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics)
Z
zyfncg 已提交
112 113 114 115 116 117 118
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, x]
  kernel :
    func : batch_norm_grad_grad
    data_type : x
119
  optional : out_mean, out_variance, grad_x_grad, grad_scale_grad, grad_bias_grad
Z
zyfncg 已提交
120 121
  inplace : (grad_out -> grad_out_grad)

122
- backward_op : batch_norm_grad
123 124
  forward : batch_norm (Tensor x, Tensor mean, Tensor variance, Tensor scale, Tensor bias, bool is_test, float momentum, float epsilon, str data_layout, bool use_global_stats, bool trainable_statistics) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor mean_out, Tensor variance_out, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor out_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics)
Z
zyfncg 已提交
125 126 127 128 129 130 131 132 133 134
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : batch_norm_grad
    data_type : out_grad
  optional : mean_out, variance_out, reserve_space
  backward : batch_norm_double_grad

135
- backward_op : bce_loss_grad
Z
zyfncg 已提交
136 137 138 139 140 141 142 143 144 145
  forward : bce_loss (Tensor input, Tensor label) -> Tensor(out)
  args : (Tensor input, Tensor label, Tensor out_grad)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : bce_loss_grad
  inplace : (out_grad -> input_grad)

146
- backward_op : bilinear_tensor_product_grad
147 148 149 150 151 152 153 154
  forward : bilinear_tensor_product (Tensor x, Tensor y, Tensor weight, Tensor bias) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor weight, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(weight_grad), Tensor(bias_grad)
  infer_meta :
    func : BilinearTensorProductGradInferMeta
  kernel :
    func : bilinear_tensor_product_grad

155
- backward_op : cast_grad
156
  forward : cast (Tensor x, DataType dtype) -> Tensor(out)
Z
zyfncg 已提交
157 158
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
159
  invoke : cast (out_grad, x.dtype())
160
  composite: cast_grad(x, out_grad)
Z
zyfncg 已提交
161 162
  no_need_buffer : x

163 164 165 166 167 168 169 170 171
- backward_op : channel_shuffle_grad
  forward : channel_shuffle (Tensor x, int groups, str data_format="NCHW") -> Tensor(out)
  args : (Tensor out_grad, int groups, str data_format="NCHW")
  output : Tensor(x_grad)
  infer_meta :
    func : ChannelShuffleGradInferMeta
  kernel :
    func : channel_shuffle_grad

172
- backward_op : concat_double_grad
Z
zyfncg 已提交
173 174 175
  forward : concat_grad (Tensor[] x, Tensor grad_out, Scalar axis) -> Tensor[](grad_x)
  args : (Tensor[] grad_x_grad, Scalar axis = 0)
  output : Tensor(grad_out_grad)
176
  invoke : concat(grad_x_grad, axis)
Z
zyfncg 已提交
177

178
- backward_op : concat_grad
Z
zyfncg 已提交
179 180 181 182 183 184 185 186
  forward : concat (Tensor[] x, Scalar axis) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, Scalar axis = 0)
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : UnchangedMultiInferMeta
    param : [x]
  kernel :
    func : concat_grad
W
wangzhen38 已提交
187
  composite : concat_grad(x, out_grad, axis, x_grad)
Z
zyfncg 已提交
188 189 190
  no_need_buffer : x
  backward : concat_double_grad

191
- backward_op : conv2d_grad
192 193
  forward : conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad,  int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format)
Z
zyfncg 已提交
194
  output : Tensor(input_grad), Tensor(filter_grad)
Z
zyfncg 已提交
195 196 197 198 199
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : conv2d_grad
Z
zyfncg 已提交
200 201
  backward : conv2d_grad_grad

202
- backward_op : conv2d_grad_grad
203 204
  forward : conv2d_grad (Tensor input, Tensor filter, Tensor grad_out,  int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format) -> Tensor(grad_input), Tensor(grad_filter)
  args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str padding_algorithm, int[] dilations, int groups, str data_format)
Z
zyfncg 已提交
205 206 207 208 209 210 211 212
  output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [input, filter, grad_out]
  kernel :
    func : conv2d_grad_grad
  optional : grad_input_grad, grad_filter_grad

213
- backward_op : conv2d_transpose_double_grad
214 215
  forward : conv2d_transpose_grad(Tensor x, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_x), Tensor(grad_filter)
  args : (Tensor x, Tensor filter, Tensor grad_out, Tensor grad_x_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
216 217 218 219 220 221
  output : Tensor(x_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : Conv2dTransposeDoubleGradInferMeta
  kernel :
    func : conv2d_transpose_grad_grad

222
- backward_op : conv2d_transpose_grad
223 224
  forward : conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
225 226
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
227
    func : Conv2dTransposeGradInferMeta
Z
zyfncg 已提交
228 229 230 231
  kernel :
    func : conv2d_transpose_grad
  backward : conv2d_transpose_double_grad

232 233 234 235 236 237 238 239 240 241 242
- backward_op : conv3d_double_grad
  forward : conv3d_grad (Tensor input, Tensor filter, Tensor grad_out,  int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_input), Tensor(grad_filter)
  args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [input, filter, grad_out]
  kernel :
    func : conv3d_double_grad
  optional : grad_input_grad, grad_filter_grad

243
- backward_op : conv3d_grad
244 245
  forward : conv3d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad,  int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
246
  output : Tensor(input_grad), Tensor(filter_grad)
Z
zyfncg 已提交
247 248 249 250 251
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : conv3d_grad
252
  backward : conv3d_double_grad
Z
zyfncg 已提交
253

254
- backward_op : conv3d_transpose_grad
Z
zyfncg 已提交
255 256 257 258 259 260 261 262
  forward : conv3d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
  kernel :
    func : conv3d_transpose_grad

263
- backward_op : cross_entropy_with_softmax_grad
Z
zyfncg 已提交
264 265 266 267 268 269 270 271 272 273
  forward : cross_entropy_with_softmax (Tensor input, Tensor label, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis) -> Tensor(softmax), Tensor(loss)
  args : (Tensor label, Tensor softmax, Tensor loss_grad, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis)
  output : Tensor(input_grad)
  infer_meta :
    func : CrossEntropyWithSoftmaxGradInferMeta
  kernel :
    func : cross_entropy_with_softmax_grad
    data_type : softmax
  inplace : (softmax -> input_grad)

274
- backward_op : cumprod_grad
Z
zyfncg 已提交
275 276 277 278 279 280 281 282 283
  forward : cumprod (Tensor x, int dim) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int dim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : cumprod_grad

284
- backward_op : cumsum_grad
W
WangZhen 已提交
285
  forward : cumsum(Tensor x, Scalar axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
286
  args : (Tensor x, Tensor out_grad, Scalar axis, bool flatten, bool exclusive, bool reverse)
Z
zyfncg 已提交
287
  output : Tensor(x_grad)
288 289 290 291 292 293
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : cumsum_grad
    data_type: x
G
GGBond8488 已提交
294
  composite: cumsum_grad(x, out_grad, axis, flatten, exclusive, reverse, x_grad)
Z
zyfncg 已提交
295

296
- backward_op : deformable_conv_grad
Z
zyfncg 已提交
297 298 299 300 301 302 303 304 305 306
  forward : deformable_conv(Tensor x, Tensor offset, Tensor filter, Tensor mask, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step) -> Tensor(out)
  args : (Tensor x, Tensor offset, Tensor filter, Tensor mask, Tensor out_grad, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step)
  output : Tensor(x_grad), Tensor(offset_grad), Tensor(filter_grad), Tensor(mask_grad)
  infer_meta :
    func : DeformableConvGradInferMeta
  kernel :
    func : deformable_conv_grad
    data_type : x
  optional : mask

307
- backward_op : depthwise_conv2d_double_grad
308
  forward : depthwise_conv2d_grad (Tensor input, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_input), Tensor(grad_filter)
309 310 311 312 313 314 315 316 317
  args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [input, filter, grad_out]
  kernel :
    func : depthwise_conv2d_double_grad
  optional : grad_input_grad, grad_filter_grad

318
- backward_op : depthwise_conv2d_grad
319 320
  forward : depthwise_conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
321 322 323 324 325 326
  output : Tensor(input_grad), Tensor(filter_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [input, filter]
  kernel :
    func : depthwise_conv2d_grad
327 328
    param : [input, filter, out_grad, strides, paddings, padding_algorithm, groups, dilations, data_format]
  backward : depthwise_conv2d_double_grad
Z
zyfncg 已提交
329

330
- backward_op : depthwise_conv2d_transpose_grad
331 332
  forward : depthwise_conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, IntArray output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
Z
zyfncg 已提交
333 334
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
335
    func : Conv2dTransposeGradInferMeta
Z
zyfncg 已提交
336 337 338
  kernel :
    func : depthwise_conv2d_transpose_grad

339
- backward_op : divide_double_grad
Z
zyfncg 已提交
340 341 342 343 344 345 346 347 348 349 350 351
  forward : divide_grad (Tensor x, Tensor y, Tensor out, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor out, Tensor grad_x, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(y_grad), Tensor(out_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [y, grad_x, grad_x]
  kernel :
    func : divide_double_grad
    data_type : out
  optional : grad_x_grad, grad_y_grad
  inplace : (grad_x_grad -> grad_out_grad)

352
- backward_op : divide_grad
Z
zyfncg 已提交
353 354 355 356 357 358 359 360
  forward : divide (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : divide_grad
361
  composite : divide_grad(x, y, out, out_grad, axis)
Z
zyfncg 已提交
362 363
  backward : divide_double_grad

364
- backward_op : dropout_grad
365 366
  forward : dropout (Tensor x, Tensor seed_tensor, Scalar p, bool is_test, str mode, int seed, bool fix_seed) -> Tensor(out), Tensor(mask)
  args : (Tensor mask, Tensor out_grad, Scalar p, bool is_test, str mode)
Z
zyfncg 已提交
367 368 369 370 371 372 373
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : dropout_grad

374
- backward_op : eigvalsh_grad
375 376 377 378 379 380 381 382 383 384 385
  forward : eigvalsh (Tensor x, str uplo, bool is_test) -> Tensor(eigenvalues), Tensor(eigenvectors)
  args : (Tensor eigenvectors, Tensor eigenvalues_grad, str uplo, bool is_test)
  output : Tensor(x_grad)
  infer_meta :
    func : EigvalshGradInferMeta
  kernel :
    func : eigvalsh_grad
    data_type : eigenvectors
  data_transform :
    skip_transform : eigenvalues_grad

386
- backward_op : einsum_grad
Z
zyfncg 已提交
387 388 389 390 391 392 393 394 395
  forward : einsum (Tensor[] x, str equation) -> Tensor(out), Tensor[](inner_cache), Tensor[](x_shape)
  args : (Tensor[] x_shape, Tensor[] inner_cache, Tensor out_grad, str equation)
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : UnchangedMultiInferMeta
    param : [x_shape]
  kernel :
    func : einsum_grad

396
- backward_op : elementwise_pow_grad
Z
zyfncg 已提交
397 398 399 400 401 402
  forward : elementwise_pow(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
403
  composite : elementwise_pow_grad(x, y, out_grad, axis)
Z
zyfncg 已提交
404 405 406
  kernel :
    func : elementwise_pow_grad

407
- backward_op : embedding_grad
Z
zyfncg 已提交
408 409 410 411
  forward : embedding (Tensor x, Tensor weight, int64_t padding_idx=-1, bool sparse=false) -> Tensor(out)
  args : (Tensor x, Tensor weight, Tensor out_grad, int64_t padding_idx=-1, bool sparse=false)
  output : Tensor(weight_grad)
  invoke : embedding_grad_impl(x, weight, out_grad, padding_idx, sparse, weight_grad)
W
wanghuancoder 已提交
412
  no_need_buffer : weight
Z
zyfncg 已提交
413

414
- backward_op : expand_as_grad
Z
zyfncg 已提交
415 416 417 418 419 420 421 422 423 424
  forward : expand_as (Tensor x, Tensor y, int[] target_shape) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] target_shape)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : expand_as_grad
  no_need_buffer : x

425
- backward_op : expand_double_grad
Z
zyfncg 已提交
426 427 428
  forward : expand_grad (Tensor x, Tensor grad_out, IntArray shape) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray shape)
  output : Tensor(grad_out_grad)
429
  invoke : expand(grad_x_grad, shape)
Z
zyfncg 已提交
430

431
- backward_op : expand_grad
Z
zyfncg 已提交
432 433 434 435 436 437 438 439 440 441
  forward : expand (Tensor x, IntArray shape) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray shape)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : expand_grad
  no_need_buffer : x
  backward : expand_double_grad
442
  composite: expand_grad(x, out_grad, shape, x_grad_p)
Z
zyfncg 已提交
443

444
- backward_op : exponential__grad
445
  forward : exponential_ (Tensor x, float lam) -> Tensor(out)
446 447 448 449
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
450
  invoke : zeros_like(out_grad)
451

452
- backward_op : fill_grad
453 454 455 456 457 458 459 460 461 462
  forward : fill (Tensor x, Scalar value) -> Tensor(out)
  args : (Tensor out_grad, Scalar value)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : fill_grad
  inplace : (out_grad -> x_grad)

463
- backward_op : flatten_grad
Z
zyfncg 已提交
464 465 466 467 468 469 470 471 472 473 474 475 476
  forward : flatten(Tensor x, int start_axis, int stop_axis) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func :  KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : flatten_grad
    data_type: out_grad
    backend: out_grad
    layout: out_grad
  inplace : (out_grad -> x_grad)

477
- backward_op : fmax_grad
478
  forward : fmax(Tensor x, Tensor y) -> Tensor(out)
Z
zhangyuqin1998 已提交
479
  args : (Tensor x, Tensor y, Tensor out_grad)
Z
zyfncg 已提交
480 481 482 483 484 485 486
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmax_grad

487
- backward_op : fmin_grad
488
  forward : fmin(Tensor x, Tensor y) -> Tensor(out)
Z
zyfncg 已提交
489
  args : (Tensor x, Tensor y, Tensor out_grad)
Z
zyfncg 已提交
490 491 492 493 494 495 496
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmin_grad

497
- backward_op : frobenius_norm_grad
Z
zyfncg 已提交
498 499 500 501 502 503 504 505 506
  forward : frobenius_norm(Tensor x, int64_t[] axis,  bool keep_dim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keep_dim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : frobenius_norm_grad

507
- backward_op : gather_grad
Z
zyfncg 已提交
508 509 510 511 512 513 514 515 516
  forward : gather(Tensor x, Tensor index, Scalar axis=0) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, Scalar axis=0, bool overwrite=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    data_type: x
    func : gather_grad
517
  composite : gather_grad(x, index, out_grad, axis, overwrite)
Z
zyfncg 已提交
518 519
  no_need_buffer : x

520
- backward_op : group_norm_grad
Z
zyfncg 已提交
521 522 523 524 525 526 527 528 529 530 531 532
  forward : group_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int groups, str data_layout) -> Tensor(y), Tensor(mean), Tensor(variance)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor y, Tensor mean, Tensor variance, Tensor y_grad, float epsilon, int groups, str data_layout)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [y, scale, bias]
  kernel :
    func : group_norm_grad
    data_type : y_grad
  optional: scale, bias
  inplace : (y_grad -> x_grad)

533
- backward_op : hardswish_grad
534 535
  forward : hardswish (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold = 6.0, float scale = 6.0, float offset = 3.0)
Z
zyfncg 已提交
536 537 538 539 540
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
Z
zyfncg 已提交
541
    func : hardswish_grad
Z
zyfncg 已提交
542 543
  inplace : (out_grad -> x_grad)

544 545 546 547 548 549 550 551 552 553
- backward_op : heaviside_grad
  forward : heaviside (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : heaviside_grad

554
- backward_op : hsigmoid_loss_grad
555 556
  forward : hsigmoid_loss (Tensor x, Tensor label, Tensor w, Tensor bias, Tensor path, Tensor code, int num_classes, bool remote_prefetch, bool is_sparse) -> Tensor(out), Tensor(pre_out), Tensor(w_out)
  args : (Tensor x, Tensor w, Tensor label, Tensor path, Tensor code, Tensor bias, Tensor pre_out, Tensor out_grad, int num_classes, bool remote_prefetch, bool is_sparse)
557 558 559 560 561 562
  output : Tensor(x_grad), Tensor(w_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x ,w, bias]
  optional: path, code, bias
  kernel :
563
    func : hsigmoid_loss_grad
564

565
- backward_op : huber_loss_grad
Z
zyfncg 已提交
566 567 568 569 570 571 572 573 574
  forward : huber_loss (Tensor input, Tensor label, float delta) -> Tensor(out), Tensor(residual)
  args : (Tensor residual, Tensor out_grad, float delta)
  output : Tensor(input_grad), Tensor(label_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [residual, residual]
  kernel :
    func : huber_loss_grad

575
- backward_op : index_add_grad
L
Li Min 已提交
576 577 578 579 580 581 582 583 584 585
  forward : index_add(Tensor x, Tensor index,  Tensor add_value, int axis) -> Tensor(out)
  args : (Tensor index, Tensor add_value, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(add_value_grad)
  infer_meta :
    func : IndexAddGradInferMeta
  kernel :
    func : index_add_grad
    data_type : out_grad
  inplace : (out_grad -> x_grad)

586
- backward_op : instance_norm_double_grad
Z
zyfncg 已提交
587 588 589 590 591 592 593 594 595 596
  forward : instance_norm_grad(Tensor x, Tensor fwd_scale, Tensor saved_mean, Tensor saved_variance, Tensor grad_y, float epsilon) -> Tensor(grad_x), Tensor(grad_scale), Tensor(grad_bias)
  args : (Tensor x, Tensor fwd_scale, Tensor saved_mean, Tensor saved_variance, Tensor grad_y, Tensor grad_x_grad, Tensor grad_scale_grad, Tensor grad_bias_grad, float epsilon)
  output : Tensor(x_grad), Tensor(fwd_scale_grad), Tensor(grad_y_grad)
  infer_meta :
    func : InstanceNormDoubleGradInferMeta
  kernel :
    func : instance_norm_double_grad
    data_type : x
  optional : fwd_scale, grad_x_grad, grad_scale_grad, grad_bias_grad

597
- backward_op : instance_norm_grad
Z
zyfncg 已提交
598 599 600 601 602 603 604 605 606 607 608
  forward : instance_norm(Tensor x, Tensor scale, Tensor bias, float epsilon) -> Tensor(y), Tensor(saved_mean), Tensor(saved_variance)
  args : (Tensor x, Tensor scale, Tensor saved_mean, Tensor saved_variance, Tensor y_grad, float epsilon)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : InstanceNormGradInferMeta
  kernel :
    func : instance_norm_grad
    data_type : x
  optional : scale
  backward : instance_norm_double_grad

609
- backward_op : kldiv_loss_grad
Z
zyfncg 已提交
610 611 612 613 614 615 616 617 618 619
  forward : kldiv_loss(Tensor x, Tensor label, str reduction) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, str reduction)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : kldiv_loss_grad
  no_need_buffer : x

620
- backward_op : layer_norm_grad
621 622
  forward : layer_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int begin_norm_axis) -> Tensor(out), Tensor(mean), Tensor(variance)
  args : (Tensor x,  Tensor scale, Tensor bias, Tensor mean, Tensor variance, Tensor out_grad, float epsilon, int begin_norm_axis)
Z
zyfncg 已提交
623 624 625 626 627 628 629 630 631 632
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : LayerNormGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : layer_norm_grad
    data_type : out_grad
  no_need_buffer : bias
  optional : scale, bias

633
- backward_op : log_softmax_grad
Z
zyfncg 已提交
634 635 636 637 638 639 640 641 642
  forward : log_softmax(Tensor x,  int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad,  int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out]
  kernel :
    func : log_softmax_grad

643
- backward_op : logcumsumexp_grad
Z
zyfncg 已提交
644 645 646 647 648 649 650 651 652
  forward : logcumsumexp(Tensor x, int axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  args : (Tensor x, Tensor out, Tensor out_grad, int axis, bool flatten, bool exclusive, bool reverse)
  output : Tensor(x_grad)
  kernel :
    func : logcumsumexp_grad

653
- backward_op : logsumexp_grad
Z
zyfncg 已提交
654 655 656 657 658 659 660 661 662
  forward : logsumexp(Tensor x, int64_t[] axis,  bool keepdim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keepdim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : logsumexp_grad

663
- backward_op : lu_grad
L
Lin Manhui 已提交
664 665 666 667 668 669 670 671
  forward : lu (Tensor x, bool pivot) -> Tensor(out), Tensor(pivots), Tensor(infos)
  args : (Tensor x, Tensor out, Tensor pivots, Tensor out_grad, bool pivot)
  output : Tensor(x_grad)
  infer_meta :
    func : LUGradInferMeta
  kernel :
    func : lu_grad

672
- backward_op : margin_cross_entropy_grad
673 674 675 676 677 678 679 680 681 682
  forward : margin_cross_entropy (Tensor logits, Tensor label, bool return_softmax, int ring_id, int rank, int nranks, float margin1, float margin2, float margin3, float scale) -> Tensor(softmax), Tensor(loss)
  args : (Tensor logits, Tensor label, Tensor softmax, Tensor loss_grad, bool return_softmax, int ring_id, int rank, int nranks, float margin1, float margin2, float margin3, float scale)
  output : Tensor(logits_grad)
  infer_meta :
    func : MarginCrossEntropyGradInferMeta
  kernel :
    func : margin_cross_entropy_grad
    data_type : softmax
  inplace : (softmax -> logits_grad)

683
- backward_op : matmul_double_grad
Z
zyfncg 已提交
684 685 686 687 688 689 690 691
  forward : matmul_grad (Tensor x, Tensor y, Tensor grad_out, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, y, grad_out]
  kernel :
    func : matmul_double_grad
692
  composite : matmul_double_grad(x, y, grad_out, grad_x_grad, grad_y_grad, transpose_x=false, transpose_y=false)
Z
zyfncg 已提交
693 694 695
  backward : matmul_triple_grad
  optional : grad_x_grad, grad_y_grad

696
- backward_op : matmul_grad
Z
zyfncg 已提交
697 698 699 700 701 702 703 704 705 706
  forward : matmul (Tensor x, Tensor y, bool transpose_x=false, bool transpose_y=false) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : matmul_grad
  backward : matmul_double_grad

707
- backward_op : matmul_triple_grad
Z
zyfncg 已提交
708 709 710 711 712 713 714 715
  forward : matmul_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
  infer_meta :
    func : GeneralQuinaryGradInferMeta
    param : [x, y, fwd_grad_out, fwd_grad_grad_x, fwd_grad_grad_y]
  kernel :
    func : matmul_triple_grad
716
  optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_x_grad, grad_y_grad, grad_grad_out_grad
Z
zyfncg 已提交
717

718
- backward_op : max_grad
719 720
  forward: max (Tensor x,  IntArray axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray axis={}, bool keepdim=false, bool reduce_all=false)
Z
zyfncg 已提交
721 722 723 724 725 726 727
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : max_grad

728
- backward_op : max_pool2d_with_index_grad
Z
zyfncg 已提交
729 730 731 732 733 734 735 736
  forward : max_pool2d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool2d_with_index_grad

737
- backward_op : max_pool3d_with_index_grad
Z
zyfncg 已提交
738 739 740 741 742 743 744 745
  forward : max_pool3d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool3d_with_index_grad

746
- backward_op : maximum_grad
Z
zyfncg 已提交
747 748 749 750 751 752 753 754 755
  forward : maximum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : maximum_grad

756
- backward_op : mean_all_grad
Z
zyfncg 已提交
757 758 759 760 761 762 763 764 765
  forward : mean_all(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_all_grad

766
- backward_op : mean_double_grad
767 768
  forward: mean_grad (Tensor x, Tensor grad_out, IntArray axis={},  bool keepdim=false, bool reduce_all = false) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray axis={},  bool keepdim=false)
Z
zyfncg 已提交
769
  output : Tensor(grad_out_grad)
770
  invoke : mean(grad_x_grad, axis, keepdim)
Z
zyfncg 已提交
771

772
- backward_op : mean_grad
773 774
  forward: mean (Tensor x,  IntArray axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray axis={},  bool keepdim=false, bool reduce_all=false)
Z
zyfncg 已提交
775 776 777 778 779 780 781 782 783
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_grad
  backward : mean_double_grad
  no_need_buffer : x

784
- backward_op : min_grad
785 786
  forward: min (Tensor x,  IntArray axis={},  bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray axis={}, bool keepdim=false, bool reduce_all=false)
Z
zyfncg 已提交
787 788 789 790 791 792 793
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : min_grad

794
- backward_op : minimum_grad
Z
zyfncg 已提交
795 796 797 798 799 800 801 802 803
  forward : minimum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : minimum_grad

804
- backward_op : mish_grad
Z
zyfncg 已提交
805 806 807 808 809 810 811 812 813 814
  forward : mish (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : mish_grad
  inplace : (out_grad -> x_grad)

815
- backward_op : multiply_double_grad
Z
zyfncg 已提交
816 817 818 819 820 821 822 823 824 825 826 827
  forward : multiply_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, y, grad_out]
  kernel :
    func : multiply_double_grad
  optional : grad_x_grad, grad_y_grad
  backward : multiply_triple_grad
  inplace : (grad_x_grad -> grad_out_grad)

828
- backward_op : multiply_grad
Z
zyfncg 已提交
829 830 831 832 833 834 835 836
  forward : multiply (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : multiply_grad
837
  composite: multiply_grad(x, y, out_grad, axis, x_grad, y_grad)
Z
zyfncg 已提交
838 839
  backward : multiply_double_grad

840
- backward_op : multiply_triple_grad
Z
zyfncg 已提交
841 842 843 844 845
  forward : multiply_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, int aixs = -1) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
  infer_meta :
    func : GeneralQuinaryGradInferMeta
846
    param : [x, y, fwd_grad_out, fwd_grad_grad_x, fwd_grad_grad_y]
Z
zyfncg 已提交
847 848
  kernel :
    func : multiply_triple_grad
849
  optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_x_grad, grad_y_grad, grad_grad_out_grad
Z
zyfncg 已提交
850

851
- backward_op : norm_grad
Z
zyfncg 已提交
852 853 854 855 856 857 858 859 860
  forward : norm (Tensor x, int axis, float epsilon, bool is_test) -> Tensor(out), Tensor(norm)
  args : (Tensor x, Tensor norm, Tensor out_grad, int axis, float epsilon, bool is_test)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : norm_grad

861
- backward_op : p_norm_grad
Z
zyfncg 已提交
862 863 864 865 866 867 868 869 870
  forward : p_norm(Tensor x,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : p_norm_grad

871
- backward_op : pad3d_double_grad
Z
zyfncg 已提交
872 873 874 875 876 877 878 879
  forward : pad3d_grad(Tensor x, Tensor grad_out, IntArray paddings, str mode, float pad_value, str data_format) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray paddings, str mode, float pad_value, str data_format)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : Pad3dInferMeta
  kernel :
    func : pad3d

880
- backward_op : pad3d_grad
Z
zyfncg 已提交
881 882 883 884 885 886 887 888 889 890 891
  forward : pad3d(Tensor x, IntArray paddings, str mode,  float pad_value, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray paddings, str mode,  float pad_value, str data_format)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pad3d_grad
  no_need_buffer : x
  backward : pad3d_double_grad

892
- backward_op : pad_double_grad
893 894
  forward : pad_grad(Tensor x, Tensor grad_out, int[] paddings, Scalar pad_value) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int[] paddings, Scalar pad_value)
Z
zyfncg 已提交
895 896 897 898 899 900
  output : Tensor(grad_out_grad)
  infer_meta :
    func : PadInferMeta
  kernel :
    func : pad

901
- backward_op : pad_grad
902 903
  forward : pad(Tensor x, int[] paddings, Scalar pad_value) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] paddings, Scalar pad_value)
Z
zyfncg 已提交
904 905 906 907 908 909 910 911 912 913
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pad_grad
    param: [out_grad, paddings, pad_value]
  no_need_buffer : x
  backward : pad_double_grad

914
- backward_op : pool2d_double_grad
915 916
  forward : pool2d_grad(Tensor x, Tensor out, Tensor grad_out, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_x_grad, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
Z
zyfncg 已提交
917 918
  output : Tensor(grad_out_grad)
  infer_meta :
919
    func : Pool2DInferMeta
920
    param : [grad_x_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
Z
zyfncg 已提交
921 922
  kernel :
    func : pool2d_double_grad
923
    param : [grad_x_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
924
  no_need_buffer : x
Z
zyfncg 已提交
925

926
- backward_op : pool2d_grad
927 928
  forward : pool2d(Tensor x, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
Z
zyfncg 已提交
929 930
  output : Tensor(x_grad)
  infer_meta :
931 932
    func : UnchangedInferMeta
    param: [x]
Z
zyfncg 已提交
933 934
  kernel :
    func : pool2d_grad
935
    param : [x, out, out_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
Z
zyfncg 已提交
936 937
  backward : pool2d_double_grad

938
- backward_op : pool3d_grad
939 940
  forward : pool3d(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
Z
zyfncg 已提交
941 942
  output : Tensor(x_grad)
  infer_meta :
943 944
    func : UnchangedInferMeta
    param: [x]
Z
zyfncg 已提交
945 946
  kernel :
    func : pool3d_grad
947
    param : [x, out, out_grad, kernel_size, strides, paddings, ceil_mode, exclusive, data_format, pooling_type, global_pooling, adaptive, padding_algorithm]
Z
zyfncg 已提交
948

949
- backward_op : prelu_grad
Z
zyfncg 已提交
950 951 952 953 954 955 956 957 958
  forward : prelu(Tensor x, Tensor alpha, str data_format, str mode) -> Tensor(out)
  args : (Tensor x, Tensor alpha, Tensor out_grad, str data_format, str mode)
  output : Tensor(x_grad), Tensor(alpha_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, alpha]
  kernel :
    func : prelu_grad

959 960 961 962 963 964 965 966 967 968
- backward_op : prod_grad
  forward : prod (Tensor x, IntArray dims, bool keep_dim, bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, IntArray dims,  bool keep_dim, bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : prod_grad

969
- backward_op : psroi_pool_grad
Z
zyfncg 已提交
970 971 972 973 974 975 976 977 978 979 980
  forward : psroi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, int output_channels, float spatial_scale) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, int output_channels, float spatial_scale)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : psroi_pool_grad
    data_type : x
  optional : boxes_num

981
- backward_op : relu6_grad
982 983
  forward : relu6 (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float threshold = 6)
984 985 986 987 988 989 990 991
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : relu6_grad
  inplace : (out_grad -> x_grad)

992
- backward_op : repeat_interleave_grad
993 994
  forward : repeat_interleave(Tensor x, int repeats, int axis) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int repeats, int axis)
S
seemingwang 已提交
995 996 997 998 999 1000 1001
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : repeat_interleave_grad

1002
- backward_op : repeat_interleave_with_tensor_index_grad
1003 1004
  forward : repeat_interleave_with_tensor_index(Tensor x, Tensor repeats, int axis) -> Tensor(out)
  args : (Tensor x, Tensor repeats, Tensor out_grad, int axis)
S
seemingwang 已提交
1005 1006 1007 1008 1009 1010 1011 1012
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : repeat_interleave_with_tensor_index_grad
    data_type : x

1013
- backward_op : reshape_double_grad
Z
zyfncg 已提交
1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024
  forward : reshape_grad (Tensor xshape, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor grad_out, Tensor grad_x_grad)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : reshape_double_grad
  no_need_buffer : grad_out
  inplace : (grad_x_grad -> grad_out_grad)

1025
- backward_op : reshape_grad
Z
zyfncg 已提交
1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
  forward : reshape (Tensor x, IntArray shape) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : reshape_grad
    param : [out_grad]
    data_type: out_grad
    backend: out_grad
    layout: out_grad
  backward : reshape_double_grad
  inplace : (out_grad -> x_grad)

1041
- backward_op : reverse_grad
1042 1043
  forward : reverse (Tensor x, IntArray axis) -> Tensor(out)
  args : (Tensor out_grad, IntArray axis)
W
wanghuancoder 已提交
1044 1045 1046
  output : Tensor(x_grad)
  invoke : reverse(out_grad, axis)

Y
YuanRisheng 已提交
1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
- backward_op : rnn_grad
  forward : rnn (Tensor x, Tensor[] pre_state, Tensor[] weight_list, Tensor sequence_length, Tensor dropout_state_in, float dropout_prob, bool is_bidirec, int input_size, int hidden_size, int num_layers, str mode, int seed, bool is_test) -> Tensor(out), Tensor(dropout_state_out), Tensor[](state), Tensor(reserve)
  args : (Tensor x, Tensor[] pre_state, Tensor[] weight_list, Tensor sequence_length, Tensor out, Tensor dropout_state_out, Tensor reserve, Tensor out_grad, Tensor[] state_grad, float dropout_prob, bool is_bidirec, int input_size, int hidden_size, int num_layers, str mode, int seed, bool is_test)
  output : Tensor(x_grad), Tensor[](pre_state_grad){pre_state.size()}, Tensor[](weight_list_grad){weight_list.size()}
  infer_meta :
    func : RnnGradInferMeta
    param : [x, pre_state, weight_list]
  kernel :
    func : rnn_grad
    data_type: out_grad
  optional : sequence_length

1059
- backward_op : roi_align_grad
Z
zyfncg 已提交
1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
  forward : roi_align (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_align_grad
    data_type : boxes
  no_need_buffer : x
  optional : boxes_num

1072
- backward_op : roi_pool_grad
Z
zyfncg 已提交
1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
  forward : roi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale) -> Tensor(out), Tensor(arg_max)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor arg_max, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_pool_grad
    data_type : x
  optional : boxes_num

W
Weilong Wu 已提交
1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094
- backward_op : rrelu_grad
  forward : rrelu (Tensor x, float lower, float upper, bool is_test) -> Tensor(out), Tensor(noise)
  args : (Tensor x, Tensor noise, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : RReluGradInferMeta
    param : [out_grad, noise]
  kernel :
    func : rrelu_grad
    data_type : x

1095
- backward_op : segment_pool_grad
Z
zyfncg 已提交
1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106
  forward : segment_pool (Tensor x, Tensor segment_ids, str pooltype) -> Tensor(out), Tensor(summed_ids)
  args : (Tensor x, Tensor segment_ids, Tensor out, Tensor summed_ids, Tensor out_grad, str pooltype)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : segment_pool_grad
    data_type : x
  optional : summed_ids

1107
- backward_op : sigmoid_cross_entropy_with_logits_grad
Z
zyfncg 已提交
1108 1109 1110 1111 1112 1113 1114 1115 1116 1117
  forward : sigmoid_cross_entropy_with_logits (Tensor x, Tensor label, bool normalize, int ignore_index) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, bool normalize, int ignore_index)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sigmoid_cross_entropy_with_logits_grad
  inplace : (out_grad -> x_grad)

1118
- backward_op : slice_double_grad
1119 1120 1121
  forward : slice_grad (Tensor input, Tensor grad_out, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(grad_input)
  args : (Tensor grad_input_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
  output : Tensor(grad_out_grad)
1122
  invoke : slice(grad_input_grad, axes, starts, ends, infer_flags, decrease_axis)
1123

1124
- backward_op : slice_grad
Z
zyfncg 已提交
1125 1126 1127 1128 1129 1130 1131 1132
  forward : slice (Tensor input, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(out)
  args : (Tensor input, Tensor out_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : slice_grad
X
xiaoguoguo626807 已提交
1133
  composite: slice_grad(input, out_grad, axes, starts, ends, infer_flags, decrease_axis)
1134
  backward : slice_double_grad
Z
zyfncg 已提交
1135 1136
  no_need_buffer : input

1137
- backward_op : softmax_grad
Z
zyfncg 已提交
1138 1139 1140 1141 1142 1143 1144 1145 1146
  forward : softmax (Tensor x, int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : softmax_grad

1147
- backward_op : spectral_norm_grad
1148 1149 1150 1151 1152 1153 1154 1155 1156
  forward : spectral_norm (Tensor weight, Tensor u, Tensor v, int dim, int power_iters, float eps) -> Tensor(out)
  args : (Tensor weight, Tensor u, Tensor v, Tensor out_grad, int dim, int power_iters, float eps)
  output : Tensor(weight_grad)
  infer_meta :
    func : SpectralNormGradInferMeta
  kernel :
    func : spectral_norm_grad
    data_type : out_grad

1157
- backward_op : split_grad
Z
zyfncg 已提交
1158 1159 1160 1161
  forward : split (Tensor x, IntArray num_or_sections, Scalar axis) -> Tensor[](out)
  args : (Tensor[] out_grad, Scalar axis = -1)
  output : Tensor(x_grad)
  invoke : concat( out_grad, axis)
C
Charles-hit 已提交
1162

1163
- backward_op : split_with_num_grad
C
Charles-hit 已提交
1164 1165 1166 1167
  forward : split_with_num (Tensor x, int num, Scalar axis) -> Tensor[](out)
  args : (Tensor[] out_grad, Scalar axis = -1)
  output : Tensor(x_grad)
  invoke : concat( out_grad, axis)
Z
zyfncg 已提交
1168

1169
- backward_op : squared_l2_norm_grad
1170 1171 1172 1173 1174 1175 1176 1177 1178
  forward : squared_l2_norm(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : squared_l2_norm_grad

1179
- backward_op : strided_slice_grad
Z
zyfncg 已提交
1180 1181 1182 1183 1184 1185 1186 1187 1188 1189
  forward : strided_slice (Tensor x, int[] axes, IntArray starts, IntArray ends, IntArray strides) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] axes, IntArray starts, IntArray ends, IntArray strides)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : strided_slice_grad
  no_need_buffer : x

1190
- backward_op : subtract_double_grad
Z
zyfncg 已提交
1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202
  forward : subtract_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : subtract_double_grad
  optional : grad_x_grad, grad_y_grad
  no_need_buffer : y, grad_out
  inplace : (grad_x_grad -> grad_out_grad)

1203
- backward_op : subtract_grad
Z
zyfncg 已提交
1204 1205 1206 1207 1208 1209 1210 1211 1212
  forward : subtract (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : subtract_grad
  no_need_buffer : x, y
1213
  composite : subtract_grad(x, y, out_grad, axis)
Z
zyfncg 已提交
1214 1215 1216
  backward : subtract_double_grad
  inplace : (out_grad -> x_grad)

1217
- backward_op : sum_double_grad
1218 1219
  forward : sum_grad (Tensor x, Tensor grad_out, IntArray axis, bool keepdim, bool reduce_all=false) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray axis={}, bool keepdim=false)
Z
zyfncg 已提交
1220
  output : Tensor(grad_out_grad)
1221
  invoke : sum(grad_x_grad, axis, grad_x_grad.dtype(), keepdim)
Z
zyfncg 已提交
1222

1223
- backward_op : sum_grad
1224 1225
  forward : sum (Tensor x, IntArray axis={}, DataType dtype=DataType::UNDEFINED, bool keepdim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray axis, bool keepdim, bool reduce_all=false)
Z
zyfncg 已提交
1226 1227 1228 1229 1230 1231
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sum_grad
1232
  composite : sum_grad(x, out_grad, axis, keepdim, reduce_all, x_grad)
Z
zyfncg 已提交
1233 1234 1235
  no_need_buffer : x
  backward : sum_double_grad

1236
- backward_op : swish_grad
1237
  forward : swish (Tensor x) -> Tensor(out)
Z
zyfncg 已提交
1238 1239 1240 1241 1242 1243 1244 1245 1246
  args : (Tensor x, Tensor out_grad, float bete=1.0)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : swish_grad
  inplace : (out_grad -> x_grad)

1247
- backward_op : sync_batch_norm_grad
1248 1249
  forward : sync_batch_norm_ (Tensor x, Tensor mean, Tensor variance, Tensor scale, Tensor bias, bool is_test, float momentum, float epsilon, str data_layout, bool use_global_stats, bool trainable_statistics) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor out_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics)
1250 1251 1252 1253 1254 1255 1256
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : sync_batch_norm_grad
    data_type : out_grad
1257
  optional : reserve_space
1258

1259
- backward_op : temporal_shift_grad
C
ccrrong 已提交
1260 1261 1262 1263 1264 1265 1266 1267 1268
  forward : temporal_shift(Tensor x, int seg_num, float shift_ratio, str data_format_str) -> Tensor(out)
  args : (Tensor out_grad, int seg_num, float shift_ratio, str data_format_str)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : temporal_shift_grad

1269
- backward_op : tile_double_grad
Z
zyfncg 已提交
1270 1271 1272
  forward : tile_grad (Tensor x, Tensor grad_out, IntArray repeat_times) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, IntArray repeat_times)
  output : Tensor(grad_out_grad)
1273
  invoke : tile(grad_x_grad, repeat_times)
Z
zyfncg 已提交
1274

1275
- backward_op : tile_grad
Z
zyfncg 已提交
1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286
  forward : tile (Tensor x, IntArray repeat_times) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray repeat_times)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : tile_grad
  no_need_buffer : x
  backward : tile_double_grad

1287
- backward_op : transpose_double_grad
1288 1289
  forward : transpose_grad (Tensor grad_out, int[] perm) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int[] perm)
Z
zyfncg 已提交
1290
  output : Tensor(grad_out_grad)
1291
  invoke : transpose(grad_x_grad, perm)
Z
zyfncg 已提交
1292

1293
- backward_op : transpose_grad
1294 1295
  forward : transpose (Tensor x, int[] perm) -> Tensor(out)
  args : (Tensor out_grad, int[] perm)
Z
zyfncg 已提交
1296 1297 1298
  output : Tensor(x_grad)
  infer_meta :
    func : TransposeGradInferMeta
1299
    param : [out_grad, perm]
Z
zyfncg 已提交
1300 1301 1302
  kernel :
    func : transpose_grad
  backward : transpose_double_grad
1303
  composite: transpose_grad(out_grad, perm)
Z
zyfncg 已提交
1304

1305
- backward_op : triangular_solve_grad
Z
zyfncg 已提交
1306 1307 1308 1309 1310 1311 1312 1313 1314
  forward : triangular_solve (Tensor x, Tensor y, bool upper, bool tranpose, bool unitriangular) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, bool upper, bool tranpose, bool unitriangular)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : triangular_solve_grad

1315
- backward_op : tril_grad
1316 1317
  forward : tril(Tensor x,  int diagonal) -> Tensor(out)
  args : (Tensor out_grad,  int diagonal)
Z
zyfncg 已提交
1318 1319 1320 1321 1322
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
1323
    func : tril_grad
Z
zyfncg 已提交
1324

1325 1326 1327 1328 1329 1330 1331 1332 1333 1334
- backward_op : triu_grad
  forward : triu(Tensor x,  int diagonal) -> Tensor(out)
  args : (Tensor out_grad,  int diagonal)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : triu_grad

1335 1336
- backward_op : uniform_inplace_grad
  forward : uniform_inplace(Tensor x, float min, float max, int seed, int diag_num, int diag_step, float diag_val) -> Tensor(out)
1337 1338 1339 1340 1341
  args : (Tensor out_grad, float min, float max, int seed, int diag_num, int diag_step, float diag_val)
  output : Tensor(x_grad)
  infer_meta :
    func : UniformRandomInplaceGradInferMeta
  kernel :
1342
    func : uniform_inplace_grad
1343 1344
  inplace : (out_grad -> x_grad)

1345
- backward_op : warpctc_grad
1346
  forward : warpctc (Tensor logits, Tensor label, Tensor logits_length, Tensor labels_length, int blank, bool norm_by_times) -> Tensor(loss), Tensor(warpctcgrad)
Z
Zhong Hui 已提交
1347 1348 1349 1350 1351 1352 1353 1354 1355 1356
  args : (Tensor logits, Tensor logits_length, Tensor warpctcgrad, Tensor loss_grad, int blank, bool norm_by_times)
  output : Tensor(logits_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [logits]
  kernel :
    func : warpctc_grad
  optional : logits_length
  no_need_buffer : logits

1357 1358
- backward_op : yolo_loss_grad
  forward : yolo_loss(Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0) -> Tensor(loss), Tensor(objectness_mask), Tensor(gt_match_mask)
1359 1360 1361
  args : (Tensor x, Tensor gt_box, Tensor gt_label, Tensor gt_score, Tensor objectness_mask, Tensor gt_match_mask, Tensor loss_grad, int[] anchors, int[] anchor_mask, int class_num, float ignore_thresh, int downsample_ratio, bool use_label_smooth=true, float scale_x_y=1.0)
  output : Tensor(x_grad), Tensor(gt_box_grad), Tensor(gt_label_grad), Tensor(gt_score_grad)
  infer_meta :
1362
    func : YoloLossGradInferMeta
1363
  kernel :
1364
    func : yolo_loss_grad
1365
  optional : gt_score
X
xiaoting 已提交
1366

1367
- backward_op: unpool3d_grad
X
xiaoting 已提交
1368 1369 1370 1371 1372 1373 1374 1375 1376 1377
  forward: unpool3d (Tensor x, Tensor indices, int[] ksize, int[] strides, int[] padding, int[] output_size, str data_format) -> Tensor(out)
  args: (Tensor x, Tensor indices, Tensor out, Tensor out_grad, int[] ksize, int[] strides, int[] padding, int[] output_size, str data_format)
  output: Tensor(x_grad)
  infer_meta:
    func: UnchangedInferMeta
    param : [x]
  kernel:
    func: unpool3d_grad
    data_type: x

1378
- backward_op: unpool_grad
1379 1380
  forward: unpool (Tensor x, Tensor indices, int[] ksize, int[] strides, int[] padding,  IntArray output_size, str data_format) -> Tensor(out)
  args: (Tensor x, Tensor indices, Tensor out, Tensor out_grad, int[] ksize, int[] strides, int[] padding, IntArray output_size, str data_format)
X
xiaoting 已提交
1381 1382 1383 1384 1385 1386 1387
  output: Tensor(x_grad)
  infer_meta:
    func: UnchangedInferMeta
    param : [x]
  kernel:
    func: unpool_grad
    data_type: x