operators.json 223.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
[
{
 "type" : "sgd",
 "comment" : "\n\nSGD operator\n\nThis operator implements one step of the stochastic gradient descent algorithm.\n\n$$param\\_out = param - learning\\_rate * grad$$\n\n",
 "inputs" : [ 
 { 
   "name" : "Param",
   "comment" : "(Tensor) Input parameter",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor) Learning rate of SGD",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Grad",
   "comment" : "(Tensor) Input gradient",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "ParamOut",
   "comment" : "(Tensor) Output parameter",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "adagrad",
32
 "comment" : "\n\nAdaptive Gradient Algorithm (Adagrad).\n\nThe update is done as follows:\n\n$$moment\\_out = moment + grad * grad \\\\\nparam\\_out = param - \\frac{learning\\_rate * grad}{\\sqrt{moment\\_out} + \\epsilon}\n$$\n\nThe original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)\ndoes not have the epsilon attribute. It is added here in our implementation\nas also proposed here: http://cs231n.github.io/neural-networks-3/#ada\nfor numerical stability to avoid the division by zero error.\n\n",
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730
 "inputs" : [ 
 { 
   "name" : "Param",
   "comment" : "(Tensor) Input parameter",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Grad",
   "comment" : "(Tensor) Input gradient",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Moment",
   "comment" : "(Tensor) Second moment",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor) Learning rate",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "ParamOut",
   "comment" : "(Tensor) Output parameter",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "MomentOut",
   "comment" : "(Tensor) Output second moment",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "epsilon",
   "type" : "float",
   "comment" : "(float, default 1.0e-6) Constant for numerical stability",
   "generated" : 0
 } ] 
},{
 "type" : "conv3d",
 "comment" : "\nConvolution3D Operator.\n\nThe convolution operation calculates the output based on the input, filter\nand strides, paddings, dilations, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and output(Output) are in NCDHW format, where N is batch\nsize, C is the number of channels,D is the depth of the feature, H is the height of\nthe feature, and W is the width of the feature.\nFilters(Input) is MCDHW format, where M is the number of output image channels,\nC is the number of input image channels, D is the depth of the filter,\nH is the height of the filter, and W is the width of the filter.\nParameters(strides, paddings, dilations) are three elements. These three elements\nrepresent depth, height and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{out}, C_{in}, D_f, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       D_{out}= \\frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{ strides[0]}+ 1 \\\\\n       H_{out}= \\frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{ strides[1]}+ 1 \\\\\n       W_{out}= \\frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{ strides[2]}+ 1\n  $$\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution operator. The format of input tensor is NCDHW. Where N is batch size, C is the number of channels, D is the depth of the feature, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution operator. The format of the filter tensor is MCDHW, where M is the number of output image channels, C is the number of input image channels, D is the depth of the filter, H is the height of the filter, and W is the width of the filter.If the groups attribute is greater than 1, C equals the number of input image channels divided by the groups.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution operator.The format of output tensor is also NCDHW.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int>, default:{1, 1, 1}), the strides(d_stride, h_stride, w_stride) of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int>, default:{0, 0, 0}), the paddings(d_pad, h_pad, w_pad) of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "groups",
   "type" : "int",
   "comment" : "(int default:1), the groups number of the convolution operator. According to grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels.",
   "generated" : 0
 }, { 
   "name" : "dilations",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1, 1}), the dilations(d_dilation, h_dilation, w_dilation) of convolution operator.",
   "generated" : 0
 } ] 
},{
 "type" : "conv2d",
 "comment" : "\nConvolution Operator.\n\nThe convolution operation calculates the output based on the input, filter\nand strides, paddings, dilations, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and Output(Output) are in NCHW format. Where N is batch\nsize, C is the number of channels, H is the height of the feature, and W is\nthe width of the feature.\nFilters(Input) is MCHW format. Where M is the number of output image channels, C is\nthe number of input image channels, H is the height of the filter, and W\nis the width of the filter.\nParameters(strides, paddings, dilations) are two elements. These two elements represent\nheight and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       Input shape: $(N, C_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{out}, C_{in}, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, H_{out}, W_{out})$\n  Where\n$$\n       H_{out}= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]}+ 1 \\\\\n       W_{out}= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]}+ 1\n$$\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution operator. The format of input tensor is NCHW, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution operator. The format of the filter tensor is MCHW, where M is the number of output image channels, C is the number of input image channels, H is the height of the filter, and W is the width of the filter. If the groups attribute is greater than 1, C equals the number of input image channels divided by the groups.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution operator. The format of output tensor is also NCHW.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1}), the strides(h_stride, w_stride) of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int> default:{0, 0}), the paddings(h_pad, w_pad) of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "groups",
   "type" : "int",
   "comment" : "(int default:1), the groups number of the convolution operator. According to grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels.",
   "generated" : 0
 }, { 
   "name" : "dilations",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1}), the dilations(h_dilation, w_dilation) of convolution operator.",
   "generated" : 0
 } ] 
},{
 "type" : "pool3d",
 "comment" : "\nPool3d Operator.\n\nThe pooling3d operation calculates the output based on\nthe input, pooling_type, ksize, strides, and paddings parameters.\nInput(X) and output(Out) are in NCDHW format, where N is batch\nsize, C is the number of channels, and D, H and W are the depth, height and\nwidth of the feature, respectively. Parameters(ksize, strides, paddings) \nare three elements. These three elements represent depth, height and \nwidth, respectively. The input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       X shape: $(N, C, D_{in}, H_{in}, W_{in})$\n  Output:\n       Out shape: $(N, C, D_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       D_{out} = \\frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\\\\n       H_{out} = \\frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\\\\n       W_{out} = \\frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1\n  $$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor of pooling operator. The format of input tensor is NCDHW, where N is batch size, C is the number of channels, and D, H and W is the depth, height and width of the feature, respectively.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of pooling operator.The format of output tensor is also NCDHW, where N is batch size, C is the number of channels, and D, H and W is the depth, height and width of the feature, respectively.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "pooling_type",
   "type" : "string",
   "comment" : "(string) Pooling type, can be \"max\" for max-pooling and \"avg\" for average-pooling.",
   "generated" : 0
 }, { 
   "name" : "ksize",
   "type" : "int array",
   "comment" : "(vector<int>) The pooling window size(depth, height, width) of pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 }, { 
   "name" : "global_pooling",
   "type" : "bool",
   "comment" : "(bool, default false) Whether to use the global pooling. If global_pooling = true, ksize and paddings wille be ignored.",
   "generated" : 0
 }, { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int>, default {1,1,1}) Strides(depth, height, width) of the pooling operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int>, default {0,0,0}), paddings(depth, height, width) of pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 } ] 
},{
 "type" : "pool2d",
 "comment" : "\nPool2d Operator.\n\nThe pooling2d operation calculates the output based on\nthe input, pooling_type and ksize, strides, paddings parameters.\nInput(X) and output(Out) are in NCHW format, where N is batch size, C is the\nnumber of channels, H is the height of the feature, and W is the width of the feature.\nParameters(ksize, strides, paddings) are two elements.\nThese two elements represent height and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:   \n  Input:\n       X shape: $(N, C, H_{in}, W_{in})$\n  Output:\n       Out shape: $(N, C, H_{out}, W_{out})$\n  Where\n       $$ \n       H_{out} = \\frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\\\\n       W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1\n       $$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor of pooling operator. The format of input tensor is NCHW, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of pooling operator. The format of output tensor is also NCHW, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "pooling_type",
   "type" : "string",
   "comment" : "(string), pooling type, can be \"max\" for max-pooling and \"avg\" for average-pooling.",
   "generated" : 0
 }, { 
   "name" : "ksize",
   "type" : "int array",
   "comment" : "(vector<int>) The pooling window size(height, width) of the pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 }, { 
   "name" : "global_pooling",
   "type" : "bool",
   "comment" : "(bool, default false) Whether to use the global pooling. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 }, { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int>, default {1, 1}), strides(height, width) of pooling operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int>, default {0,0}), paddings(height, width) of pooling operator.If global_pooling = true, paddings and ksize will be ignored.",
   "generated" : 0
 } ] 
},{
 "type" : "max_pool3d_with_index",
 "comment" : "\nMaxPool3d Operator.\n\nThe maxpooling3d with index operation calculates the output and the mask\nbased on the input and ksize, strides, paddings parameters.\nInput(X) and output(Out, Mask) are in NCDHW format, where N is batch\nsize, C is the number of channels, and D, H and W are the depth, height and\nwidth of the feature, respectively. \nParameters(ksize, strides, paddings) are three elements.\nThese three elements represent depth, height and width, respectively.\nThe input(X) size and output(Out, Mask) size may be different.\n\nExample:\n  Input:\n       X shape: $(N, C, D_{in}, H_{in}, W_{in})$\n  Output:\n       Out shape: $(N, C, D_{out}, H_{out}, W_{out})$\n       Mask shape: $(N, C, D_{out}, H_{out}, W_{out})$\n  Where\n       $$\n       D_{out} = \\frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\\\\n       H_{out} = \\frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\\\\n       W_{out} = \\frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1\n       $$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor of pooling operator. The format of input tensor is NCDHW, where N is batch size, C is the number of channels, and D, H and W are the depth, height and width of the image, respectively",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of pooling operator. The format of output tensor is also NCDHW, where N is the batch size, C is the number of channels, and D, H and W are the depth, height and width of the image, respectively.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Mask",
   "comment" : "(Tensor) The Mask tensor of pooling operator. The format of output tensor is also NCDHW, where N is the batch size, C is the number of channels, and D, H and W are the depth, height and width of the image, respectively. It represents the index in the current feature map.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "ksize",
   "type" : "int array",
   "comment" : "(vector<int>) The pooling window size(depth, height, width) of pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 }, { 
   "name" : "global_pooling",
   "type" : "bool",
   "comment" : "(bool, default false) Whether to use the global pooling. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 }, { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int>, default {1,1,1}), strides(depth, height, width) of pooling operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector, default {0,0,0}), paddings(depth, height, width) of pooling operator. If global_pooling = true, paddings and ksize will be ignored.",
   "generated" : 0
 } ] 
},{
 "type" : "lod_rank_table",
 "comment" : "Create LoDRanTable by LoDTensor\n\nLoD Rank Table stores the `level` of `lod` which is ordered by sequence\nlength in descending order. It is useful when implement dynamic RNN and is\nshared by dynamic RNN memory, dynamic RNN slice input and dynamic RNN slice\noutput operators.\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(LoDTensor) input lod tensor, must contain lod information.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDRankTable) The rank table of specific level.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "level",
   "type" : "int",
   "comment" : "(int) the specific lod level to rank.",
   "generated" : 0
 } ] 
},{
 "type" : "array_to_lod_tensor",
 "comment" : "This Op build a big LoDTensor from a std::vector<LoDTensor> \n          and a LoDRankTable. It is supposed to be used in getting dynamic RNN's\n          outputs back to a normal LoDTensor. The std::vector<LoDTensor> \n          would be the output of RNN Op and the LoDRankTable would be build \n          with RNN's input.",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(std::vector<LodTensor>) A vector of tensors that is going to be casted to a big LoDTensor.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "RankTable",
   "comment" : "(LoDRankTable) RankTable provides the coarse lod infomation to build the output LoDTensor. See 'paddle/framework/lod_rank_table.h' for more details.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDTensor) The LoDTensor formed by input tensor array.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "sequence_conv",
 "comment" : "\nSequence Conv Operator.\n\nSequenceConvOp performs convolution operation on features of contextLength\ntime-steps of each instance. The convolution operation calculates the output\nbased on the input, filter, strides and paddings parameters.\nThe size of each dimension of the parameters is checked during infer-shape.\nIn order to ensure the equal length of sequence before and after convolution,\nit is necessary to fill the top and bottom of each sequence based on\ncontext_length, context_stride and context_start.\n\n    ",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(LoDTensor) the input(X) is a LodTensor, which supports variable-time length input sequence. The underlying tensor in this LoDTensor is a matrix with shape (T, N), where T is the total time steps in this mini-batch and N is the input_hidden_size.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "PaddingData",
   "comment" : "(Tensor, optional) the input(PaddingData) is an optional parameter, and it is learnable. This is a tensor with shape (P, N), where P is the top_pad + bottom_pad, N is the input_hidden_size. In order to ensure the equal length of sequence before and after convolution, it is necessary to fill the top and bottom of each sequence according to context_length, context_stride and context_start",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor) the input(Filter) is an learnable parameter.This is a tensor with shape (K, M), where K is the context_length * input_hidden_size, M is the output feature size.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDTensor) the output(Out) is a LodTensor, which support variable-time length output sequence. The underlying tensor in this LoDTensor is a matrix with shape (T, M), where, T is the total time steps in this mini-batch, M is the output feature size.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "paddingTrainable",
   "type" : "bool",
   "comment" : "(bool, default:false) the padding data of SequenceConvOp is trainable or not.",
   "generated" : 0
 }, { 
   "name" : "contextLength",
   "type" : "int",
   "comment" : "(int) the contextLength of SequenceConvOp is the height of the convolution kernel.",
   "generated" : 0
 }, { 
   "name" : "contextStart",
   "type" : "int",
   "comment" : "(int, default:0) the contextStart of SequenceConvOp represents the beginning of the convolution of the number of rows of sequence, which can be negative. The negative number means to pad contextStart time-steps of zeros or learnable parameters at the beginning of each instance. The positive number means to skip contextStart time-steps of each instance.",
   "generated" : 0
 }, { 
   "name" : "contextStride",
   "type" : "int",
   "comment" : "(int, default:1) the contextStride of SequenceConvOp represents the stride length of convolution kernel. Currently, SequenceConvOp only supportscontextStride=1.",
   "generated" : 0
 } ] 
},{
 "type" : "sequence_pool",
 "comment" : "\nSequence Pool Operator.\n\nThe SequencePoolOp pools features of all time-steps of each instance.\nIt supports six pooling types:\n1. AVERAGE: $$Out[i] = \\frac{\\sum_i X_i}{N}$$\n2. SUM:     $$Out[i] = \\sum_jX_{ij}$$\n3. SQRT:    $$Out[i] = \\frac{\\sum_jX_{ij}}{\\sqrt{len(X_i)}}$$\n4. LAST:    Out[i] = last instance in i-th sequence X[i]\n5. FIRST:   Out[i] = first instance in i-th sequence X[i]\n6. MAX:     $$Out[i] = max(X_i)$$\n\nThe following example explains how this works:\nFor a mini-batch of 3 variable-length sentences,\ncontaining 2, 3, and 2 time-steps:\n\nAssume X is a [7,M,N] LoDTensor, and X->lod()[0] = [0, 2, 5, 7], 7=2+3+2.\nBesides, for the sake of simplicity, we assume M=1 and N=1,\nand the value of X = [[1, 3], [2, 4, 6], [5, 1]].\n\nThus, Out is a [3,1,1] Tensor without LoD infomation.\nAnd for different pooltype, the value of Out is as follows:\n\n- AVERAGE: [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2\n- SUM: [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1\n- SQRT: [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2),\n           6.93=(2+4+6)/sqrt(3), 4.24=(5+1)/sqrt(2)\n- MAX: [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1)\n- LAST: [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)\n- FIRST: [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)\n\n    ",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(LoDTensor) The variable-length input of SequencePoolOp",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output of SequencePoolOp does not contain LoD infomation.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "MaxIndex",
   "comment" : "(Tensor<int>) This tensor is used for the sequence max-pooling to record the max indexes.",
   "duplicable" : 0,
   "intermediate" : 1
 } ], 
 "attrs" : [ 
 { 
   "name" : "pooltype",
   "type" : "string",
   "comment" : "(int, default AVERAGE) the pooling pooltype of SequencePoolOp.",
   "generated" : 0
 } ] 
},{
 "type" : "lstm",
 "comment" : "\nLong-Short Term Memory (LSTM) Operator.\n\nThe defalut implementation is diagonal/peephole connection\n(https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows:\n\n$$\ni_t = \\sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i) \\\\\n\nf_t = \\sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f) \\\\\n\n\\tilde{c_t} = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c) \\\\\n\no_t = \\sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o) \\\\\n\nc_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} \\\\\n\nh_t = o_t \\odot act_h(c_t)\n$$\n\nwhere the W terms denote weight matrices (e.g. $W_{xi}$ is the matrix\nof weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$\nare diagonal weight matrices for peephole connections. In our implementation,\nwe use vectors to reprenset these diagonal weight matrices. The b terms\ndenote bias vectors ($b_i$ is the input gate bias vector), $\\sigma$\nis the non-line activations, such as logistic sigmoid function, and\n$i, f, o$ and $c$ are the input gate, forget gate, output gate,\nand cell activation vectors, respectively, all of which have the same size as\nthe cell output activation vector $h$.\n\nThe $\\odot$ is the element-wise product of the vectors. $act_g$ and $act_h$\nare the cell input and cell output activation functions and `tanh` is usually\nused for them. $\\tilde{c_t}$ is also called candidate hidden state,\nwhich is computed based on the current input and the previous hidden state.\n\nSet `use_peepholes` False to disable peephole connection. The formula\nis omitted here, please refer to the paper\nhttp://www.bioinf.jku.at/publications/older/2604.pdf for details.\n\nNote that these $W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}$\noperations on the input $x_{t}$ are NOT included in this operator.\nUsers can choose to use fully-connect operator before LSTM operator.\n\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(LoDTensor) the first input is a LodTensor, which support variable-time length input sequence. The underlying tensor in this LoDTensor is a matrix with shape (T X 4D), where T is the total time steps in this mini-batch, D is the hidden size.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "H0",
   "comment" : "(Tensor, optional) the initial hidden state is an optional input. This is a tensor with shape (N x D), where N is the batch size and D is the hidden size.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "C0",
   "comment" : "(Tensor, optional) the initial cell state is an optional input. This is a tensor with shape (N x D), where N is the batch size. `H0` and `C0` can be NULL but only at the same time",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Weight",
   "comment" : "(Tensor) the learnable hidden-hidden weights. - The shape is (D x 4D), where D is the hidden size.  - Weight = {W_ch, W_ih, W_fh, W_oh}",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Bias",
   "comment" : "(Tensor) the learnable weights, which contains two parts: input-hidden bias weight and peephole connections weight if setting `use_peepholes` True. 1. `use_peepholes = False`  - The shape is (1 x 4D).  - Bias = {b_c, b_i, b_f, b_o}.2. `use_peepholes = True`  - The shape is (1 x 7D).  - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Hidden",
   "comment" : "(LoDTensor) the hidden state of LSTM operator. The shape is (T x D), and lod is the same with the `Input`.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Cell",
   "comment" : "(LoDTensor) the cell state of LSTM operator. The shape is (T x D), and lod is the same with the `Input`.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "BatchGate",
   "comment" : "(LoDTensor) This LoDTensor contains input gate, forget gate and output gate after the nonlinear computation. This LoDTensor has the same shape as the reorganized input, which is also be called batch input. The LoD size is 2. The first LoD is the batch offsets and the second LoD contains the indexes, which denote the position of reorganized sequence in the raw input.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "BatchCellPreAct",
   "comment" : "(LoDTensor) This LoDTensor is obtained in the forward and used in the backward.",
   "duplicable" : 0,
   "intermediate" : 1
 } ], 
 "attrs" : [ 
 { 
   "name" : "use_peepholes",
   "type" : "bool",
   "comment" : "(bool, defalut: True) whether to enable diagonal/peephole connections.",
   "generated" : 0
 }, { 
   "name" : "is_reverse",
   "type" : "bool",
   "comment" : "(bool, defalut: False) whether to compute reversed LSTM.",
   "generated" : 0
 }, { 
   "name" : "gate_activation",
   "type" : "string",
   "comment" : "(string, default: sigmoid)The activation for input gate, forget gate and output gate, `sigmoid` by default.",
   "generated" : 0
 }, { 
   "name" : "cell_activation",
   "type" : "string",
   "comment" : "(string, default: tanh)The activation for cell output, `tanh` by defalut.",
   "generated" : 0
 }, { 
   "name" : "candidate_activation",
   "type" : "string",
   "comment" : "(string, default: tanh)The activation for candidate hidden state, `tanh` by default.",
   "generated" : 0
 } ] 
},{
 "type" : "conv3d_transpose",
 "comment" : "\nConvolution3D Transpose Operator.\n\nThe convolution transpose operation calculates the output based on the input, filter\nand strides, paddings, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and output(Output) are in NCDHW format. Where N is batch size, C is the\nnumber of channels, D is the depth of the feature, H is the height of the feature,\nand W is the width of the feature.\nFilter(Input) is in MCDHW format. Where M is the number of input feature channels,\nC is the number of output feature channels, D is the depth of the filter,H is the\nheight of the filter, and W is the width of the filter.\nParameters(strides, paddings) are three elements. These three elements represent\ndepth, height and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:   \n  Input:\n       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{in}, C_{out}, D_f, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + D_f \\\\\n       H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + H_f \\\\\n       W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + W_f\n  $$\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution transpose operator.The format of input tensor is NCDHW. Where N is batch size, C is the number of channels, D is the depth of the feature, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution transpose operator.The format of the filter tensor is MCDHW, where M is the number of input feature channels, C is the number of output feature channels, D is the depth of the filter, H is the height of the filter, and W is the width of the filter.We enforce groups number == 1 and padding == 0 in the convolution3d transpose scenario.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution transpose operator.The format of output tensor is also NCDHW.Where N is batch size, C is the number of channels, D is the depth of the feature, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1, 1}), the strides{d_stride, h_stride, w_stride} of convolution transpose operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int> default:{0, 0, 0}), paddings(d_pad, h_pad, w_pad) of convolution transpose operator.",
   "generated" : 0
 } ] 
},{
 "type" : "conv2d_transpose",
 "comment" : "\nConvolution2D Transpose Operator.\n\nThe convolution transpose operation calculates the output based on the input, filter\nand strides, paddings, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and output(Output) are in NCHW format. Where N is batchsize, C is the\nnumber of channels, H is the height of the feature, and W is the width of the feature.\nFilter(Input) is in MCHW format. Where M is the number of input feature channels,\nC is the number of output feature channels, H is the height of the filter,\nand W is the width of the filter.\nParameters(strides, paddings) are two elements. These two elements represent height\nand width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       Input shape: $(N, C_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{in}, C_{out}, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + H_f \\\\\n       W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + W_f\n  $$\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution transpose operator. The format of input tensor is NCHW. Where N is batch size, C is the number of input channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution transpose operator. The format of the filter tensor is MCHW, where M is the number of input feature channels, C is the number of output feature channels,H is the height of the filter, and W is the width of the filter. We enforce groups number == 1 in the convolution transpose scenario.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution transpose operator. The format of output tensor is also NCHW.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1}), the strides(h_stride, w_stride) of convolution transpose operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int> default:{0, 0}), the paddings(h_pad, w_pad) of convolution transpose operator.",
   "generated" : 0
 } ] 
},{
 "type" : "gru",
 "comment" : "\nGRU Operator implements part calculations of the complete GRU as following:\n\n\\f[\nupdate \\ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\\\\nreset \\ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r)  \\\\\noutput \\ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\\\\noutput: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t)\n\\f]\n\n@note To implement the complete GRU, fully-connected operator must be used  \nbefore to feed xu, xr and xc as the Input of GRU operator.\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(LoDTensor) The first input is a LodTensor, which supports variable-time length input sequence. The underlying tensor in this LoDTenosr is a matrix with shape (T X 3D), where, T is the total time steps in this mini-batch, D is the hidden size.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "H0",
   "comment" : "(Tensor, optional) The initial hidden state is an optional input. This is a tensor with shape (N x D), where N is the batch size, D is the hidden size.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Weight",
   "comment" : "(Tensor) The learnable hidden-hidden weight matrix with shape (D x 3D), where D is the hidden size. The elements continuous in memory can be divided into two parts. The first part are weights of the update gate and reset gate with shape (D x 2D), and the second part are weights of output candidate with shape (D x D).",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Bias",
   "comment" : "(Tensor, optional) Bias vector with shape (1 x 3D) concating bias of the update gate, reset gate and output candidate.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "BatchGate",
   "comment" : "(LoDTensor) To compute with batches, sequence data will be reorganized into several successive batches each containing data from the same time step. The LoDTensor BatchGate contains the update gate, reset gate and output candidate values organized in batches. The LoD size is 2. The first LoD contains the batch offsets and the second LoD contains the indexes in the raw sequence data.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "BatchResetHiddenPrev",
   "comment" : "(LoDTensor) The reseted hidden state LoDTensor organized in batches. This LoDTensor is a matrix with shape (T X D) and has the same LoD with `BatchGate`.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "BatchHidden",
   "comment" : "(LoDTensor) The hidden state LoDTensor organized in batches.  This LoDTensor is a matrix with shape (T X D) and has the same LoD with `BatchGate`.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "Hidden",
   "comment" : "(LoDTensor) the hidden state LoDTensor organized in sequences. This LoDTensor is a matrix with shape (T X D) and has the same LoD with `BatchGate`.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "activation",
   "type" : "string",
   "comment" : "(string, default tanh) The activation type used for output candidate {h}_t.",
   "generated" : 0
 }, { 
   "name" : "gate_activation",
   "type" : "string",
   "comment" : "(string, default sigmoid) The activation type used in update gate and reset gate.",
   "generated" : 0
 }, { 
   "name" : "is_reverse",
   "type" : "bool",
   "comment" : "(bool, defalut: False) whether to compute reversed GRU.",
   "generated" : 0
 } ] 
},{
 "type" : "recurrent",
 "comment" : "\nStatic Length Recurrent Operator.\n\nThe static length recurrent operator can only operate on fixed size sequence\ndata, i.e. in each mini-batch, the sequence length of all inputs are the same.\n\n",
 "inputs" : [ 
 { 
   "name" : "inputs",
   "comment" : "rnn inputs",
   "duplicable" : 1,
   "intermediate" : 0
 }, { 
   "name" : "initial_states",
   "comment" : "rnn initial states",
   "duplicable" : 1,
   "intermediate" : 0
 }, { 
   "name" : "parameters",
   "comment" : "Parameters are used by step block as its input. However, the input is not a sequence tensor. Every time step, each operator in step block just use the parameter directly.",
   "duplicable" : 1,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "outputs",
   "comment" : "The output sequence of RNN. The sequence length must be same.",
   "duplicable" : 1,
   "intermediate" : 0
 }, { 
   "name" : "step_scopes",
   "comment" : "StepScopes contain all local variables in each time step.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "ex_states",
   "type" : "string array",
   "comment" : "The ex-state variable names.\nThe ex-state means the state value in the ex-timestep or the previous time step\n[ex_states, states, initial_states@GRAD] must be the same order",
   "generated" : 0
 }, { 
   "name" : "states",
   "type" : "string array",
   "comment" : "The state variable names. [ex_states, states, initial_states@GRAD] must be the same order",
   "generated" : 0
 }, { 
   "name" : "step_block",
   "type" : "block id",
   "comment" : "The step block inside RNN",
   "generated" : 0
 }, { 
   "name" : "reverse",
   "type" : "bool",
   "comment" : "Calculate RNN reversely or not.\nBy default reverse=False\n\nAssume the input data is [A, B, C, D]\n\nif reverse is False:\n  the computation of RNN is like\n      A          B          C         D\n      |          |          |         |\n      v          v          v         v\n     rnn -----> rnn -----> rnn ----> rnn\n      |          |          |         |\n      v          v          v         v\n      o          o          o         o\n\nif reverse is True\n  the computation of RNN is like\n      A          B          C         D\n      |          |          |         |\n      v          v          v         v\n     rnn <----- rnn <----- rnn <---- rnn\n      |          |          |         |\n      v          v          v         v\n      o          o          o         o\n",
   "generated" : 0
 }, { 
   "name" : "is_train",
   "type" : "bool",
   "comment" : "",
   "generated" : 0
 } ] 
},{
 "type" : "save",
 "comment" : "\nSave operator\n\nThis operator will serialize and write a tensor variable to file on disk.\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor ) Input tensor to be saved",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [  ], 
 "attrs" : [ 
 { 
   "name" : "overwrite",
   "type" : "bool",
   "comment" : "(boolean, default true)Overwrite the output file if exist",
   "generated" : 0
 }, { 
   "name" : "file_path",
   "type" : "string",
   "comment" : "(string)The \"file_path\" where the variable will be saved.",
   "generated" : 0
 } ] 
},{
 "type" : "load",
 "comment" : "\nLoad Operator.\n\nLoad operator will load a tensor variable from disk file.\n\n",
 "inputs" : [  ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The tensor need to be loaded",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "file_path",
   "type" : "string",
   "comment" : "(string) Variable will be loaded from \"file_path\".",
   "generated" : 0
 } ] 
},{
731 732
 "type" : "accuracy",
 "comment" : "\nAccuracy Operator. \n\nIt will print accuracy rate for classification.\nThe accuracy is calculated as follows:\n\n$$accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples}$$\n\nBoth the input Out and Label can carry the LoD (Level of Details)\ninformation, or not. But the output only shares the LoD information \nwith the input Out(Inference).\n\n",
733 734 735
 "inputs" : [ 
 { 
   "name" : "Out",
736
   "comment" : "The network output of topk (inferences)",
737 738 739 740
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Indices",
741
   "comment" : "The the network output of topk (indices)",
742 743 744 745
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Label",
746
   "comment" : "Label of the training data",
747 748 749 750 751
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
752 753
   "name" : "Accuracy",
   "comment" : "The accuracy of current batch",
754 755 756
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
757 758 759 760 761 762 763 764 765 766 767
   "name" : "Correct",
   "comment" : "The correct samples count of current batch",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Total",
   "comment" : "The samples count of current batch",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994
},{
 "type" : "hard_sigmoid",
 "comment" : "\nHardSigmoid Activation Operator.\n\nSegment-wise linear approximation of sigmoid(https://arxiv.org/abs/1603.00391), \nwhich is much faster than sigmoid.\n\n$y = \\max(0, \\min(1, slope * x + shift))$\n\nThe slope should be positive. The offset can be either positive or negative.\nThe default slope and shift are set according to the above reference.\nIt is recommended to use the defaults for this activation.\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of HardSigmoid operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of HardSigmoid operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "slope",
   "type" : "float",
   "comment" : "Slope for linear approximation of sigmoid",
   "generated" : 0
 }, { 
   "name" : "offset",
   "type" : "float",
   "comment" : "Offset for linear approximation of sigmoid",
   "generated" : 0
 } ] 
},{
 "type" : "cond",
 "comment" : "\nSample Dependent Conditional Operator.\n\nGiven Cond[i] as a 1/0 vector to indicate true/false:\nOut[i] = subnet_true[i], if Cond[i] == true\nOut[i] = subnet_false[i], if Cond[i] == false\n\n",
 "inputs" : [ 
 { 
   "name" : "Cond",
   "comment" : "The condition, which is a bool vector",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Xs",
   "comment" : "Inputs of Subnets",
   "duplicable" : 1,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Outs",
   "comment" : "Outputs of Cond_Op after merge",
   "duplicable" : 1,
   "intermediate" : 0
 }, { 
   "name" : "SubScopes",
   "comment" : "sub scopes for true and false branches",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "IndexTensors",
   "comment" : "Index Tensors contains indices for true/false",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "max_pool2d_with_index",
 "comment" : "\nMaxPool2d Operator.\n\nThe maxPooling2d with index operation calculates the output and the mask\nbased on the input, ksize, strides, and paddings parameters. Input(X) and\noutput(Out, Mask) are in NCHW format, where N is batch size, C is the\nnumber of channels, H is the height of the feature, \nand W is the width of the feature.\nParameters(ksize, strides, paddings) are two elements.\nThese two elements represent height and width, respectively.\nThe input(X) size and output(Out, Mask) size may be different.\n\nExample:\n  Input:\n       X shape: $(N, C, H_{in}, W_{in})$\n  Output:\n       Out shape: $(N, C, H_{out}, W_{out})$\n       Mask shape: $(N, C, H_{out}, W_{out})$\n  Where\n       $$\n       H_{out} = \\frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\\\\n       W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1\n       $$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor of pooling operator. The format of input tensor is NCHW, where N is batch size, C is the number of channels, H is the height of the image, and W is the width of the image.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of pooling operator. The format of output tensor is also NCHW, where N is batch size, C is the number of channels, H is the height of the image and W is the width of the image.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Mask",
   "comment" : "(Tensor) The Mask tensor of pooling operator.The format of output tensor is also NCHW, where N is batch size, C is the number of channels, H is the height of the image, and W is the width of the image. It represents the index in the current feature map.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "ksize",
   "type" : "int array",
   "comment" : "(vector<int>) The pooling window size(height, width) of pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 }, { 
   "name" : "global_pooling",
   "type" : "bool",
   "comment" : "(bool, default:false) Whether to use the global pooling. If global_pooling = true, ksize and paddings will be ignored.",
   "generated" : 0
 }, { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int>, default {1, 1}), strides(height, width) of pooling operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int>, default:{0, 0}), paddings(height, width) of pooling operator. If global_pooling = true, paddings and will be ignored.",
   "generated" : 0
 } ] 
},{
 "type" : "thresholded_relu",
 "comment" : "\nThresholdedRelu Activation Operator.\n\n$$\ny = \\begin{cases} \n    x, \\text{if } x > threshold \\\\\n    0,  \\text{otherwise}\n    \\end{cases}\n$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of ThresholdedRelu operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of ThresholdedRelu operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "threshold",
   "type" : "float",
   "comment" : "The threshold location of activation",
   "generated" : 0
 } ] 
},{
 "type" : "hard_shrink",
 "comment" : "\nHardShrink Activation Operator.\n\n$$\ny = \\begin{cases} \n    x, \\text{if } x > \\lambda \\\\\n    x, \\text{if } x < -\\lambda \\\\\n    0,  \\text{otherwise}\n    \\end{cases}\n$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of HardShrink operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of HardShrink operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "threshold",
   "type" : "float",
   "comment" : "The value of threshold for HardShrink",
   "generated" : 0
 } ] 
},{
 "type" : "relu6",
 "comment" : "\nRelu6 Activation Operator.\n\n$y = \\min(\\max(0, x), 6)$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Relu6 operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of Relu6 operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "threshold",
   "type" : "float",
   "comment" : "The threshold value of Relu6",
   "generated" : 0
 } ] 
},{
 "type" : "elu",
 "comment" : "\nELU Activation Operator.\n\nApplies the following element-wise computation on the input according to\nhttps://arxiv.org/abs/1511.07289.\n\n$y = \\max(0, x) + \\min(0, \\alpha * (e^x - 1))$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of ELU operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of ELU operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "alpha",
   "type" : "float",
   "comment" : "The alpha value of ELU",
   "generated" : 0
 } ] 
},{
 "type" : "leaky_relu",
 "comment" : "\nLeakyRelu Activation Operator.\n\n$y = \\max(x, \\alpha * x)$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of LeakyRelu operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of LeakyRelu operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "alpha",
   "type" : "float",
   "comment" : "The small negative slope",
   "generated" : 0
 } ] 
},{
995 996
 "type" : "softsign",
 "comment" : "\nSoftsign Activation Operator.\n\n$$y = \\frac{x}{1 + |x|}$$\n\n",
997 998 999
 "inputs" : [ 
 { 
   "name" : "X",
1000
   "comment" : "Input of Softsign operator",
1001 1002 1003 1004 1005
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
   "name" : "Y",
   "comment" : "Output of Softsign operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "logical_xor",
 "comment" : "logical_xor Operator\n\nIt operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = (X || Y) \\, \\&\\& \\, !(X \\&\\& Y)$$\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(LoDTensor) Left hand operand of logical_xor operator",
1019 1020 1021
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1022 1023
   "name" : "Y",
   "comment" : "(LoDTensor) Right hand operand of logical_xor operator",
1024 1025 1026
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
1027
 "outputs" : [ 
1028
 { 
1029 1030 1031 1032 1033 1034
   "name" : "Out",
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is $$Out = (X || Y) \\, \\&\\& \\, !(X \\&\\& Y)$$",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
1035
},{
1036 1037
 "type" : "softmax",
 "comment" : "\nSoftmax Operator.\n\nThe input of the softmax operator is a 2-D tensor with shape N x K (N is the\nbatch_size, K is the dimension of input feature). The output tensor has the\nsame shape as the input tensor.\n\nFor each row of the input tensor, the softmax operator squashes the\nK-dimensional vector of arbitrary real values to a K-dimensional vector of real\nvalues in the range [0, 1] that add up to 1.\nIt computes the exponential of the given dimension and the sum of exponential\nvalues of all the other dimensions in the K-dimensional vector input.\nThen the ratio of the exponential of the given dimension and the sum of\nexponential values of all the other dimensions is the output of the softmax\noperator.\n\nFor each row $i$ and each column $j$ in Input(X), we have:\n    $$Y[i, j] = \\frac{\\exp(X[i, j])}{\\sum_j(exp(X[i, j])}$$\n\n",
1038 1039 1040
 "inputs" : [ 
 { 
   "name" : "X",
1041
   "comment" : "The input tensor of softmax. 2-D with shape [batch_size, input_feature_dimensions].",
1042 1043 1044 1045 1046
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1047 1048
   "name" : "Y",
   "comment" : "The normalized values with the same shape as X.",
1049 1050 1051 1052 1053
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
1054 1055
 "type" : "lod_array_length",
 "comment" : "\nLoDArrayLength Operator.\n\nThis operator obtains the length of lod tensor array:\n\n$$Out = len(X)$$\n\nNOTE: The output is a CPU Tensor since the control variable should be only in\nCPU and the length of LoDTensorArray should be used as control variables.\n\n",
1056 1057
 "inputs" : [ 
 { 
1058 1059
   "name" : "X",
   "comment" : "(LoDTensorArray) The input tensor array.",
1060 1061
   "duplicable" : 0,
   "intermediate" : 0
1062 1063 1064 1065 1066
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) 1x1 CPU Tensor of length, int64_t",
1067 1068
   "duplicable" : 0,
   "intermediate" : 0
1069 1070 1071 1072 1073 1074 1075 1076 1077
 } ], 
 "attrs" : [  ] 
},{
 "type" : "top_k",
 "comment" : "\nTop K operator\n\nIf the input is a vector (1d tensor), this operator finds the k largest \nentries in the vector and outputs their values and indices as vectors. \nThus values[j] is the j-th largest entry in input, and its index is indices[j].\n\nFor matrices, this operator computes the top k entries in each row. ",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input of Topk op",
1078 1079 1080 1081 1082
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1083 1084
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of Topk op",
1085 1086 1087
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1088 1089
   "name" : "Indices",
   "comment" : "(Tensor) The indices of Topk elements of input",
1090 1091 1092 1093 1094
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1095 1096 1097
   "name" : "k",
   "type" : "int",
   "comment" : "(int, default 1) Number of top elements to look for along the last dimension (along each row for matrices).",
1098 1099 1100
   "generated" : 0
 } ] 
},{
1101 1102
 "type" : "clip",
 "comment" : "\nClip Operator.\n\nThe clip operator limits the value of given input within an interval. The interval is\nspecified with arguments 'min' and 'max':\n\n$$\nOut = \\min(\\max(X, min), max)\n$$\n\n",
1103 1104 1105
 "inputs" : [ 
 { 
   "name" : "X",
1106
   "comment" : "(Tensor)The input of clip op.The number of dimensions must be between [1, 9].",
1107 1108 1109 1110 1111 1112
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1113
   "comment" : "(Tensor)The output of clip op with shape as input(X)",
1114 1115 1116 1117 1118
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1119
   "name" : "min",
1120
   "type" : "float",
1121 1122 1123 1124 1125 1126
   "comment" : "(float)Minimum value, under which element is replaced by min.",
   "generated" : 0
 }, { 
   "name" : "max",
   "type" : "float",
   "comment" : "(float)Maximum value, above which element is replaced by max",
1127 1128 1129
   "generated" : 0
 } ] 
},{
1130 1131
 "type" : "margin_rank_loss",
 "comment" : "\nMarginRankLoss Operator.\n\nThis operator measures the loss given a pair of training sample\n{`X1`, `X2`} and the `Label` with attribute `margin`, where `Label = +1` \nindicating X1 is ranked higher than `X2` and `Label = -1` otherwise. The loss \nis calculated as:\n\n$loss(X1, X2, Label) = \\max(0, -Label * (X1 - X2) + margin)$\n\nThe attribute `margin` here helps make the predictions more robust.\nDenote the item ranked higher as the positive sample, otherwise the negative \nsample. If the score of the two samples satisfies \n\n$positive sample - negative sample < margin$\n\nthe pair of samples will contribute to the final loss, which will backpropagate \nand train the ranking model to enlarge the difference between the two scores.\n\nFor batch input with size `batch_size`, `X1`, `X2` and `Label`\nall have the same shape [batch_size x 1].\n\n",
1132 1133
 "inputs" : [ 
 { 
1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145
   "name" : "X1",
   "comment" : "(2-D tensor with shape [batch_size x 1]) The score for one item X1 to be ranked, from pairwise ranking model.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "X2",
   "comment" : "(2-D tensor with shape [batch_size x 1]) The score for another item X2 to be ranked, from pairwise ranking model.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Label",
   "comment" : "(2-D tensor with shape [batch_size x 1]) The label indicating X1 ranked higher than X2 or not, can only be +1 or -1.",
1146 1147 1148 1149 1150
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1151 1152 1153 1154 1155
   "name" : "Activated",
   "comment" : "(2-D tensor with shape [batch_size x 1]) Intermediate tensor to indicate whether each element of Output(Out) is activated.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
1156
   "name" : "Out",
1157
   "comment" : "(2-D tensor with shape [batch_size x 1]) The output loss of MarginRankLoss operator.",
1158 1159 1160 1161 1162
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1163
   "name" : "margin",
1164
   "type" : "float",
1165
   "comment" : "(scalar, default 0) Margin for MarginRankLossOp.",
1166 1167 1168
   "generated" : 0
 } ] 
},{
1169 1170
 "type" : "mul",
 "comment" : "\nMul Operator. \n\nThis operator is used to perform matrix multiplication for input X and Y.\n\nThe equation is:\n\n    $$Out = X * Y$$\n\nBoth the input `X` and `Y` can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input `X`.\n\n",
1171 1172 1173
 "inputs" : [ 
 { 
   "name" : "X",
1174 1175 1176 1177 1178 1179
   "comment" : "The first input of mul op",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
   "comment" : "The second input of mul op",
1180 1181 1182 1183 1184 1185
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1186
   "comment" : "The output of mul op",
1187 1188 1189 1190 1191
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1192 1193 1194 1195 1196 1197 1198 1199
   "name" : "x_num_col_dims",
   "type" : "int",
   "comment" : "(int, default 1) mul_op can take tensors with more than two dimensions as input `X`,\n            in that case, tensors will be reshaped to a matrix. The matrix's first\n            dimension(column length) will be the product of tensor's last\n            `num_col_dims` dimensions, and the matrix's second dimension(row length)\n            will be the product of tensor's first `rank - num_col_dims` dimensions.\n        ",
   "generated" : 0
 }, { 
   "name" : "y_num_col_dims",
   "type" : "int",
   "comment" : "(int, default 1) mul_op can take tensors with more than two dimensions as input `Y`,\n             in that case, tensors will be reshaped to a matrix. Just like input `X`.\n        ",
1200 1201 1202
   "generated" : 0
 } ] 
},{
1203 1204
 "type" : "minus",
 "comment" : "\nMinus Operator.\n\nEquation:\n\n    $Out = X - Y$\n\nBoth the input `X` and `Y` can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input `X`.\n\n",
1205 1206 1207
 "inputs" : [ 
 { 
   "name" : "X",
1208 1209 1210 1211 1212 1213
   "comment" : "The left tensor of minus operator.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
   "comment" : "The right tensor of minus operator.",
1214 1215 1216 1217 1218 1219
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1220
   "comment" : "The output tensor of minus operator.",
1221 1222 1223 1224 1225
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
1226 1227
 "type" : "max_sequence_len",
 "comment" : "Calculate the max sequence length through lod_rank_table.",
1228 1229
 "inputs" : [ 
 { 
1230 1231
   "name" : "RankTable",
   "comment" : "The lod_rank_table.",
1232 1233 1234 1235 1236 1237
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1238
   "comment" : "The max sequence length.",
1239 1240 1241
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
1242
 "attrs" : [  ] 
1243
},{
1244 1245
 "type" : "multiplex",
 "comment" : "\nMultiplex Operator.\n\nMultiplex multiple tensors according to the index provided by the index tensor.\n\nIds: the index tensor.\nX[0 : N - 1]: the candidate tensors for output (N >= 2).\nFor each index i from 0 to batchSize - 1, the output is the i-th row of the\nthe (Ids[i])-th tensor.\n\nFor i-th row of the output tensor:\n\n$$y[i] = x_{k}[i]$$\n\nwhere `y` is the output tensor, `x_{k}` is the k-th input tensor,\nand `k = Ids[i]`.\n\n",
1246 1247
 "inputs" : [ 
 { 
1248 1249
   "name" : "Ids",
   "comment" : "The index tensor of multiplex operator.",
1250 1251
   "duplicable" : 0,
   "intermediate" : 0
1252 1253 1254 1255 1256
 }, { 
   "name" : "X",
   "comment" : "The candidate tensors of multiplex operator.",
   "duplicable" : 1,
   "intermediate" : 0
1257 1258 1259
 } ], 
 "outputs" : [ 
 { 
1260 1261
   "name" : "Out",
   "comment" : "The output tensor of multiplex operator.",
1262 1263 1264 1265 1266
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
1267 1268
 "type" : "pool3d_cudnn",
 "comment" : "\nPool3d Operator.\n\nThe pooling3d operation calculates the output based on\nthe input, pooling_type, ksize, strides, and paddings parameters.\nInput(X) and output(Out) are in NCDHW format, where N is batch\nsize, C is the number of channels, and D, H and W are the depth, height and\nwidth of the feature, respectively. Parameters(ksize, strides, paddings) \nare three elements. These three elements represent depth, height and \nwidth, respectively. The input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       X shape: $(N, C, D_{in}, H_{in}, W_{in})$\n  Output:\n       Out shape: $(N, C, D_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       D_{out} = \\frac{(D_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\\\\n       H_{out} = \\frac{(H_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1 \\\\\n       W_{out} = \\frac{(W_{in} - ksize[2] + 2 * paddings[2])}{strides[2]} + 1\n  $$\n\n",
1269 1270
 "inputs" : [ 
 { 
1271 1272
   "name" : "X",
   "comment" : "(Tensor) The input tensor of pooling operator. The format of input tensor is NCDHW, where N is batch size, C is the number of channels, and D, H and W is the depth, height and width of the feature, respectively.",
1273 1274 1275 1276 1277
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1278 1279
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of pooling operator.The format of output tensor is also NCDHW, where N is batch size, C is the number of channels, and D, H and W is the depth, height and width of the feature, respectively.",
1280 1281 1282 1283 1284
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1285 1286 1287
   "name" : "pooling_type",
   "type" : "string",
   "comment" : "(string) Pooling type, can be \"max\" for max-pooling and \"avg\" for average-pooling.",
1288 1289
   "generated" : 0
 }, { 
1290 1291 1292
   "name" : "ksize",
   "type" : "int array",
   "comment" : "(vector<int>) The pooling window size(depth, height, width) of pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
1293 1294
   "generated" : 0
 }, { 
1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307
   "name" : "global_pooling",
   "type" : "bool",
   "comment" : "(bool, default false) Whether to use the global pooling. If global_pooling = true, ksize and paddings wille be ignored.",
   "generated" : 0
 }, { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int>, default {1,1,1}) Strides(depth, height, width) of the pooling operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int>, default {0,0,0}), paddings(depth, height, width) of pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
1308 1309 1310
   "generated" : 0
 } ] 
},{
1311 1312
 "type" : "pow",
 "comment" : "\nPow Activation Operator.\n\n$y = x^{factor}$\n\n",
1313 1314 1315
 "inputs" : [ 
 { 
   "name" : "X",
1316
   "comment" : "Input of Pow operator",
1317 1318 1319 1320 1321
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1322 1323
   "name" : "Y",
   "comment" : "Output of Pow operator",
1324 1325 1326 1327 1328
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1329 1330 1331
   "name" : "factor",
   "type" : "float",
   "comment" : "The exponential factor of Pow",
1332 1333 1334
   "generated" : 0
 } ] 
},{
1335 1336
 "type" : "l1_norm",
 "comment" : "\nL1 Norm Operator.\n\nComputes the L1 norm of a tensor.\n\n$$Out = \\sum{|X|}$$\n\n",
1337 1338 1339
 "inputs" : [ 
 { 
   "name" : "X",
1340
   "comment" : "(Tensor) The input of l1_norm op.",
1341 1342 1343 1344 1345 1346
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1347
   "comment" : "(Scalar) The output of l1_norm op.",
1348 1349 1350
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
1351
 "attrs" : [  ] 
1352
},{
1353 1354
 "type" : "sqrt",
 "comment" : "\nSqrt Activation Operator.\n\n$y = \\sqrt{x}$\n\n",
1355 1356 1357
 "inputs" : [ 
 { 
   "name" : "X",
1358
   "comment" : "Input of Sqrt operator",
1359 1360 1361 1362 1363
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1364 1365
   "name" : "Y",
   "comment" : "Output of Sqrt operator",
1366 1367 1368 1369 1370
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
1371 1372
 "type" : "lookup_table",
 "comment" : "\nLookup Table Operator.\n\nThis operator is used to perform lookups on the parameter W,\nthen concatenated into a dense tensor.\n\nThe input Ids can carry the LoD (Level of Details) information,\nor not. And the output only shares the LoD information with input Ids.\n\n",
1373 1374
 "inputs" : [ 
 { 
1375 1376 1377 1378 1379 1380 1381
   "name" : "W",
   "comment" : "An input represents embedding tensors, which is a learnable parameter.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Ids",
   "comment" : "An input with type int32 or int64 contains the ids to be looked up in W. Ids must be a column vector with rank = 2. The 2nd dimension size must be 1.",
1382 1383 1384 1385 1386 1387
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1388
   "comment" : "The lookup results, which have the same type as W.",
1389 1390 1391 1392 1393
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1394 1395 1396
   "name" : "is_sparse",
   "type" : "bool",
   "comment" : "(boolean, default false) Sparse update",
1397 1398 1399
   "generated" : 0
 } ] 
},{
1400 1401 1402
 "type" : "positive_negative_pair",
 "comment" : "\n        PositiveNegativePairOp can be used to evaluate Learning To Rank(LTR) \n        model performance. \n        Within some context, e.g. the \"query\", a LTR model generates scores\n        for a list of items, which gives a partial order of the items.\n        PositiveNegativePairOp takes a list of reference rank order \n        (Input(\"Label\")) and the model generated scores (Input(Score)) as \n        inputs and counts the pairs that ranked correctly and incorrectly.\n",
 "inputs" : [ 
1403
 { 
1404 1405
   "name" : "Score",
   "comment" : "(Tensor, float) Model Score on an item (with respect to QueryID). It's a 2-D tensor with shape [batch_size, depth], where the column specified by the attribute \"column\" is used as item score.",
1406 1407 1408
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1409 1410 1411 1412
   "name" : "Label",
   "comment" : "(Tensor, float) Label of an item (with repsect to QueryId). It's a 2-D tensor with shape [batch_size, 1].",
   "duplicable" : 0,
   "intermediate" : 0
1413
 }, { 
1414 1415 1416 1417
   "name" : "QueryID",
   "comment" : "(Tensor, int64) Query ID that indicates the context. Its shape should be the same as Label.",
   "duplicable" : 0,
   "intermediate" : 0
1418
 }, { 
1419 1420
   "name" : "AccumulatePositivePair",
   "comment" : "(float) Optional. The accumulated number of positive pairs over a stream of data. If provided, the output PositivePair will be initialized with this number rather than 0. it won't be modified in place.",
1421 1422 1423
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435
   "name" : "AccumulateNegativePair",
   "comment" : "(float) Optional. The accumulated number of negative pairs over a stream of data. If provided, the output NegativePair will be initialized with this number rather than 0. it won't be modified in place.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "AccumulateNeutralPair",
   "comment" : "(float) Optional. The accumulated number of neutral pairs over a stream of data. If provided, the output NeutralPair will be initialized with this number rather than 0. it won't be modified in place.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Weight",
   "comment" : "(float) Optional. Weight of current item. If specified, its shape should be the same as Label, and the meaning of the output changes from numbers of pairs to the total sum of pairs' weights. Weight of a pair of items is the average of their weights.",
1436 1437 1438 1439 1440
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452
   "name" : "PositivePair",
   "comment" : "(float) Number of positive pairs, i.e. the pairs of items that are ranked correctly.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "NegativePair",
   "comment" : "(float) Number of negative pairs, i.e. the pairs of items that are ranked incorrectly.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "NeutralPair",
   "comment" : "(float) Number of neutral pairs, i.e. the pairs of items that have the same score.",
1453 1454 1455 1456 1457
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1458
   "name" : "column",
1459
   "type" : "int",
1460
   "comment" : "(int, default -1) The column position of Score used to rank items in descending order. It must be in the range of [-rank(Score), rank(Score)). If `dim < 0`, the dim to reduce is `rank + dim`. Noting that reducing on the first dim will make the LoD info lost.",
1461 1462 1463
   "generated" : 0
 } ] 
},{
1464 1465
 "type" : "proximal_gd",
 "comment" : "\nProximalGD Operator.\n\nOptimizer that implements the proximal gradient descent algorithm:\n\n$$\nprox\\_param = param - learning\\_rate * grad \\\\\nparam = sign(prox\\_param) / (1 + learning\\_rate * l2) *\n        \\max(|prox\\_param| - learning\\_rate * l1, 0)\n$$        \n\nThe paper that proposed Proximal Gradient Descent:\n(http://papers.nips.cc/paper/3793-efficient-learning-using-forward-backward-splitting.pdf)\n\n",
1466 1467
 "inputs" : [ 
 { 
1468 1469
   "name" : "Param",
   "comment" : "(Tensor, default Tensor<float>) Input parameter value that has to be updated.",
1470 1471 1472
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1473 1474
   "name" : "Grad",
   "comment" : "(Tensor, default Tensor<float>) Input gradient of the parameter.",
1475 1476
   "duplicable" : 0,
   "intermediate" : 0
1477 1478 1479 1480 1481 1482 1483
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor, default Tensor<float>) The learning rate should be a tensor of size 1.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
1484
 { 
1485 1486
   "name" : "ParamOut",
   "comment" : "(Tensor) Output updated parameter value.",
1487 1488 1489 1490 1491
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1492 1493 1494
   "name" : "l1",
   "type" : "float",
   "comment" : "(float, default 0.0) L1 regularization strength.",
1495 1496
   "generated" : 0
 }, { 
1497 1498 1499
   "name" : "l2",
   "type" : "float",
   "comment" : "(float, default 0.0) L2 regularization strength.",
1500 1501 1502
   "generated" : 0
 } ] 
},{
1503 1504
 "type" : "prelu",
 "comment" : "\nPRelu Operator.\n\nThe equation is:\n\n$$\nf(x) =\n\\begin{cases}\n\\alpha * x, \\quad  \\text{if} \\ x < 0 \\\\\nx,         \\qquad  \\text{if} \\ x >= 0\n\\end{cases}\n$$\n\nThe input `X` can carry the LoD (Level of Details) information,\nor not. And the output shares the LoD information with input `X`.\n\n",
1505 1506 1507
 "inputs" : [ 
 { 
   "name" : "X",
1508
   "comment" : "The input tensor of prelu operator.",
1509 1510 1511
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1512 1513
   "name" : "Alpha",
   "comment" : "The alpha weight of prelu operator.",
1514 1515 1516 1517 1518 1519
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1520
   "comment" : "The output tensor of prelu operator.",
1521 1522 1523
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
1524
 "attrs" : [  ] 
1525
},{
1526 1527
 "type" : "proximal_adagrad",
 "comment" : "\nProximal Adagrad Optimizer.\n\nOptimizer that implements the proximal adagrad algorithm:\n\n$$\nmoment = moment + grad * grad \\\\\nprox\\_param = param - learning\\_rate * grad * (1 / \\sqrt{moment}) \\\\\nparam = sign(prox\\_param) / (1 + learning\\_rate * l2) *\n        \\max(|prox\\_param| - learning\\_rate * l1 , 0)\n$$\n\nThe paper that proposed Proximal GD: \n(http://papers.nips.cc/paper/3793-efficient-learning-using-forward-backward-splitting.pdf)\nHere, we use the adagrad learning rate as specified here: \n(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)\n\n",
1528 1529
 "inputs" : [ 
 { 
1530 1531
   "name" : "Param",
   "comment" : "(Tensor, default Tensor<float>) Input parameter that has to be updated.",
1532 1533 1534
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1535 1536
   "name" : "Moment",
   "comment" : "(Tensor, default Tensor<float>) Moment parameter that has to be updated.",
1537 1538 1539
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1540 1541 1542 1543 1544 1545 1546
   "name" : "Grad",
   "comment" : "(Tensor, default Tensor<float>) Input gradient of the parameter.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor, default Tensor<float>) The learning rate should be a tensor of size 1.",
1547 1548 1549 1550 1551
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1552 1553
   "name" : "ParamOut",
   "comment" : "(Tensor) Output updated parameter value.",
1554
   "duplicable" : 0,
1555
   "intermediate" : 0
1556
 }, { 
1557 1558
   "name" : "MomentOut",
   "comment" : "(Tensor) Output updated moment value.",
1559 1560 1561 1562 1563
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1564
   "name" : "l1",
1565
   "type" : "float",
1566
   "comment" : "(float, default 0.0) L1 regularization strength.",
1567 1568
   "generated" : 0
 }, { 
1569 1570 1571 1572 1573
   "name" : "l2",
   "type" : "float",
   "comment" : "(float, default 0.0) L2 regularization strength.",
   "generated" : 0
 } ] 
1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592
},{
 "type" : "reciprocal",
 "comment" : "\nReciprocal Activation Operator.\n\n$$y = \\frac{1}{x}$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Reciprocal operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of Reciprocal operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
1593 1594
 "type" : "merge_lod_tensor",
 "comment" : "\n        Merge True and False branches of LoDTensor into a single Output,\n        with a mask at certain lod level. X is used to obtain complete\n        lod information. Please refer to SplitLoDTensorOp.",
1595 1596 1597
 "inputs" : [ 
 { 
   "name" : "X",
1598
   "comment" : "The input LoDTensor, contains complete lod information to construct the output",
1599 1600
   "duplicable" : 0,
   "intermediate" : 0
1601 1602 1603
 }, { 
   "name" : "Mask",
   "comment" : "A bool column vector which mask the input",
1604 1605 1606
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1607 1608
   "name" : "InTrue",
   "comment" : "The True branch to be merged",
1609 1610 1611
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1612 1613
   "name" : "InFalse",
   "comment" : "The False branch to be merged",
1614 1615 1616 1617 1618 1619
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1620
   "comment" : "The merged output LoDTensor",
1621 1622 1623
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
1624 1625 1626 1627 1628 1629 1630
 "attrs" : [ 
 { 
   "name" : "level",
   "type" : "int",
   "comment" : "(int) the specific lod level to rank.",
   "generated" : 0
 } ] 
1631
},{
1632 1633
 "type" : "precision_recall",
 "comment" : "\nPrecision Recall Operator.\n\nWhen given Input(Indices) and Input(Labels), this operator can be used\nto compute various metrics including:\n1. macro average precision\n2. macro average recall\n3. macro f1 score\n4. micro average precision\n5. micro average recall\n6. micro f1 score\n\nTo compute the above metrics, we need to do statistics for true positives,\nfalse positives and false negatives. Here the count of true negatives is not\nnecessary, but counting it may provide potential usage and the cost is\ntrivial, so the operator also provides the count of true negatives.\n\nWe define state as a 2-D tensor with shape [class_number, 4]. Each row of a\nstate contains statistic variables for corresponding class. Layout of each row\nis: TP(true positives), FP(false positives), TN(true negatives),\nFN(false negatives). If Input(Weights) is provided, TP, FP, TN, FN will be\ncalculated by given weight instead of the instance count.\n\nThis operator also supports metrics computing for cross-batch situation. To\nachieve this, Input(StatesInfo) should be provided. State of current batch\ndata will be accumulated to Input(StatesInfo) and Output(AccumStatesInfo)\nis the accumulation state.\n\nOutput(BatchMetrics) is metrics of current batch data while\nOutput(AccumStatesInfo) is metrics of accumulation data.\n\n",
1634 1635
 "inputs" : [ 
 { 
1636 1637 1638
   "name" : "MaxProbs",
   "comment" : "(Tensor, default Tensor<float>) A 2-D tensor with shape N x 1, where N is the batch size. Each row contains the max probability of an instance which computed by the previous top_k (k=1) operator.",
   "duplicable" : 0,
1639 1640
   "intermediate" : 0
 }, { 
1641 1642 1643
   "name" : "Indices",
   "comment" : "(Tensor, default Tensor<int>) A 2-D tensor with shape N x 1, where N is the batch size. Each row contains the corresponding index which computed by the previous top_k (k=1) operator.",
   "duplicable" : 0,
1644 1645
   "intermediate" : 0
 }, { 
1646 1647
   "name" : "Labels",
   "comment" : "(Tensor, default Tensor<int>) A 2-D tensor with shape N x 1, where N is the batch size. Each element is a label and the value should be in [0, class_number - 1].",
1648 1649
   "duplicable" : 0,
   "intermediate" : 0
1650 1651 1652
 }, { 
   "name" : "Weights",
   "comment" : "(Tensor, default Tensor<float>) A 2-D tensor with shape N x 1, where N is the batch size. This input is optional. If provided, weight of instance would be considered when computing metrics.",
1653 1654 1655
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1656 1657
   "name" : "StatesInfo",
   "comment" : "(Tensor, default Tensor<int>) A 2-D tensor with shape D x 4, where D is the number of classes. This input is optional. If provided, current state will be accumulated to this state and the accumulation state will be the output state.",
1658 1659 1660 1661 1662
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674
   "name" : "BatchMetrics",
   "comment" : "(Tensor, default Tensor<float>) A 1-D tensor with shape {6}. This output tensor contains metrics for current batch data. The layout is [macro average precision, macro average recall, macro f1 score, micro average precision, micro average recall, micro f1 score].",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "AccumMetrics",
   "comment" : "(Tensor, default Tensor<float>) A 1-D tensor with shape {6}. This output tensor contains metrics for accumulated data. The layout is [macro average precision, macro average recall, macro f1 score, micro average precision, micro average recall, micro f1 score].",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "AccumStatesInfo",
   "comment" : "(Tensor, default Tensor<float>) A 2-D tensor with shape D x 4, where D is equal to class number. This output tensor contains accumulated state variables used to compute metrics. The layout for each class is [true positives, false positives, true negatives, false negatives].",
1675 1676 1677 1678 1679
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1680 1681 1682
   "name" : "class_number",
   "type" : "int",
   "comment" : "(int) Number of classes to be evaluated.",
1683 1684 1685
   "generated" : 0
 } ] 
},{
1686 1687
 "type" : "conv3d_cudnn",
 "comment" : "\nConvolution3D Operator.\n\nThe convolution operation calculates the output based on the input, filter\nand strides, paddings, dilations, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and output(Output) are in NCDHW format, where N is batch\nsize, C is the number of channels,D is the depth of the feature, H is the height of\nthe feature, and W is the width of the feature.\nFilters(Input) is MCDHW format, where M is the number of output image channels,\nC is the number of input image channels, D is the depth of the filter,\nH is the height of the filter, and W is the width of the filter.\nParameters(strides, paddings, dilations) are three elements. These three elements\nrepresent depth, height and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{out}, C_{in}, D_f, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       D_{out}= \\frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{ strides[0]}+ 1 \\\\\n       H_{out}= \\frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{ strides[1]}+ 1 \\\\\n       W_{out}= \\frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{ strides[2]}+ 1\n  $$\n",
1688 1689
 "inputs" : [ 
 { 
1690 1691 1692 1693 1694 1695 1696
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution operator. The format of input tensor is NCDHW. Where N is batch size, C is the number of channels, D is the depth of the feature, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution operator. The format of the filter tensor is MCDHW, where M is the number of output image channels, C is the number of input image channels, D is the depth of the filter, H is the height of the filter, and W is the width of the filter.If the groups attribute is greater than 1, C equals the number of input image channels divided by the groups.",
1697 1698 1699 1700 1701
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1702 1703
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution operator.The format of output tensor is also NCDHW.",
1704 1705 1706 1707 1708
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1709
   "name" : "strides",
1710
   "type" : "int array",
1711
   "comment" : "(vector<int>, default:{1, 1, 1}), the strides(d_stride, h_stride, w_stride) of convolution operator.",
1712 1713
   "generated" : 0
 }, { 
1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int>, default:{0, 0, 0}), the paddings(d_pad, h_pad, w_pad) of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "groups",
   "type" : "int",
   "comment" : "(int default:1), the groups number of the convolution operator. According to grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels.",
   "generated" : 0
 }, { 
   "name" : "dilations",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1, 1}), the dilations(d_dilation, h_dilation, w_dilation) of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "workspace_size_MB",
   "type" : "int",
   "comment" : "workspace size for cudnn, in MB, workspace is a section of GPU memory which will be allocated/freed each time the operator runs, larger workspace size can increase performance but also requires better hardware. This size should be chosen carefully.",
1732 1733 1734
   "generated" : 0
 } ] 
},{
1735 1736
 "type" : "sigmoid_cross_entropy_with_logits",
 "comment" : "\nSigmoidCrossEntropyWithLogits Operator.\n\nThis measures the element-wise probability error in classification tasks\nin which each class is independent. This can be thought of as predicting labels\nfor a data-point, where labels are not mutually exclusive.\nFor example, a news article can be about politics, technology or sports\nat the same time or none of these.\n\nThe logistic loss is given as follows:\n\n       $$loss = -Labels * \\log(\\sigma(X)) - (1 - Labels) * \\log(1 - \\sigma(X))$$\n\nWe know that $$\\sigma(X) = (1 / (1 + \\exp(-X)))$$. By substituting this we get:\n\n       $$loss = X - X * Labels + \\log(1 + \\exp(-X))$$\n\nFor stability and to prevent overflow of $$\\exp(-X)$$ when X < 0,\nwe reformulate the loss as follows:\n\n       $$loss = \\max(X, 0) - X * Labels + \\log(1 + \\exp(-|X|))$$\n\nBoth the input `X` and `Labels` can carry the LoD (Level of Details) information.\nHowever the output only shares the LoD with input `X`.\n\n",
1737 1738 1739
 "inputs" : [ 
 { 
   "name" : "X",
1740
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor with shape N x D, where N is the batch size and D is the number of classes. This input is a tensor of logits computed by the previous  operator. Logits are unscaled log probabilities given as log(p/(1-p)).",
1741 1742 1743
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1744 1745
   "name" : "Label",
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor of the same type and shape as X. This input is a tensor of probabalistic labels for each logit",
1746 1747 1748 1749 1750
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1751 1752
   "name" : "Out",
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor with shape N x D  of elementwise logistic losses.",
1753 1754
   "duplicable" : 0,
   "intermediate" : 0
1755 1756 1757 1758 1759 1760 1761 1762 1763 1764
 } ], 
 "attrs" : [  ] 
},{
 "type" : "fill",
 "comment" : "Fill operator\n\nFill an tensor with `value` and `shape`. The type of the tensor is specify by\n`dtype`.\n",
 "inputs" : [  ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDTensor) The output tensor.",
1765 1766 1767 1768 1769
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780
   "name" : "value",
   "type" : "float array",
   "comment" : "The float values of tensor, which are flatten in row major",
   "generated" : 0
 }, { 
   "name" : "shape",
   "type" : "int array",
   "comment" : "The shape of output tensor",
   "generated" : 0
 }, { 
   "name" : "dtype",
1781
   "type" : "int",
1782 1783 1784 1785 1786 1787
   "comment" : "The data type of output tensor, Default is float",
   "generated" : 0
 }, { 
   "name" : "force_cpu",
   "type" : "bool",
   "comment" : "Whether the output tensor must be at CPU memory or not. Default is false.",
1788 1789 1790
   "generated" : 0
 } ] 
},{
1791 1792
 "type" : "rnn_memory_helper",
 "comment" : "",
1793 1794
 "inputs" : [ 
 { 
1795 1796
   "name" : "X",
   "comment" : "",
1797 1798 1799 1800 1801 1802
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1803
   "comment" : "",
1804 1805 1806
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
1807
 "attrs" : [ 
1808
 { 
1809 1810 1811 1812 1813
   "name" : "dtype",
   "type" : "int",
   "comment" : "(int, default 5 (FP32)) Output data type",
   "generated" : 0
 } ] 
1814
},{
1815 1816
 "type" : "pad",
 "comment" : "\nPad Operator.\n\nPad input into output, as specified by paddings and pad_value. \nThe input should be a k-D tensor(k > 0 and k < 7). As an example:\n\nGiven:\n\nX = [[1, 2],\n     [3, 4]],\n\npaddings = [0, 1, 1, 2],\n\nand\n\npad_value = 0,\n\nwe have:\n\nOut = [[0, 1, 2, 0, 0]\n       [0, 3, 4, 0, 0]\n       [0, 0, 0, 0, 0]]\n\n",
1817 1818 1819
 "inputs" : [ 
 { 
   "name" : "X",
1820
   "comment" : "The input of pad op. The input should be a k-D tensor(k > 0 and k < 7)",
1821 1822 1823 1824 1825
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1826 1827
   "name" : "Out",
   "comment" : "The output of pad op. A tensor with the same shape as X.",
1828 1829 1830 1831 1832
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1833 1834 1835
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int>) A list<int> to describe the padding rules for each dimension. For 2-D image tensor, paddings=[0, 1, 2, 3] means padding 0 row to top, 1 row to bottom, 2 columns to left and 3 columns to right. Size of paddings should be equal to 2 * dimension size of the input tensor.",
1836 1837
   "generated" : 0
 }, { 
1838
   "name" : "pad_value",
1839
   "type" : "float",
1840
   "comment" : "(float, default 0.0) The value to fill the padded areas.",
1841 1842 1843
   "generated" : 0
 } ] 
},{
1844 1845
 "type" : "lstm_unit",
 "comment" : "\nLstm Unit Operator\n\nEquation:\n\n$$\ni, f, o, j = split(X) \\\\\nC = C_{prev} * sigm(f + forget\\_bias) + sigm(i) * tanh(j) \\\\\nH = C * sigm(o)\n$$\n\n",
1846 1847
 "inputs" : [ 
 { 
1848 1849
   "name" : "X",
   "comment" : "FC input before the non-linear activation.",
1850 1851 1852
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1853 1854
   "name" : "C_prev",
   "comment" : "The cell state tensor of last time-step in the Lstm Unit operator.",
1855 1856 1857 1858 1859
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
1860 1861
   "name" : "C",
   "comment" : "The cell tensor of Lstm Unit operator.",
1862 1863 1864
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
1865 1866
   "name" : "H",
   "comment" : "The hidden state tensor of Lstm Unit operator.",
1867 1868 1869 1870 1871
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1872
   "name" : "forget_bias",
1873
   "type" : "float",
1874
   "comment" : "(float, default 0.0) The forget bias of Lstm Unit.",
1875 1876 1877
   "generated" : 0
 } ] 
},{
1878 1879
 "type" : "squared_l2_norm",
 "comment" : "\nSquaredL2Norm Operator.\n\nComputes the squared L2 norm of a tensor.\n\n$$Out = \\sum_{i} X_{i}^2$$\n\n",
1880 1881 1882
 "inputs" : [ 
 { 
   "name" : "X",
1883
   "comment" : "(Tensor) The input of squared_l2_norm op.",
1884 1885 1886 1887 1888 1889
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1890
   "comment" : "(Scalar) The output of squared_l2_norm op.",
1891 1892 1893 1894 1895
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
1896 1897 1898
 "type" : "uniform_random",
 "comment" : "\nUniform random operator.\n\nThis operator initializes a tensor with random values sampled from a \nuniform distribution.\n\n",
 "inputs" : [  ], 
1899 1900 1901
 "outputs" : [ 
 { 
   "name" : "Out",
1902
   "comment" : "(Tensor) The output tensor of uniform random op",
1903 1904 1905 1906 1907
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
1908 1909 1910 1911 1912 1913
   "name" : "shape",
   "type" : "int array",
   "comment" : "(vector<int>) The shape of the output tensor",
   "generated" : 0
 }, { 
   "name" : "min",
1914
   "type" : "float",
1915
   "comment" : "(float, default -1.0) Minimum value of uniform random",
1916 1917
   "generated" : 0
 }, { 
1918 1919 1920
   "name" : "max",
   "type" : "float",
   "comment" : "(float, default 1.0) Maximun value of uniform random",
1921 1922 1923 1924
   "generated" : 0
 }, { 
   "name" : "seed",
   "type" : "int",
1925 1926 1927 1928 1929 1930
   "comment" : "(int, default 0) Random seed used for generating samples. 0 means use a seed generated by the system.",
   "generated" : 0
 }, { 
   "name" : "dtype",
   "type" : "int",
   "comment" : "(int, default 5(FP32)) Output tensor data type",
1931 1932 1933
   "generated" : 0
 } ] 
},{
1934 1935
 "type" : "sign",
 "comment" : "\nSign operator\n\n$$Out = X.sign()$$\n",
1936 1937 1938
 "inputs" : [ 
 { 
   "name" : "X",
1939
   "comment" : "(Tensor) Input tensor of sign operator.",
1940 1941 1942 1943 1944 1945
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1946
   "comment" : "(Tensor) Output tensor of sign operator.",
1947 1948 1949 1950 1951
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
1952 1953
 "type" : "roi_pool",
 "comment" : "\nROIPool operator\n\nROI Pooling for Faster-RCNN. The link below is a further introduction: \nhttps://stackoverflow.com/questions/43430056/what-is-roi-layer-in-fast-rcnn\n    ",
1954 1955 1956
 "inputs" : [ 
 { 
   "name" : "X",
1957 1958 1959 1960 1961 1962
   "comment" : "(Tensor), the input of ROIPoolOp. The format of input tensor is NCHW. Where N is batch size, C is the number of input channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "ROIs",
   "comment" : "(Tensor), ROIs (Regions of Interest) to pool over. should be a 2-D tensor of shape (num_rois, 5)given as [[batch_id, x1, y1, x2, y2], …]. Where batch_id is the id of the data, (x1, y1) is the top left coordinates, and (x2, y2) is the bottom right coordinates.",
1963 1964 1965 1966 1967 1968
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
1969
   "comment" : "(Tensor), The output of ROIPoolOp is a 4-D tensor with shape (num_rois, channels, pooled_h, pooled_w).",
1970 1971
   "duplicable" : 0,
   "intermediate" : 0
1972 1973 1974 1975 1976
 }, { 
   "name" : "Argmax",
   "comment" : "(Tensor), Argmaxes corresponding to indices in X used for gradient computation. Only output if arg “is_test” is false.",
   "duplicable" : 0,
   "intermediate" : 1
1977 1978 1979
 } ], 
 "attrs" : [ 
 { 
1980 1981 1982
   "name" : "spatial_scale",
   "type" : "float",
   "comment" : "(float, default 1.0), Multiplicative spatial scale factor to translate ROI coords from their input scale to the scale used when pooling.",
1983 1984
   "generated" : 0
 }, { 
1985 1986 1987
   "name" : "pooled_height",
   "type" : "int",
   "comment" : "(int, default 1), The pooled output height.",
1988 1989
   "generated" : 0
 }, { 
1990 1991 1992
   "name" : "pooled_width",
   "type" : "int",
   "comment" : "(int, default 1), The pooled output width.",
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039
   "generated" : 0
 } ] 
},{
 "type" : "conv2d_transpose_cudnn",
 "comment" : "\nConvolution2D Transpose Operator.\n\nThe convolution transpose operation calculates the output based on the input, filter\nand strides, paddings, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and output(Output) are in NCHW format. Where N is batchsize, C is the\nnumber of channels, H is the height of the feature, and W is the width of the feature.\nFilter(Input) is in MCHW format. Where M is the number of input feature channels,\nC is the number of output feature channels, H is the height of the filter,\nand W is the width of the filter.\nParameters(strides, paddings) are two elements. These two elements represent height\nand width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       Input shape: $(N, C_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{in}, C_{out}, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       H_{out} = (H_{in} - 1) * strides[0] - 2 * paddings[0] + H_f \\\\\n       W_{out} = (W_{in} - 1) * strides[1] - 2 * paddings[1] + W_f\n  $$\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution transpose operator. The format of input tensor is NCHW. Where N is batch size, C is the number of input channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution transpose operator. The format of the filter tensor is MCHW, where M is the number of input feature channels, C is the number of output feature channels,H is the height of the filter, and W is the width of the filter. We enforce groups number == 1 in the convolution transpose scenario.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution transpose operator. The format of output tensor is also NCHW.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1}), the strides(h_stride, w_stride) of convolution transpose operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int> default:{0, 0}), the paddings(h_pad, w_pad) of convolution transpose operator.",
   "generated" : 0
 }, { 
   "name" : "dilations",
   "type" : "int array",
   "comment" : "dilations of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "workspace_size_MB",
   "type" : "int",
   "comment" : "workspace size for cudnn, in MB, workspace is a section of GPU memory which will be allocated/freed each time the operator runs, larger workspace size can increase performance but also requires better hardward. This size should be carefully setted.",
   "generated" : 0
 } ] 
},{
2040 2041 2042 2043 2044 2045 2046 2047 2048
 "type" : "reduce_min",
 "comment" : "\n{ReduceOp} Operator.\n\nThis operator computes the min of input tensor along the given dimension. \nThe result tensor has 1 fewer dimension than the input unless keep_dim is true.\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor. Tensors with rank at most 6 are supported.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
2049 2050 2051
 "outputs" : [ 
 { 
   "name" : "Out",
2052
   "comment" : "(Tensor) The result tensor.",
2053 2054 2055 2056 2057
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2058
   "name" : "dim",
2059
   "type" : "int",
2060
   "comment" : "(int, default 0) The dimension to reduce. Must be in the range [-rank(input), rank(input)). If `dim < 0`, the dim to reduce is `rank + dim`. Note that reducing on the first dim will make the LoD info lost.",
2061 2062
   "generated" : 0
 }, { 
2063 2064 2065
   "name" : "keep_dim",
   "type" : "bool",
   "comment" : "(bool, default false) If true, retain the reduced dimension with length 1.",
2066 2067 2068
   "generated" : 0
 } ] 
},{
2069 2070
 "type" : "smooth_l1_loss",
 "comment" : "\nSmooth L1 Loss Operator.\n\nThis operator computes the smooth l1 loss for X and Y.\nThe operator takes the first dimension of X and Y as batch size.\nFor each instance, it computes the smooth l1 loss element by element first\nand then sums all the losses. So the shape of Out is [batch_size, 1].\n\nThe equation is:\n$$\nOut_{\\sigma}(X, Y)_i = \\begin{cases}\n0.5 * (\\sigma * (X_i - Y_i)) ^ 2\n\\quad |X_i - Y_i| \\lt \\frac{1} {{\\sigma} ^ 2} \\\\\n\\frac{|X_i - Y_i| - 0.5}{{\\sigma}^2},\n\\quad otherwise\n\\end{cases}\n$$\n\nIn the above equation, $Out_{\\sigma}(X, Y)_i$, $X_i$ and $Y_i$ represent the ith\nelement of Out, X and Y.\n\n",
2071 2072 2073
 "inputs" : [ 
 { 
   "name" : "X",
2074
   "comment" : "(Tensor, default Tensor<float>) A tensor with rank at least 2. The input value of smooth l1 loss op with shape [batch_size, dim1, ..., dimN].",
2075 2076 2077
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089
   "name" : "Y",
   "comment" : "(Tensor, default Tensor<float>) A tensor with rank at least 2. The target value of smooth l1 loss op with same shape as X.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "InsideWeight",
   "comment" : "(Tensor, default Tensor<float>) A tensor with rank at least 2. This input is optional and should have same shape with X. If provided, the result of (X - Y) will be multiplied by this tensor element by element.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "OutsideWeight",
   "comment" : "(Tensor, default Tensor<float>) A tensor with rank at least 2. This input is optional and should have same shape with X. If provided, the out smooth l1 loss will be multiplied by this tensor element by element.",
2090 2091 2092 2093 2094
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2095 2096
   "name" : "Diff",
   "comment" : "Intermediate variable to cache InsideWeight * (X - Y).",
2097
   "duplicable" : 0,
2098
   "intermediate" : 1
2099
 }, { 
2100 2101
   "name" : "Out",
   "comment" : "(Tensor, default Tensor<float>) A tensor with rank be 2. The output smooth l1 loss with shape [batch_size, 1].",
2102 2103 2104 2105 2106
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2107
   "name" : "sigma",
2108
   "type" : "float",
2109
   "comment" : "Hyper parameter of smooth l1 loss op.A float scalar with default value 3.0.",
2110 2111 2112
   "generated" : 0
 } ] 
},{
2113 2114
 "type" : "log_loss",
 "comment" : "\nLogLoss Operator.\n\nLog loss is a loss function used for binary classification. Log Loss quantifies\nthe accuracy of a classifier by penalising false classifications. Minimising the\nLog Loss is equivalent to maximising the accuracy of the classifier. We define\nPredicted as the values predicted by our model and Labels as the target ground\ntruth value. Log loss can evaluate how close the predicted values are to the\ntarget. The shapes of Predicted and Labels are both [batch_size, 1].\nThe equation is:\n\n$$\nLoss = - Labels * log(Predicted + \\epsilon) -\n        (1 - Labels) * log(1 - Predicted + \\epsilon)\n$$\n\n",
2115 2116
 "inputs" : [ 
 { 
2117 2118
   "name" : "Predicted",
   "comment" : "The input value (Predicted) of Log loss op.Predicted is a 2-D tensor with shape [batch_size, 1].",
2119 2120
   "duplicable" : 0,
   "intermediate" : 0
2121 2122 2123
 }, { 
   "name" : "Labels",
   "comment" : "The target value (Labels) of Log loss op.Labels is a 2-D tensor with shape [batch_size, 1].",
2124 2125 2126 2127 2128
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2129 2130
   "name" : "Loss",
   "comment" : "The output tensor with shape [batch_size, 1] which represents the log loss.",
2131 2132 2133 2134 2135
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2136
   "name" : "epsilon",
2137
   "type" : "float",
2138
   "comment" : "Epsilon in log loss.",
2139 2140 2141
   "generated" : 0
 } ] 
},{
2142 2143
 "type" : "unpool",
 "comment" : "\n        \"Input shape: $(N, C_{in}, H_{in}, W_{in})$\n        Output shape: $(N, C_{out}, H_{out}, W_{out})$\n        Where\n          $$\n            H_{out} = (H_{in}−1) * strides[0] − 2 * paddings[0] + ksize[0] \\\\\n            W_{out} = (W_{in}−1) * strides[1] − 2 * paddings[1] + ksize[1]\n          $$\n        Paper: http://www.matthewzeiler.com/wp-content/uploads/2017\n        /07/iccv2011.pdf\n        ",
2144 2145 2146
 "inputs" : [ 
 { 
   "name" : "X",
2147 2148 2149 2150 2151 2152
   "comment" : "(Tensor) The input tensor of unpool operator. The format of input tensor is NCHW. Where N is batch size, C is the number of channels, H and W is the height and width of feature.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Indices",
   "comment" : "(Tensor) The input tensor of the indices given out by MaxPool2d. The format of input tensor is NCHW. Where N is batch size, C is the number of channels, H and W is the height and width of feature.",
2153 2154 2155 2156 2157 2158
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2159
   "comment" : "(Tensor) The output tensor of unpool operator.The format of output tensor is also NCHW.Where N is batch size, C is the number of channels, H and W is the height and width of feature.",
2160 2161 2162 2163 2164
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2165 2166 2167
   "name" : "ksize",
   "type" : "int array",
   "comment" : "(vector), the unpooling window size(height, width) of unpooling operator.",
2168 2169
   "generated" : 0
 }, { 
2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector, default:{1, 1}), strides (height, width) of unpooling operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector defalut:{0,0}), paddings (height, width) of unpooling operator.",
   "generated" : 0
 }, { 
   "name" : "unpooling_type",
   "type" : "string",
   "comment" : "(string), unpooling type, can be \"max\" for max-unpooling ",
2183 2184 2185
   "generated" : 0
 } ] 
},{
2186 2187
 "type" : "transpose",
 "comment" : "\nTranspose Operator.\n\nThe input tensor will be permuted according to the axis values given.\nThe op functions similar to how numpy.transpose works in python.\nFor example:\n >> input = numpy.arange(6).reshape((2,3))\n >> input\n array([[0, 1, 2],\n        [3, 4, 5]])\n >> axis = [1, 0]\n >> output = input.transpose(axis)\n >> output\n array([[0, 3],\n        [1, 4],\n\t\t[2, 5]])\nSo, given a input tensor of shape(N, C, H, W) and the axis is {0, 2, 3, 1},\nthe output tensor shape will be (N, H, W, C)\n\n",
2188 2189 2190
 "inputs" : [ 
 { 
   "name" : "X",
2191
   "comment" : "(Tensor)The input tensor, tensors with rank at most 6 are supported",
2192 2193 2194 2195 2196 2197
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2198
   "comment" : "(Tensor)The output tensor",
2199 2200 2201 2202 2203
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2204 2205 2206
   "name" : "axis",
   "type" : "int array",
   "comment" : "(vector<int>)A list of values, and the size of the list should be the same with the input tensor rank, the tensor will permute the axes according the the values given",
2207 2208 2209
   "generated" : 0
 } ] 
},{
2210 2211
 "type" : "rnn_memory_helper_grad",
 "comment" : "",
2212 2213
 "inputs" : [ 
 { 
2214 2215
   "name" : "Out@GRAD",
   "comment" : "",
2216 2217 2218
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2219 2220
   "name" : "X",
   "comment" : "",
2221 2222 2223
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2224 2225
   "name" : "Out",
   "comment" : "",
2226 2227 2228 2229 2230
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2231 2232
   "name" : "X@GRAD",
   "comment" : "",
2233 2234 2235 2236 2237
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2238 2239 2240
   "name" : "dtype",
   "type" : "int",
   "comment" : "(int, default 5 (FP32)) Output data type",
2241 2242 2243
   "generated" : 0
 } ] 
},{
2244 2245
 "type" : "swish",
 "comment" : "\nSwish Activation Operator.\n\n$$y = \\frac{x}{1 + e^{- \\beta x}}$$\n\n",
2246 2247
 "inputs" : [ 
 { 
2248 2249
   "name" : "X",
   "comment" : "Input of Swish operator",
2250 2251 2252 2253 2254
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2255 2256
   "name" : "Y",
   "comment" : "Output of Swish operator",
2257 2258 2259 2260 2261
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2262 2263 2264
   "name" : "beta",
   "type" : "float",
   "comment" : "Constant beta of swish operator",
2265 2266
   "generated" : 0
 } ] 
2267
},{
2268 2269
 "type" : "is_empty",
 "comment" : "\nIsEmpty Operator which checks whether a tensor is empty.\n\nIt will just return product(tensor.ddims()) > 0;\n              ",
2270 2271 2272
 "inputs" : [ 
 { 
   "name" : "X",
2273 2274
   "comment" : "(Tensor) Tensor which is to be checked.",
   "duplicable" : 0,
2275 2276 2277 2278 2279
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2280
   "comment" : "(Tensor) a boolean Tensor that indicate empty or not.",
2281 2282 2283 2284 2285
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
2286 2287
 "type" : "reduce_max",
 "comment" : "\n{ReduceOp} Operator.\n\nThis operator computes the max of input tensor along the given dimension. \nThe result tensor has 1 fewer dimension than the input unless keep_dim is true.\n\n",
2288 2289 2290
 "inputs" : [ 
 { 
   "name" : "X",
2291 2292
   "comment" : "(Tensor) The input tensor. Tensors with rank at most 6 are supported.",
   "duplicable" : 0,
2293 2294 2295 2296 2297
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2298
   "comment" : "(Tensor) The result tensor.",
2299 2300 2301 2302 2303
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2304
   "name" : "dim",
2305
   "type" : "int",
2306 2307 2308 2309 2310 2311
   "comment" : "(int, default 0) The dimension to reduce. Must be in the range [-rank(input), rank(input)). If `dim < 0`, the dim to reduce is `rank + dim`. Note that reducing on the first dim will make the LoD info lost.",
   "generated" : 0
 }, { 
   "name" : "keep_dim",
   "type" : "bool",
   "comment" : "(bool, default false) If true, retain the reduced dimension with length 1.",
2312 2313
   "generated" : 0
 } ] 
2314
},{
2315 2316
 "type" : "rank_loss",
 "comment" : "\nRankLoss Operator.\n\nRankLoss operator for RankNet\n(http://icml.cc/2015/wp-content/uploads/2015/06/icml_ranking.pdf). \nRankNet is a pairwise ranking model with\none training sample consisting of a pair of doc A and B, and the label P\nindicating that A is ranked higher than B or not:\n\nP = {0, 1} or {0, 0.5, 1}, where 0.5 means no information about the rank of\nthe input pair.\n\nThe RankLoss operator takes three inputs: Left (o_i), Right (o_j) and Label\n(P_{i,j}), which represent the output score of RankNet for the two docs and \nthe label respectively, and yields the rank loss C_{i,j} using the following \nequation:\n\n$$\n  C_{i,j} = -\\tilde{P_{ij}} * o_{i,j} + \\log(1 + e^{o_{i,j}}) \\\\\n  o_{i,j} =  o_i - o_j  \\\\\n  \\tilde{P_{i,j}} = \\left \\{0, 0.5, 1 \\right \\} \\ or \\ \\left \\{0, 1 \\right \\}\n$$\n\nThe operator can take batch inputs with size batch_size (batch_size >= 1).\n\n",
2317 2318
 "inputs" : [ 
 { 
2319 2320
   "name" : "Label",
   "comment" : "(2-D Tensor with shape [batch_size x 1]) The label indicating A ranked higher than B or not.",
2321 2322 2323
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2324 2325 2326 2327 2328 2329 2330
   "name" : "Left",
   "comment" : "(2-D Tensor with shape [batch_size x 1]) The output of RankNet for doc A.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Right",
   "comment" : "(2-D Tensor with shape [batch_size x 1]) The output of RankNet for doc B.",
2331 2332 2333 2334 2335
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2336 2337
   "name" : "Out",
   "comment" : "(2-D Tensor with shape [batch_size x 1]) The output loss of RankLoss operator.",
2338
   "duplicable" : 0,
2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "greater_than",
 "comment" : "greater_than Operator\n\nIt operates element-wise on X and Y, and returns the Out. Each of them is a\nN-dim tensor. X and Y could be any type.  The each element of the Out tensor is\ncalculated by Out = X > Y\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(LoDTensor) the left hand operand of greater_than operator",
   "duplicable" : 0,
   "intermediate" : 0
2351
 }, { 
2352 2353
   "name" : "Y",
   "comment" : "(LoDTensor) the right hand operand of greater_than operator",
2354 2355 2356
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
2357
 "outputs" : [ 
2358
 { 
2359 2360 2361 2362 2363 2364
   "name" : "Out",
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is Out = X > Y",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
2365
},{
2366 2367
 "type" : "sequence_softmax",
 "comment" : "\nSequence Softmax Operator.\n\nSequenceSoftmaxOp computes the softmax activation among all time-steps for each\nsequence. The dimension of each time-step should be 1. Thus, the shape of\ninput Tensor can be either [N, 1] or [N], where N is the sum of the length\nof all sequences.\n\nThe algorithm works as follows:\n    for i-th sequence in a mini-batch:\n        $$Out(X[lod[i]:lod[i+1]], :) =\n            \\frac{\\exp(X[lod[i]:lod[i+1], :])}\n            {\\sum(\\exp(X[lod[i]:lod[i+1], :]))}$$\n\nFor example, for a mini-batch of 3 sequences with variable-length,\neach containing 2, 3, 2 time-steps, the lod of which is [0, 2, 5, 7],\nthen softmax will be computed among X[0:2, :], X[2:5, :], X[5:7, :]\nand N turns out to be 7.\n\n",
2368 2369
 "inputs" : [ 
 { 
2370 2371
   "name" : "X",
   "comment" : "(LoDTensor) 1-D or 2-D input LoDTensor with the 2-nd dimension of length 1.",
2372 2373 2374 2375 2376 2377
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2378
   "comment" : "(LoDTensor) 1-D or 2-D output LoDTensor with the 2-nd dimension of length 1.",
2379 2380 2381
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
2382
 "attrs" : [  ] 
2383
},{
2384 2385
 "type" : "momentum",
 "comment" : "\nMomentum Optimizer.\n\nThis optimizer has a flag for Nestrov Momentum.\nThe update equations are as follows:\n\n$$\nvelocity = mu * velocity + gradient \\\\\nif (use\\_nesterov):   \\\\\n  param = param - gradient * learning\\_rate + mu * velocity * learning\\_rate \\\\\nelse:   \\\\\n  param = param - learning\\_rate * velocity. \\\\\n$$\n\n",
2386 2387 2388
 "inputs" : [ 
 { 
   "name" : "Param",
2389
   "comment" : "(Tensor, default Tensor<float>) Input parameter that has to be updated",
2390 2391 2392 2393
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Grad",
2394
   "comment" : "(Tensor, default Tensor<float>) Input gradient of the parameter",
2395 2396 2397
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2398 2399
   "name" : "Velocity",
   "comment" : "(Tensor, default Tensor<float>) Input velocity (corresponding to the parameter) that has to be updated",
2400 2401 2402
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2403 2404
   "name" : "LearningRate",
   "comment" : "(Tensor, default Tensor<float>) Input learning rate",
2405 2406 2407 2408 2409 2410
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "ParamOut",
2411
   "comment" : "(Tensor) This output is updated parameter. It shared memory with Input(Param).",
2412 2413 2414
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2415 2416
   "name" : "VelocityOut",
   "comment" : "(Tensor) This output is updated velocity. It shared memory with Input(Velocity).",
2417 2418 2419 2420 2421
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2422
   "name" : "mu",
2423
   "type" : "float",
2424
   "comment" : "(float) Momentum coefficient",
2425 2426
   "generated" : 0
 }, { 
2427 2428 2429
   "name" : "use_nesterov",
   "type" : "bool",
   "comment" : "(bool, default false) Use Nesterov Momentum",
2430 2431 2432
   "generated" : 0
 } ] 
},{
2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462
 "type" : "scatter",
 "comment" : "\nScatter Operator.\n\nThis operator obtains output by updating the input on selected indices on the first axis:\n\n$$\nOut = Ref \\\\\nOut[Index] = Ref[Index] + Updates\n$$\n\n",
 "inputs" : [ 
 { 
   "name" : "Ref",
   "comment" : "The source input of scatter op",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Index",
   "comment" : "The index input of scatter op where Ref will be updated",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Updates",
   "comment" : "The updated value of updates op",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "The output of add op",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "shrink_rnn_memory",
 "comment" : "\n        In dynamic RNN, we are able to handle sequences of different lengths. \n        Because of the multiple lengths, the size of each step input can be \n        different, which may lead to a mismatching between the input of\n        the current step and the memory generated by the previous one. This \n        operator shrinks memory according to the size of the next step input, \n        to make sure that they can match each other.\n        ",
2463 2464 2465
 "inputs" : [ 
 { 
   "name" : "X",
2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476
   "comment" : "(LoDTensor) The RNN step memory to be shrinked.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "RankTable",
   "comment" : "(LoDRankTable) The lod_rank_table of dynamic RNN.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "I",
   "comment" : "(LoDTensor) The step index. The RNN step memory 'X' will be shrinked to match the size of the input of the index'th step.",
2477 2478 2479 2480 2481
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2482 2483
   "name" : "Out",
   "comment" : "(LoDTensor) The shrinked RNN step memory.",
2484 2485 2486 2487 2488
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
2489 2490
 "type" : "lod_reset",
 "comment" : "LoDReset operator\n\nReset LoD of Input(X) into a new one specified by Input(TargetLoD) or\nAttr(target_lod), or set LoD for Input(X) if it doesn't have one.\nCurrently the lod_reset operator only supports the reset of level 0 LoD.\nAt least one of Input(TargetLoD) and Attr(target_lod) must be set,\nand if both of them are set, Input(TargetLoD) will be chosen as the\ntarget LoD.\n\nAn example:\nGiven a float LoDTensor X with shape (6, 1), its transpose form represents\n\n    [1.0, 2.0, 3.0, 4.0, 5.0, 6.0],\n\nwith LoD = [[0, 2, 5, 6]] and the three (transposed) sequences look like\n\n    [1.0, 2.0], [3.0, 4.0, 5.0], [6.0].\n\nIf target LoD = [0, 4, 6], the lod_reset operator will reset the LoD and\nthe sequences that the LoDTensor Output(Out) contains becomes:\n\n    [1.0, 2.0, 3.0, 4.0], [5.0, 6.0].\n\n",
2491 2492
 "inputs" : [ 
 { 
2493 2494
   "name" : "X",
   "comment" : "(LoDTensor) The input tensor of lod_reset operator.",
2495 2496 2497
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2498 2499
   "name" : "TargetLoD",
   "comment" : "(Tensor, optional) The target level 0 LoD from Input().",
2500 2501 2502 2503 2504
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2505 2506
   "name" : "Out",
   "comment" : "(LoDTensor) The output tensor of lod_reset operator.",
2507 2508 2509 2510 2511
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2512
   "name" : "target_lod",
2513
   "type" : "int array",
2514
   "comment" : "The target level 0 LoD from Attr().",
2515 2516 2517
   "generated" : 0
 } ] 
},{
2518 2519
 "type" : "write_to_array",
 "comment" : "\nWriteToArray Operator.\n\nThis operator writes a LoDTensor to a LoDTensor array.\n\nAssume $T$ is LoDTensor, $i$ is the subscript of the array, and $A$ is the array. The\nequation is\n\n$$A[i] = T$$\n\n",
2520 2521
 "inputs" : [ 
 { 
2522 2523
   "name" : "X",
   "comment" : "(LoDTensor) the tensor will be written to tensor array",
2524 2525 2526
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2527 2528
   "name" : "I",
   "comment" : "(Tensor) the subscript index in tensor array. The number of element should be 1",
2529 2530 2531 2532 2533
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2534 2535
   "name" : "Out",
   "comment" : "(TensorArray) the tensor array will be written",
2536 2537 2538
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
2539 2540 2541 2542 2543
 "attrs" : [  ] 
},{
 "type" : "logical_or",
 "comment" : "logical_or Operator\n\nIt operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = X || Y$$\n",
 "inputs" : [ 
2544
 { 
2545 2546 2547 2548
   "name" : "X",
   "comment" : "(LoDTensor) Left hand operand of logical_or operator",
   "duplicable" : 0,
   "intermediate" : 0
2549
 }, { 
2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562
   "name" : "Y",
   "comment" : "(LoDTensor) Right hand operand of logical_or operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is $$Out = X || Y$$",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
2563
},{
2564 2565
 "type" : "seq_expand",
 "comment" : "\nSeq Expand Operator.\n\nThis operator expands input(X) according to LOD of input(Y).\nFollowing are cases to better explain how this works:\nCase 1:\n\nGiven 2-level a LoDTensor input(X)\n    X.lod = [[0,       2, 3],\n             [0, 1,    3, 4]]\n    X.data = [a, b, c, d]\n    X.dims = [4, 1]\nand input(Y)\n    Y.lod = [[0,    2,    4],\n             [0, 3, 6, 7, 8]]\nwith condition len(Y.lod[-1]) -1 == X.dims[0]\nthen we get 2-level LoDTensor\n    Out.lod = [[0,                2,    4],\n               [0,       3,       6, 7, 8]]\n    Out.data = [a, a, a, b, b, b, c, d]\n    Out.dims = [8, 1]\n\nCase 2:\n\nGiven a 0-level LoDTensor input(X)\n    X.data = [a, b, c]\n    X.lod = NULL\n    X.dims = [3, 1]\nand input(Y)\n    Y.lod = [[0, 2, 3, 6]]\nwith condition len(Y.lod[-1]) -1 == X.dims[0]\nthen we get 1-level LoDTensor\n    Out.lod = [[0,    2, 3,      6]]\n    Out.data = [a, a, b, c, c, c]\n    Out.dims = [6, 1]\n\nCase 3:\n\nGiven a 0-level LoDTensor input(X)\n    X.data = [[a, b], [c, d], [e, f]]\n    X.lod = NULL\n    X.dims = [3, 2]\nand input(Y)\n    Y.lod = [[0, 2, 3, 6]]\nwith condition len(Y.lod[-1]) -1 == X.dims[0]\nthen we get 1-level LoDTensor\n    Out.lod = [[0,           2,     3,                     6]]\n    Out.data = [[a,b], [a,b] [c,d], [e, f], [e, f], [e, f]]\n    Out.dims = [6, 2]\n\nCase 4:\n\nGiven 2-level a LoDTensor input(X)\n    X.lod = [[0,       2, 3],\n             [0, 1,    3, 4]]\n    X.data = [a, b, c, d]\n    X.dims = [4, 1]\nand input(Y)\n    Y.lod = [[0,    2,    4],\n             [0, 3, 6, 6, 8]]\nwith condition len(Y.lod[-1]) -1 == X.dims[0]\nthen we get 2-level LoDTensor\n    Out.lod = [[0,                2,    4],\n               [0,       3,       6, 6, 8]]\n    Out.data = [a, a, a, b, b, b, d, d]\n    Out.dims = [8, 1]\n\n\n",
2566 2567 2568
 "inputs" : [ 
 { 
   "name" : "X",
2569
   "comment" : "(Tensor or LoDTensor) The input(X) of this operator can be a LoDTensor or a base Tensor.",
2570 2571 2572
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2573 2574
   "name" : "Y",
   "comment" : "(LoDTensor)The reference input(Y) of seq_expand op.It must be a LoDTensor with k-level(k>0).The input(X) will be expanded according to LOD of input(Y).The element numbers of last level in input(Y) must be equal to dims[0] of input(X).",
2575 2576 2577 2578 2579
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599
   "name" : "Out",
   "comment" : "(LodTensor)The output of seq_expand op.The lod of output will be as same as input(Y)'s lod.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "scale",
 "comment" : "\nScale operator\n\n$$Out = scale*X$$\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) Input tensor of scale operator.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) Output tensor of scale operator.",
2600 2601 2602 2603 2604
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2605 2606 2607
   "name" : "scale",
   "type" : "float",
   "comment" : "(float, default 0)The scaling factor of the scale operator.",
2608 2609 2610
   "generated" : 0
 } ] 
},{
2611 2612
 "type" : "reduce_sum",
 "comment" : "\n{ReduceOp} Operator.\n\nThis operator computes the sum of input tensor along the given dimension. \nThe result tensor has 1 fewer dimension than the input unless keep_dim is true.\n\n",
2613 2614 2615
 "inputs" : [ 
 { 
   "name" : "X",
2616
   "comment" : "(Tensor) The input tensor. Tensors with rank at most 6 are supported.",
2617 2618 2619 2620 2621 2622
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2623
   "comment" : "(Tensor) The result tensor.",
2624 2625 2626 2627 2628
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2629 2630 2631
   "name" : "dim",
   "type" : "int",
   "comment" : "(int, default 0) The dimension to reduce. Must be in the range [-rank(input), rank(input)). If `dim < 0`, the dim to reduce is `rank + dim`. Note that reducing on the first dim will make the LoD info lost.",
2632 2633
   "generated" : 0
 }, { 
2634
   "name" : "keep_dim",
2635
   "type" : "bool",
2636
   "comment" : "(bool, default false) If true, retain the reduced dimension with length 1.",
2637 2638
   "generated" : 0
 } ] 
2639
},{
2640 2641
 "type" : "stanh",
 "comment" : "\nSTanh Activation Operator.\n\n$$y = b * \\frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$\n\n",
2642 2643 2644
 "inputs" : [ 
 { 
   "name" : "X",
2645
   "comment" : "Input of STanh operator",
2646 2647 2648 2649 2650 2651
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
2652
   "comment" : "Output of STanh operator",
2653 2654 2655 2656 2657
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2658
   "name" : "scale_a",
2659
   "type" : "float",
2660
   "comment" : "The scale parameter of a for the input",
2661 2662
   "generated" : 0
 }, { 
2663
   "name" : "scale_b",
2664
   "type" : "float",
2665
   "comment" : "The scale parameter of b for the input",
2666 2667 2668
   "generated" : 0
 } ] 
},{
2669 2670
 "type" : "adamax",
 "comment" : "\nAdamax Optimizer.\n\nWe implement the Adamax optimizer from Section 7 of the Adam\npaper: https://arxiv.org/abs/1412.6980. Adamax is a variant of the\nAdam algorithm based on the infinity norm.\n\nAdamax updates:\n\n$$\nmoment\\_out = \\beta_1 * moment + (1 - \\beta_1) * grad \\\\\ninf\\_norm\\_out = max(\\beta_2 * inf\\_norm + \\epsilon, |grad|) \\\\\nlearning\\_rate = \\frac{learning\\_rate}{1 - \\beta_{1\\_pow}} \\\\\nparam\\_out = param - learning\\_rate * \\frac{moment\\_out}{inf\\_norm\\_out}\n$$\n\nThe original paper does not have an epsilon attribute.\nHowever, it is added here for numerical stability to prevent the\ndivision by 0 error.\n\n",
2671 2672
 "inputs" : [ 
 { 
2673 2674
   "name" : "Param",
   "comment" : "(Tensor) Input parameter",
2675 2676 2677
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2678 2679
   "name" : "Grad",
   "comment" : "(Tensor) Input gradient",
2680 2681 2682
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699
   "name" : "LearningRate",
   "comment" : "(Tensor) Learning rate",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Moment",
   "comment" : "(Tensor) First moment",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "InfNorm",
   "comment" : "(Tensor) Input exponentially weighted infinity norm",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Beta1Pow",
   "comment" : "(Tensor) Input beta1 power accumulator",
2700 2701 2702 2703 2704
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716
   "name" : "ParamOut",
   "comment" : "(Tensor) Output parameter",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "MomentOut",
   "comment" : "(Tensor) Output first moment",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "InfNormOut",
   "comment" : "(Tensor) Output exponentially weighted infinity norm",
2717 2718 2719
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736
 "attrs" : [ 
 { 
   "name" : "beta1",
   "type" : "float",
   "comment" : "(float, default 0.9) Exponential decay rate for the 1st moment estimates.",
   "generated" : 0
 }, { 
   "name" : "beta2",
   "type" : "float",
   "comment" : "(float, default 0.999) exponential decay rate for the weighted infinity norm estimates.",
   "generated" : 0
 }, { 
   "name" : "epsilon",
   "type" : "float",
   "comment" : "(float, default 1.0e-8) Constant for numerical stability",
   "generated" : 0
 } ] 
2737
},{
2738 2739
 "type" : "tanh_shrink",
 "comment" : "\nTanhShrink Activation Operator.\n\n$$y = x - \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$\n\n",
2740 2741 2742
 "inputs" : [ 
 { 
   "name" : "X",
2743
   "comment" : "Input of TanhShrink operator",
2744 2745 2746 2747 2748
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
2749 2750
   "name" : "Y",
   "comment" : "Output of TanhShrink operator",
2751 2752 2753
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
2754
 "attrs" : [  ] 
2755
},{
2756 2757
 "type" : "mean",
 "comment" : "\nMean Operator.\n\nOut is a scalar which is the mean of all elements in X. \n\n",
2758 2759 2760
 "inputs" : [ 
 { 
   "name" : "X",
2761
   "comment" : "The input of mean op",
2762 2763 2764 2765 2766 2767
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2768
   "comment" : "The output of mean op",
2769 2770 2771 2772 2773
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
2774 2775
 "type" : "reduce_mean",
 "comment" : "\n{ReduceOp} Operator.\n\nThis operator computes the mean of input tensor along the given dimension. \nThe result tensor has 1 fewer dimension than the input unless keep_dim is true.\n\n",
2776 2777 2778
 "inputs" : [ 
 { 
   "name" : "X",
2779
   "comment" : "(Tensor) The input tensor. Tensors with rank at most 6 are supported.",
2780 2781 2782 2783 2784 2785
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2786
   "comment" : "(Tensor) The result tensor.",
2787 2788 2789 2790 2791
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823
   "name" : "dim",
   "type" : "int",
   "comment" : "(int, default 0) The dimension to reduce. Must be in the range [-rank(input), rank(input)). If `dim < 0`, the dim to reduce is `rank + dim`. Note that reducing on the first dim will make the LoD info lost.",
   "generated" : 0
 }, { 
   "name" : "keep_dim",
   "type" : "bool",
   "comment" : "(bool, default false) If true, retain the reduced dimension with length 1.",
   "generated" : 0
 } ] 
},{
 "type" : "pool2d_cudnn",
 "comment" : "\nPool2d Operator.\n\nThe pooling2d operation calculates the output based on\nthe input, pooling_type and ksize, strides, paddings parameters.\nInput(X) and output(Out) are in NCHW format, where N is batch size, C is the\nnumber of channels, H is the height of the feature, and W is the width of the feature.\nParameters(ksize, strides, paddings) are two elements.\nThese two elements represent height and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:   \n  Input:\n       X shape: $(N, C, H_{in}, W_{in})$\n  Output:\n       Out shape: $(N, C, H_{out}, W_{out})$\n  Where\n       $$ \n       H_{out} = \\frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\\\\n       W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1\n       $$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor of pooling operator. The format of input tensor is NCHW, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of pooling operator. The format of output tensor is also NCHW, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "pooling_type",
   "type" : "string",
   "comment" : "(string), pooling type, can be \"max\" for max-pooling and \"avg\" for average-pooling.",
2824 2825 2826 2827
   "generated" : 0
 }, { 
   "name" : "ksize",
   "type" : "int array",
2828
   "comment" : "(vector<int>) The pooling window size(height, width) of the pooling operator. If global_pooling = true, ksize and paddings will be ignored.",
2829 2830 2831 2832
   "generated" : 0
 }, { 
   "name" : "global_pooling",
   "type" : "bool",
2833
   "comment" : "(bool, default false) Whether to use the global pooling. If global_pooling = true, ksize and paddings will be ignored.",
2834 2835 2836 2837
   "generated" : 0
 }, { 
   "name" : "strides",
   "type" : "int array",
2838
   "comment" : "(vector<int>, default {1, 1}), strides(height, width) of pooling operator.",
2839 2840 2841 2842
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
2843
   "comment" : "(vector<int>, default {0,0}), paddings(height, width) of pooling operator.If global_pooling = true, paddings and ksize will be ignored.",
2844 2845 2846
   "generated" : 0
 } ] 
},{
2847 2848
 "type" : "lod_tensor_to_array",
 "comment" : "",
2849 2850 2851
 "inputs" : [ 
 { 
   "name" : "X",
2852
   "comment" : "",
2853 2854 2855
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2856 2857
   "name" : "RankTable",
   "comment" : "",
2858 2859 2860 2861 2862 2863
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2864 2865 2866 2867 2868 2869 2870
   "comment" : "",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "reshape",
2871
 "comment" : "\nReshape Operator.\n\nReshape Input(X) into the shape specified by Attr(shape).\n\nAn example:\nGiven a 2-D tensor X with 2 rows and 2 columns\n\n    [[1, 2], [3, 4]]\n\nand target shape = [1, 4], the reshape operator will transform\nthe tensor X into a 2-D tensor:\n\n    [[1, 2, 3, 4]]\n\n",
2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "The input tensor of reshape operator.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "The output tensor of reshape operator.",
2883 2884 2885 2886 2887 2888 2889
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "shape",
   "type" : "int array",
2890
   "comment" : "(vector<int>) Target shape of reshape operator.",
2891 2892 2893
   "generated" : 0
 } ] 
},{
2894 2895
 "type" : "modified_huber_loss",
 "comment" : "\nModified Huber Loss Operator.\n\nThis operator is used in binary classification problem. The shape of\ninput X and target Y are both [N, 1] and so is the shape of the output loss.\nSince target Y is not differentiable, calculating gradient for Y is illegal.\nThe formula of modified huber loss is:\n\n$$\nL(y, f(x)) = \n\\begin{cases}\n(\\max(0, 1 - yf(x)))^2,  \\text{if} \\  yf(x) >= -1    \\\\\n             -4yf(x),    \\quad \\text{otherwise}\n\\end{cases}\n$$\n\nMake sure the values of target label Y are in {0, 1} here. This operator will\nscale values of Y to {-1, +1} when computing losses and gradients.\n\n",
2896 2897 2898
 "inputs" : [ 
 { 
   "name" : "X",
2899
   "comment" : "The input tensor of modified huber loss op. X is 2-D tensor with shape [batch_size, 1].",
2900 2901 2902
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2903 2904
   "name" : "Y",
   "comment" : "The target labels of modified huber loss op. The shape of Y is the same as X. Values of Y must be 0 or 1.",
2905 2906
   "duplicable" : 0,
   "intermediate" : 0
2907 2908 2909 2910 2911 2912 2913
 } ], 
 "outputs" : [ 
 { 
   "name" : "IntermediateVal",
   "comment" : "Variable to save intermediate result which will be reused in backward processing.",
   "duplicable" : 0,
   "intermediate" : 1
2914
 }, { 
2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927
   "name" : "Out",
   "comment" : "Classification loss for X.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "elementwise_sub",
 "comment" : "\nLimited Elementwise Sub Operator.\n\nThe equation is:\n\n$Out = X - Y$\n\nX is a tensor of any dimension and the dimensions of tensor Y must be smaller than\nor equal to the dimensions of X. \n\nThere are two cases for this operator:\n1. The shape of Y is same with X;\n2. The shape of Y is a subset of X.\n\nFor case 2:\nY will be broadcasted to match the shape of X and axis should be \nthe starting dimension index for broadcasting Y onto X.\n\nexample:\n  shape(X) = (2, 3, 4, 5), shape(Y) = (,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (5,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1\n  shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0\n\nBoth the input X and Y can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input X.\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The first input tensor of elementwise op",
2928 2929 2930
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
2931 2932
   "name" : "Y",
   "comment" : "(Tensor) The second input tensor of elementwise op",
2933 2934 2935 2936 2937 2938
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2939
   "comment" : "The output of elementwise op",
2940 2941 2942
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
2943 2944
 "attrs" : [ 
 { 
2945
   "name" : "axis",
2946
   "type" : "int",
2947
   "comment" : "(int, default -1) The starting dimension index for broadcasting Y onto X",
2948 2949
   "generated" : 0
 } ] 
2950
},{
2951 2952
 "type" : "logical_and",
 "comment" : "logical_and Operator\n\nIt operates element-wise on X and Y, and returns the Out. X, Y and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = X \\&\\& Y$$\n",
2953 2954 2955
 "inputs" : [ 
 { 
   "name" : "X",
2956
   "comment" : "(LoDTensor) Left hand operand of logical_and operator",
2957 2958 2959 2960
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
2961
   "comment" : "(LoDTensor) Right hand operand of logical_and operator",
2962 2963 2964 2965 2966 2967
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is $$Out = X \\&\\& Y$$",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "logical_not",
 "comment" : "logical_not Operator\n\nIt operates element-wise on X, and returns the Out. X and Out are N-dim boolean tensors.\nEach element of Out is calculated by $$Out = !X$$\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(LoDTensor) Operand of logical_not operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is $$Out = !X$$",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "abs",
 "comment" : "\nAbs Activation Operator.\n\n$y = |x|$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Abs operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of Abs operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "square",
 "comment" : "\nSquare Activation Operator.\n\n$y = x^2$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Square operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of Square operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "increment",
 "comment" : "\nIncrement Operator.\n\nThe equation is: \n$$Out = X + step$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor of increment operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of increment operator.",
3041 3042 3043 3044 3045
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3046 3047 3048
   "name" : "step",
   "type" : "float",
   "comment" : "(float, default 1.0) The step size by which the input tensor will be incremented.",
3049 3050 3051
   "generated" : 0
 } ] 
},{
3052 3053
 "type" : "rmsprop",
 "comment" : "\nRmsprop Optimizer. \n\n$$\nMeanSquareOut = decay * MeanSquare + (1 - decay) * Grad * Grad \\\\\nMomentOut = momentum * Moment +\n            \\frac{LearningRate * Grad}{\\sqrt{MeanSquareOut + epsilon}} \\\\\nParamOut = Param -  MomentOut\n$$\n\nThe original slides that proposed Rmsprop: Slide 29 of\nhttp://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)\n\n",
3054 3055
 "inputs" : [ 
 { 
3056 3057
   "name" : "Param",
   "comment" : "(Tensor, default Tensor<float>) Input parameter value that has to be updated.",
3058 3059 3060
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077
   "name" : "MeanSquare",
   "comment" : "(Tensor, default Tensor<float>) The mean square value that gets updated.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor, default Tensor<float>) The learning rate should be a tensor of size 1.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Grad",
   "comment" : "(Tensor, default Tensor<float>) Input gradient of the parameter.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Moment",
   "comment" : "(Tensor, default Tensor<float>) The moment that gets updated.",
3078 3079 3080 3081 3082
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094
   "name" : "ParamOut",
   "comment" : "(Tensor) Output updated parameter value.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "MomentOut",
   "comment" : "(Tensor) Output updated moment.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "MeanSquareOut",
   "comment" : "(Tensor) Output Mean squared updated value.",
3095 3096 3097 3098 3099
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3100 3101 3102
   "name" : "epsilon",
   "type" : "float",
   "comment" : "(float, default 1e-10) Constant for numerical stability.",
3103 3104
   "generated" : 0
 }, { 
3105 3106 3107 3108 3109 3110 3111 3112
   "name" : "decay",
   "type" : "float",
   "comment" : "(float, default 0.9) Discounting factor for coming gradient.",
   "generated" : 0
 }, { 
   "name" : "momentum",
   "type" : "float",
   "comment" : "(float, default 0.0) Constant value.",
3113 3114 3115
   "generated" : 0
 } ] 
},{
3116 3117
 "type" : "elementwise_mul",
 "comment" : "\nLimited Elementwise Mul Operator.\n\nThe equation is:\n\n$Out = X \\odot\\ Y$\n\nX is a tensor of any dimension and the dimensions of tensor Y must be smaller than\nor equal to the dimensions of X. \n\nThere are two cases for this operator:\n1. The shape of Y is same with X;\n2. The shape of Y is a subset of X.\n\nFor case 2:\nY will be broadcasted to match the shape of X and axis should be \nthe starting dimension index for broadcasting Y onto X.\n\nexample:\n  shape(X) = (2, 3, 4, 5), shape(Y) = (,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (5,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1\n  shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0\n\nBoth the input X and Y can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input X.\n\n",
3118 3119
 "inputs" : [ 
 { 
3120 3121
   "name" : "X",
   "comment" : "(Tensor) The first input tensor of elementwise op",
3122 3123 3124
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3125 3126
   "name" : "Y",
   "comment" : "(Tensor) The second input tensor of elementwise op",
3127 3128 3129 3130 3131
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3132 3133
   "name" : "Out",
   "comment" : "The output of elementwise op",
3134 3135 3136 3137 3138
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3139 3140 3141
   "name" : "axis",
   "type" : "int",
   "comment" : "(int, default -1) The starting dimension index for broadcasting Y onto X",
3142 3143 3144
   "generated" : 0
 } ] 
},{
3145 3146
 "type" : "huber_loss",
 "comment" : "\nHuberLoss Operator.\n\nHuber loss is a loss function used in robust regression. We define X as the\ninput value and Y as the target value. Huber loss can evaluate the fitness of\nX to Y. Different from MSE loss, Huber loss is more robust for outliers. The\nshape of X and Y are [batch_size, 1]. The equation is:\n\n$$\nOut_{\\delta}(X, Y)_i =\n\\begin{cases}\n0.5 * (Y_i - X_i)^2,\n\\quad |Y_i - X_i| \\leq \\delta \\\\\n\\delta * (|Y_i - X_i| - 0.5 * \\delta),\n\\quad otherwise\n\\end{cases}\n$$\n\nIn the above equation, $Out_\\delta(X, Y)_i$, $X_i$ and $Y_i$ represent the ith\nelement of Out, X and Y.\n\n",
3147 3148
 "inputs" : [ 
 { 
3149 3150
   "name" : "X",
   "comment" : "The input value of huber loss op.X is a 2-D tensor with shape [batch_size, 1].",
3151 3152 3153
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3154 3155
   "name" : "Y",
   "comment" : "The target value of huber loss op.Y is a 2-D tensor with shape [batch_size, 1].",
3156 3157 3158 3159 3160
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3161 3162
   "name" : "Residual",
   "comment" : "Intermediate tensor to cache residual value between Y and X.The shape is same as Input(X) and will be reused in backward.",
3163
   "duplicable" : 0,
3164
   "intermediate" : 1
3165
 }, { 
3166 3167
   "name" : "Out",
   "comment" : "The output tensor with shape [batch_size, 1] which represents the huber loss.",
3168 3169 3170 3171 3172
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3173 3174 3175
   "name" : "delta",
   "type" : "float",
   "comment" : "Hyper parameter in huber loss.",
3176 3177 3178
   "generated" : 0
 } ] 
},{
3179 3180
 "type" : "sequence_slice",
 "comment" : "\nSequence slice operator\n\nThe operator crops a subsequence from given sequence with given start offset and subsequence length.\nIt only supports sequence (LoD Tensor with level number is 1).\n- Case:\n    X = [[a1, a2;\n        b1, b2;\n        c1, c2]\n       [d1, d2;\n        e1, e2]]\n    LoD(X) = {{0, 3, 5}}; Dims(X) = (5, 2)\n    Offset = [[0], [1]]; Length = [[2], [1]]\n\n    Out = [[a1, a2;\n            b1, b2]\n            [e1, e2]]\n    LoD(Out) = {{0, 2, 3}}; Dims(Out) = (3, 2)\nNOTE: The first dimension size of input, the size of offset and Length, should be equal. The offset start from 0.\n    ",
3181 3182
 "inputs" : [ 
 { 
3183 3184
   "name" : "X",
   "comment" : "(LoDTensor), the input of SequenceSliceOp.",
3185 3186 3187
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3188 3189 3190 3191 3192 3193 3194
   "name" : "Offset",
   "comment" : "(Tensor), a vector<int> to describe the offset of every input sequence for sub sequence item.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Length",
   "comment" : "(Tensor), a vector<int> to describe the length of every input sequence for sub sequence item.",
3195 3196 3197 3198 3199
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3200 3201
   "name" : "Out",
   "comment" : "(LoDTensor), the output of SequenceSliceOp.",
3202 3203 3204
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3205
 "attrs" : [  ] 
3206
},{
3207 3208
 "type" : "hinge_loss",
 "comment" : "\nHingeLoss Operator.\n\nLet x be a logit (prediction) and y be the actual label. The logit can\ntake any values from (-inf, inf), but the labels should be either -1 or 1.\nThen, the hinge loss is computed as follows:\n\n$$\nL_(x, y) = max(1 - y.x, 0) \n$$\n\nNote that the labels passed as input will have values as either 0 or 1.\n\n",
3209 3210
 "inputs" : [ 
 { 
3211 3212 3213 3214 3215 3216 3217
   "name" : "Logits",
   "comment" : "The input value (Logits) of Hinge loss op.Logits is a 2-D tensor with shape [batch_size, 1].",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Labels",
   "comment" : "The target value (Labels) of Hinge loss op.Labels is a 2-D tensor with shape [batch_size, 1].",
3218 3219 3220 3221 3222
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3223 3224
   "name" : "Loss",
   "comment" : "The output tensor with shape [batch_size, 1] which represents the hinge loss.",
3225 3226 3227 3228 3229
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
3230 3231
 "type" : "less_than",
 "comment" : "less_than Operator\n\nIt operates element-wise on X and Y, and returns the Out. Each of them is a\nN-dim tensor. X and Y could be any type.  The each element of the Out tensor is\ncalculated by Out = X < Y\n",
3232 3233 3234
 "inputs" : [ 
 { 
   "name" : "X",
3235
   "comment" : "(LoDTensor) the left hand operand of less_than operator",
3236 3237 3238 3239
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
3240
   "comment" : "(LoDTensor) the right hand operand of less_than operator",
3241 3242 3243 3244 3245 3246
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3247
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is Out = X < Y",
3248 3249 3250
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3251
 "attrs" : [  ] 
3252
},{
3253 3254
 "type" : "gru_unit",
 "comment" : "\nGRUUnit Operator implements partial calculations of the GRU unit as following:\n\n$$\nupdate \\ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\\\\nreset \\ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r)  \\\\\noutput \\ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\\\\noutput: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t)\n$$\n\nwhich is same as one time step of GRU Operator.\n\n@note To implement the complete GRU unit, fully-connected operator must be \nused before to feed xu, xr and xc as the Input of GRUUnit operator.\n\n",
3255 3256
 "inputs" : [ 
 { 
3257 3258
   "name" : "Input",
   "comment" : "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the input.",
3259
   "duplicable" : 0,
3260
   "intermediate" : 0
3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273
 }, { 
   "name" : "HiddenPrev",
   "comment" : "(Tensor) Matrix with shape [batch_size, frame_size] for the states of previous time step.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Weight",
   "comment" : "(Tensor) Weight matrix with shape [frame_size, frame_size * 3]. The elements continuous in memory can be divided into two parts. The first part are weights of the update gate and reset gate with shape [frame_size, frame_size * 2], and the second part are weights of output candidate with shape [frame_size, frame_size].",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Bias",
   "comment" : "(Tensor) Bias vector with shape [1, frame_size * 3] concatenating bias of the update gate, reset gate and output candidate.",
3274 3275 3276
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3277
 "outputs" : [ 
3278
 { 
3279 3280
   "name" : "Gate",
   "comment" : "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the output of update gate, reset gate and output candidate.",
3281
   "duplicable" : 0,
3282
   "intermediate" : 1
3283
 }, { 
3284 3285 3286 3287 3288 3289 3290
   "name" : "ResetHiddenPrev",
   "comment" : "(Tensor) Matrix with shape [batch_size, frame_size] for the reseted hidden state of previous time step.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "Hidden",
   "comment" : "(Tensor) The GRU hidden state of the current time step with shape [batch_size, frame_size].",
3291
   "duplicable" : 0,
3292 3293
   "intermediate" : 0
 } ], 
3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309
 "attrs" : [ 
 { 
   "name" : "activation",
   "type" : "int",
   "comment" : "(enum int, default tanh) The activation type used for output candidate {h}_t.",
   "generated" : 0
 }, { 
   "name" : "gate_activation",
   "type" : "int",
   "comment" : "(enum int, default sigmoid) The activation type used in update gate and reset gate.",
   "generated" : 0
 } ] 
},{
 "type" : "gaussian_random",
 "comment" : "\nGaussianRandom Operator.\n\nUsed to initialize tensors with gaussian random generator.\n\n",
 "inputs" : [  ], 
3310 3311 3312
 "outputs" : [ 
 { 
   "name" : "Out",
3313
   "comment" : "Output matrix of gaussian random op",
3314 3315 3316
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343
 "attrs" : [ 
 { 
   "name" : "shape",
   "type" : "int array",
   "comment" : "(vector<int>) The dimension of random tensor.",
   "generated" : 0
 }, { 
   "name" : "mean",
   "type" : "float",
   "comment" : "(float, default 0.0) mean of random tensor.",
   "generated" : 0
 }, { 
   "name" : "std",
   "type" : "float",
   "comment" : "(float, default 1.0) std of random tensor.",
   "generated" : 0
 }, { 
   "name" : "seed",
   "type" : "int",
   "comment" : "(int, default 0) Random seed of generator.0 means use system wide seed.",
   "generated" : 0
 }, { 
   "name" : "dtype",
   "type" : "int",
   "comment" : "(int, default 5(FP32)) Output data type.",
   "generated" : 0
 } ] 
3344
},{
3345 3346
 "type" : "fill_constant",
 "comment" : "\nFillConstantBatchSizeLike Operator.\n\nFill up a variable with specified constant value.\n\n",
3347 3348 3349 3350
 "inputs" : [  ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3351
   "comment" : "(Tensor) Tensor of specified shape will be filled with the specified value",
3352 3353 3354 3355 3356
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3357 3358 3359
   "name" : "dtype",
   "type" : "int",
   "comment" : "(int, default 5 (FP32)) Output data type",
3360 3361 3362 3363
   "generated" : 0
 }, { 
   "name" : "shape",
   "type" : "int array",
3364
   "comment" : "(vector<int>) The shape of the output",
3365 3366
   "generated" : 0
 }, { 
3367 3368 3369
   "name" : "value",
   "type" : "float",
   "comment" : "(float, default 0) The value to be filled",
3370 3371 3372 3373
   "generated" : 0
 }, { 
   "name" : "force_cpu",
   "type" : "bool",
3374
   "comment" : "(bool, default false) Force fill output variable to cpu memory. Otherwise, fill output variable to the running device",
3375 3376 3377
   "generated" : 0
 } ] 
},{
3378 3379
 "type" : "fill_zeros_like",
 "comment" : "\nFillZerosLike Operator.\n\nFill up a variable with zeros.\nThe output will have the same size as the input.\n\n",
3380 3381 3382
 "inputs" : [ 
 { 
   "name" : "X",
3383
   "comment" : "The input of fill-zeros-like op.",
3384 3385 3386 3387 3388
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3389 3390
   "name" : "Y",
   "comment" : "The variable will be filled up with zeros.",
3391 3392 3393 3394 3395
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
3396 3397
 "type" : "softmax_with_cross_entropy",
 "comment" : "\nSoftmax With Cross Entropy Operator.\n\nCross entropy loss with softmax is used as the output layer extensively. This\noperator computes the softmax normalized values for each row of the input\ntensor, after which cross-entropy loss is computed. This provides a more\nnumerically stable gradient.\n\nBecause this operator performs a softmax on logits internally, it expects\nunscaled logits. This operator should not be used with the output of\nsoftmax operator since that would produce incorrect results.\n\nWhen the attribute soft_label is set false, this operators expects mutually\nexclusive hard labels, each sample in a batch is in exactly one class with a\nprobability of 1.0. Each sample in the batch will have a single label.\n\nThe equation is as follows:\n\n1) Hard label (one-hot label, so every sample has exactly one class)\n\n$$Loss_j =  -\\text{Logit}_{Label_j} +\n\\log\\left(\\sum_{i=0}^{K}\\exp(\\text{Logit}_i)\\right),\nj = 1,..., K$$\n\n2) Soft label (each sample can have a distribution over all classes)\n\n$$Loss_j =  -\\sum_{i=0}^{K}\\text{Label}_i \\left(\\text{Logit}_i -\n\\log\\left(\\sum_{i=0}^{K}\\exp(\\text{Logit}_i)\\right)\\right),\nj = 1,...,K$$\n\n",
3398 3399
 "inputs" : [ 
 { 
3400 3401
   "name" : "Logits",
   "comment" : "(Tensor, default: Tensor<float>), The unscaled log probabilities which is a 2-D tensor with shape [N x K]. N is the batch_size, and K is the class number.",
3402 3403 3404
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3405 3406
   "name" : "Label",
   "comment" : "(Tensor) The ground truth which is a 2-D tensor. If soft_label is set to false, Label is a Tensor<int64> with shape [N x 1]. If soft_label is set to true, Label is a Tensor<float/double> with shape [N x K].",
3407 3408 3409 3410 3411
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3412 3413
   "name" : "Softmax",
   "comment" : "(Tensor, default: Tensor<float>), A 2-D tensor with shape [N x K]. The outputs value of softmax activation by given the input batch, which will be used in backward calculation.",
3414 3415 3416
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
3417 3418
   "name" : "Loss",
   "comment" : "(Tensor, default: Tensor<float>), A 2-D tensor. The cross entropy loss with shape [N x 1].",
3419 3420 3421 3422 3423
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3424 3425 3426
   "name" : "soft_label",
   "type" : "bool",
   "comment" : "(bool, default: false), A flag to indicate whether to interpretate the given labels as soft labels.",
3427 3428
   "generated" : 0
 } ] 
3429
},{
3430 3431
 "type" : "fill_constant_batch_size_like",
 "comment" : "\nFillConstantBatchSizeLike Operator.\n\nFill up a variable with specified constant value.\n\n",
3432 3433
 "inputs" : [ 
 { 
3434 3435
   "name" : "Input",
   "comment" : "(Tensor) Tensor whose dim_idx th dimension is used to specify the batch_size",
3436 3437 3438 3439 3440 3441
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3442
   "comment" : "(Tensor) Tensor of specified shape will be filled with the specified value",
3443 3444 3445
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3446 3447
 "attrs" : [ 
 { 
3448
   "name" : "dtype",
3449
   "type" : "int",
3450
   "comment" : "(int, default 5 (FP32)) Output data type",
3451 3452
   "generated" : 0
 }, { 
3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470
   "name" : "shape",
   "type" : "int array",
   "comment" : "(vector<int>) The shape of the output",
   "generated" : 0
 }, { 
   "name" : "input_dim_idx",
   "type" : "int",
   "comment" : "(int, default 0) The index of input's batch size dimension",
   "generated" : 0
 }, { 
   "name" : "output_dim_idx",
   "type" : "int",
   "comment" : "(int, default 0) The index of output's batch size dimension",
   "generated" : 0
 }, { 
   "name" : "value",
   "type" : "float",
   "comment" : "(float, default 0) The value to be filled",
3471 3472
   "generated" : 0
 } ] 
3473
},{
3474 3475
 "type" : "tanh",
 "comment" : "\nTanh Activation Operator.\n\n$$y = \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$\n\n",
3476 3477 3478
 "inputs" : [ 
 { 
   "name" : "X",
3479
   "comment" : "Input of Tanh operator",
3480 3481 3482 3483 3484 3485
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
3486
   "comment" : "Output of Tanh operator",
3487 3488 3489 3490 3491
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
3492 3493
 "type" : "feed",
 "comment" : "\nFeed Operator.\n\nIt should not be configured by users directly.\n\n",
3494 3495 3496
 "inputs" : [ 
 { 
   "name" : "X",
3497
   "comment" : "The input of feed op",
3498 3499 3500 3501 3502 3503
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3504
   "comment" : "The output of feed op",
3505 3506 3507 3508 3509
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3510
   "name" : "col",
3511
   "type" : "int",
3512
   "comment" : "(int) The column of feed",
3513 3514 3515
   "generated" : 0
 } ] 
},{
3516 3517
 "type" : "expand",
 "comment" : "\nExpand operator tiles the input by given times number. You should set times\nnumber for each dimension by providing attribute 'expand_times'. The rank of X\nshould be in [1, 6]. Please notice that size of 'expand_times' must be same with\nX's rank. Following is a using case:\n\nInput(X) is a 3-D tensor with shape [2, 3, 1]:\n\n        [\n           [[1], [2], [3]],\n           [[4], [5], [6]]\n        ]\n\nAttr(expand_times):  [1, 2, 2]\n\nOutput(Out) is a 3-D tensor with shape [2, 6, 2]:\n\n        [\n            [[1, 1], [2, 2], [3, 3], [1, 1], [2, 2], [3, 3]],\n            [[4, 4], [5, 5], [6, 6], [4, 4], [5, 5], [6, 6]]\n        ]\n\n",
3518 3519 3520
 "inputs" : [ 
 { 
   "name" : "X",
3521
   "comment" : "(Tensor, default Tensor<float>) A tensor with rank in [1, 6].X is the input tensor to be expanded.",
3522 3523 3524 3525 3526 3527
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3528
   "comment" : "(Tensor, default Tensor<float>) A tensor with rank in [1, 6].The rank of Output(Out) is same as Input(X) except that each dimension size of Output(Out) is equal to corresponding dimension size of Input(X) multiplying corresponding value of Attr(expand_times).",
3529 3530 3531
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3532 3533 3534 3535 3536 3537 3538
 "attrs" : [ 
 { 
   "name" : "expand_times",
   "type" : "int array",
   "comment" : "Expand times number for each dimension.",
   "generated" : 0
 } ] 
3539
},{
3540 3541
 "type" : "elementwise_div",
 "comment" : "\nLimited Elementwise Div Operator.\n\nThe equation is:\n\n$Out = X / Y$\n\nX is a tensor of any dimension and the dimensions of tensor Y must be smaller than\nor equal to the dimensions of X. \n\nThere are two cases for this operator:\n1. The shape of Y is same with X;\n2. The shape of Y is a subset of X.\n\nFor case 2:\nY will be broadcasted to match the shape of X and axis should be \nthe starting dimension index for broadcasting Y onto X.\n\nexample:\n  shape(X) = (2, 3, 4, 5), shape(Y) = (,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (5,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1\n  shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0\n\nBoth the input X and Y can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input X.\n\n",
3542 3543 3544
 "inputs" : [ 
 { 
   "name" : "X",
3545
   "comment" : "(Tensor) The first input tensor of elementwise op",
3546 3547 3548 3549
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
3550
   "comment" : "(Tensor) The second input tensor of elementwise op",
3551 3552 3553 3554 3555 3556
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3557
   "comment" : "The output of elementwise op",
3558 3559 3560
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3561 3562 3563 3564 3565 3566 3567
 "attrs" : [ 
 { 
   "name" : "axis",
   "type" : "int",
   "comment" : "(int, default -1) The starting dimension index for broadcasting Y onto X",
   "generated" : 0
 } ] 
3568
},{
3569 3570
 "type" : "elementwise_add",
 "comment" : "\nLimited Elementwise Add Operator.\n\nThe equation is:\n\n$Out = X + Y$\n\nX is a tensor of any dimension and the dimensions of tensor Y must be smaller than\nor equal to the dimensions of X. \n\nThere are two cases for this operator:\n1. The shape of Y is same with X;\n2. The shape of Y is a subset of X.\n\nFor case 2:\nY will be broadcasted to match the shape of X and axis should be \nthe starting dimension index for broadcasting Y onto X.\n\nexample:\n  shape(X) = (2, 3, 4, 5), shape(Y) = (,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (5,)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)\n  shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1\n  shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0\n\nBoth the input X and Y can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input X.\n\n",
3571 3572
 "inputs" : [ 
 { 
3573 3574
   "name" : "X",
   "comment" : "(Tensor) The first input tensor of elementwise op",
3575 3576 3577
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3578 3579
   "name" : "Y",
   "comment" : "(Tensor) The second input tensor of elementwise op",
3580 3581 3582 3583 3584
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3585 3586
   "name" : "Out",
   "comment" : "The output of elementwise op",
3587 3588 3589 3590 3591
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3592
   "name" : "axis",
3593
   "type" : "int",
3594
   "comment" : "(int, default -1) The starting dimension index for broadcasting Y onto X",
3595 3596 3597
   "generated" : 0
 } ] 
},{
3598 3599
 "type" : "cross_entropy",
 "comment" : "\nCrossEntropy Operator.\n\nIt supports both standard cross-entropy and soft-label cross-entropy loss\ncomputation.\n1) One-hot cross-entropy:\n    soft_label = false, Label[i, 0] indicates the class index for sample i:\n\n                $Y[i] = -\\log(X[i, Label[i]])$\n\n2) Soft-label cross-entropy:\n    soft_label = true, Label[i, j] indicates the soft label of class j\n    for sample i:\n\n                $Y[i] = \\sum_j{-Label[i, j] * log(X[i, j])}$\n\n   Please make sure that in this case the summuation of each row of Label\n   equals one.\n\n3) One-hot cross-entropy with vecterized Input(Label):\n     As a special case of 2), when each row of Input(Label) has only one\n     non-zero element (equals 1), soft-label cross-entropy degenerates to a\n     one-hot cross-entropy with one-hot label representation.\n\nBoth the input X and Label can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input X.\n\n",
3600 3601 3602
 "inputs" : [ 
 { 
   "name" : "X",
3603 3604 3605 3606 3607 3608
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor with shape N x D, where N is the batch size and D is the number of classes. This input is a probability computed by the previous operator, which is almost always the result of a softmax operator.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Label",
   "comment" : "(Tensor), the ground truth which is a 2-D tensor. When soft_label is set to false, Label is a Tensor<int64> with shape [N x 1]. When soft_label is set to true, Label is a Tensor<float/double> with shape [N x K].",
3609 3610 3611 3612 3613 3614
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
3615
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor with shape [N x 1]. The cross entropy loss.",
3616 3617 3618 3619 3620
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3621 3622 3623
   "name" : "soft_label",
   "type" : "bool",
   "comment" : "(bool, default false), a flag indicating whether to interpretate the given labels as soft labels.",
3624 3625 3626
   "generated" : 0
 } ] 
},{
3627 3628
 "type" : "matmul",
 "comment" : "\nMatMul Operator.\n\n\nThis operator is used to perform (batched) matrix multiplication\nover the last two dimensions of the input tensors `X` and `Y`.\n\nIf a transpose flag is specified, the last two dimensions of the\ntensor are transposed. If the tensor is rank-1 of shape [D], then\nfor `X` it is treated as [1, D] in nontransposed form and as [D, 1]\nin transposed form, whereas for `Y` it is the opposite: It is treated\nas [D, 1] in nontransposed form and as [1, D] in transposed form.\n\nExamples without transpose:\n- X: [K], Y: [K] => Out: [1]\n- X: [K], Y: [K, N] => Out: [N]\n- X: [B, M, K], Y: [K] => Out: [B, M]\n- X: [M, K], Y: [B, K, N] => Out: [B, M, N]\n- X: [B, M, K], Y: [B, K, N] => Out: [B, M, N]\n\nThe behavior is designed to be similar to the `numpy.matmul` function.\nThe differences are:\n- Currently only rank 1 to rank 3 input tensors are supported.\n- We add `transpose_X` and `transpose_Y` flags.\n\nBoth the input `X` and `Y` can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input `X`.\n\n",
3629 3630 3631
 "inputs" : [ 
 { 
   "name" : "X",
3632 3633 3634 3635 3636 3637
   "comment" : "The first input of MatMul op",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
   "comment" : "The second input of MatMul op",
3638 3639 3640 3641 3642 3643
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3644
   "comment" : "The output of MatMul op",
3645 3646 3647
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659
 "attrs" : [ 
 { 
   "name" : "transpose_X",
   "type" : "bool",
   "comment" : "If true, use the transpose of `X`.\n        ",
   "generated" : 0
 }, { 
   "name" : "transpose_Y",
   "type" : "bool",
   "comment" : "If true, use the transpose of `Y`.\n        ",
   "generated" : 0
 } ] 
3660
},{
3661 3662
 "type" : "dropout",
 "comment" : "\nDropout Operator.\n\nDropout refers to randomly dropping out units in a nerual network. It is a\nregularization technique for reducing overfitting by preventing neuron\nco-adaption during training. The dropout operator randomly set (according to\nthe given dropout probability) the outputs of some units to zero, while others\nare set equal to their corresponding inputs.\n\n",
3663 3664 3665
 "inputs" : [ 
 { 
   "name" : "X",
3666 3667
   "comment" : "The input of dropout op.",
   "duplicable" : 0,
3668 3669 3670 3671 3672
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3673
   "comment" : "The output of dropout op.",
3674 3675
   "duplicable" : 0,
   "intermediate" : 0
3676 3677 3678 3679 3680
 }, { 
   "name" : "Mask",
   "comment" : "The random sampled dropout mask.",
   "duplicable" : 0,
   "intermediate" : 1
3681 3682 3683
 } ], 
 "attrs" : [ 
 { 
3684 3685 3686
   "name" : "dropout_prob",
   "type" : "float",
   "comment" : "Probability of setting units to zero.",
3687 3688
   "generated" : 0
 }, { 
3689 3690 3691 3692 3693 3694
   "name" : "is_test",
   "type" : "bool",
   "comment" : "True if in test phase.",
   "generated" : 0
 }, { 
   "name" : "seed",
3695
   "type" : "int",
3696
   "comment" : "Dropout random seed.",
3697 3698 3699
   "generated" : 0
 } ] 
},{
3700 3701
 "type" : "fetch",
 "comment" : "\nFetch Operator.\n\nIt should not be configured by users directly.\n\n",
3702 3703 3704
 "inputs" : [ 
 { 
   "name" : "X",
3705
   "comment" : "The input of fetch op",
3706 3707 3708 3709 3710
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3711 3712
   "name" : "Out",
   "comment" : "The output of fetch op",
3713 3714 3715
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3716 3717 3718 3719 3720 3721 3722
 "attrs" : [ 
 { 
   "name" : "col",
   "type" : "int",
   "comment" : "(int) The column of fetch",
   "generated" : 0
 } ] 
3723
},{
3724 3725
 "type" : "squared_l2_distance",
 "comment" : "\nSquaredL2Distance operator\n\nThis operator will cacluate the squared L2 distance for the input and \nthe target. Number of distance value will be equal to the first dimension \nof input. First dimension of the target could be equal to the input or to 1. \nIf the first dimension of target is 1, the operator will broadcast target's \nfirst dimension to input's first dimension. During backward propagation, \nthe user can decide whether to calculate the gradient of the input or \nthe target or both.\n\nBoth the input X and Y can carry the LoD (Level of Details) information. \nHowever, the output only shares the LoD information with input X.\n    ",
3726 3727 3728
 "inputs" : [ 
 { 
   "name" : "X",
3729 3730 3731 3732 3733 3734
   "comment" : "(Tensor) Input of SquaredL2DistanceOp.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
   "comment" : "(Tensor) Target of SquaredL2DistanceOp.",
3735 3736 3737 3738 3739
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3740 3741 3742 3743 3744
   "name" : "sub_result",
   "comment" : "(Tensor) Buffering subtraction result which will be reused in backward.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
3745
   "name" : "Out",
3746
   "comment" : "(Tensor) Squared l2 distance between input and target.",
3747 3748 3749
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3750
 "attrs" : [  ] 
3751
},{
3752 3753
 "type" : "while",
 "comment" : "\n",
3754 3755 3756
 "inputs" : [ 
 { 
   "name" : "X",
3757 3758 3759 3760 3761 3762 3763
   "comment" : "A set of variables, which are required by operators inside the block of While Op.",
   "duplicable" : 1,
   "intermediate" : 0
 }, { 
   "name" : "Condition",
   "comment" : "(Bool) An scalar. When it's False, the While Op will be terminated.",
   "duplicable" : 1,
3764 3765 3766 3767
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3768 3769 3770 3771 3772 3773 3774
   "name" : "Out",
   "comment" : "A set of variables, which will be assigned with values generated by the operators inside the block of While Op.",
   "duplicable" : 1,
   "intermediate" : 0
 }, { 
   "name" : "StepScopes",
   "comment" : "(StepScopeVar) A vector of local scope, which size equals the step number of While Op. The i'th scope storages temporary variables generated in the i'th step.",
3775 3776 3777
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3778 3779 3780 3781 3782 3783 3784
 "attrs" : [ 
 { 
   "name" : "step_block",
   "type" : "block id",
   "comment" : "The step block inside WhileOp",
   "generated" : 0
 } ] 
3785
},{
3786 3787
 "type" : "relu",
 "comment" : "\nRelu Activation Operator.\n\n$y = \\max(x, 0)$\n\n",
3788 3789 3790
 "inputs" : [ 
 { 
   "name" : "X",
3791
   "comment" : "Input of Relu operator",
3792 3793 3794 3795 3796 3797
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
3798
   "comment" : "Output of Relu operator",
3799 3800 3801 3802 3803
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
3804 3805
 "type" : "decayed_adagrad",
 "comment" : "\nDecayed Adagrad Optimizer.\n\nThe update is done as follows:\n\n$$\nmoment\\_out = decay * moment + (1 - decay) * grad * grad \\\\\nparam\\_out = param - \\frac{learning\\_rate * grad}{\\sqrt{moment\\_out} + epsilon}\n$$\n\nThe original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)\ndoes not have an epsilon attribute. It is added here for numerical\nstability to avoid the division by zero error.\n\n",
3806 3807
 "inputs" : [ 
 { 
3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824
   "name" : "Param",
   "comment" : "(Tensor) Input parameter",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Grad",
   "comment" : "(Tensor) Input gradient",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Moment",
   "comment" : "(Tensor) Second moment",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor) Learning rate",
3825 3826 3827 3828 3829
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3830 3831 3832 3833 3834 3835 3836
   "name" : "ParamOut",
   "comment" : "(Tensor) Output parameter",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "MomentOut",
   "comment" : "(Tensor) Output second moment",
3837 3838 3839 3840 3841
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3842 3843 3844 3845 3846 3847 3848 3849
   "name" : "decay",
   "type" : "float",
   "comment" : "(float, default 0.95) Discounting factor for coming gradient",
   "generated" : 0
 }, { 
   "name" : "epsilon",
   "type" : "float",
   "comment" : "(float, default 1.0e-6) Constant for numerical stability",
3850 3851
   "generated" : 0
 } ] 
3852
},{
3853 3854
 "type" : "beam_search",
 "comment" : "This is a beam search operator that help to generate sequences.",
3855 3856
 "inputs" : [ 
 { 
3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868
   "name" : "pre_ids",
   "comment" : "ids in previous step",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "ids",
   "comment" : "a LoDTensor of shape of [None,k]",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "scores",
   "comment" : "a LoDTensor that has the same shape and LoD with `ids`",
3869 3870 3871 3872 3873
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3874 3875 3876 3877 3878 3879 3880
   "name" : "selected_ids",
   "comment" : "a LoDTensor that stores the IDs selected by beam search",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "selected_scores",
   "comment" : "a LoDTensor that has the same shape and LoD with `selected_ids`",
3881 3882 3883 3884 3885
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3886
   "name" : "level",
3887
   "type" : "int",
3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898
   "comment" : "the level of LoDTensor",
   "generated" : 0
 }, { 
   "name" : "beam_size",
   "type" : "int",
   "comment" : "beam size for beam search",
   "generated" : 0
 }, { 
   "name" : "end_id",
   "type" : "int",
   "comment" : "the token id which indicates the end of a sequence",
3899 3900 3901
   "generated" : 0
 } ] 
},{
3902 3903
 "type" : "split_lod_tensor",
 "comment" : "\n        Split a LoDTensor with a Mask at certain level. The input LoDTensor\n        has 3 sequence at certain lod level. The Mask is a bool column vector,\n        such as [0, 1, 0] at the same level. The first and third sequence will\n        be send to False Output LoDTensor; whereas the second sequence will\n        be send to True Output LoDTensor. Please refer to MergeLoDTensorOp.",
3904 3905 3906
 "inputs" : [ 
 { 
   "name" : "X",
3907
   "comment" : "The input LoDTensor",
3908 3909 3910
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3911 3912
   "name" : "Mask",
   "comment" : "A bool column vector which mask the input",
3913 3914 3915 3916 3917
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3918 3919 3920 3921 3922 3923 3924
   "name" : "OutTrue",
   "comment" : "True branch of input LoDTensor",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "OutFalse",
   "comment" : "False branch of input LoDTensor",
3925 3926 3927 3928 3929
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3930 3931 3932
   "name" : "level",
   "type" : "int",
   "comment" : "(int) the specific lod level to split.",
3933 3934 3935
   "generated" : 0
 } ] 
},{
3936 3937
 "type" : "greater_equal",
 "comment" : "greater_equal Operator\n\nIt operates element-wise on X and Y, and returns the Out. Each of them is a\nN-dim tensor. X and Y could be any type.  The each element of the Out tensor is\ncalculated by Out = X >= Y\n",
3938 3939 3940
 "inputs" : [ 
 { 
   "name" : "X",
3941 3942 3943 3944 3945 3946
   "comment" : "(LoDTensor) the left hand operand of greater_equal operator",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
   "comment" : "(LoDTensor) the right hand operand of greater_equal operator",
3947 3948 3949 3950 3951 3952
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
3953
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is Out = X >= Y",
3954 3955 3956
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
3957
 "attrs" : [  ] 
3958
},{
3959 3960
 "type" : "crop",
 "comment" : "\nCrop Operator.\n\nCrop input into output, as specified by offsets and shape.\n\nThere are two ways to set shape:\n1. reference input: crop input X into the same shape as reference input.\n                    The dimension of reference input should\n                    be the same as the dimension of input X.\n2. shape list: crop input X into the shape described by a list<int>.\n               The size of shape list should be the same as\n               the dimension size of input X.\n\nThe input should be a k-D tensor(k > 0 and k < 7). As an example:\n\nGiven:\n\n    X = [[0, 1, 2, 0, 0]\n         [0, 3, 4, 0, 0]\n         [0, 0, 0, 0, 0]],\n\nand\n\n    offsets = [0, 1],\n\nand\n\n    shape = [2, 2],\n\nwe get:\n\n    Out = [[1, 2],\n           [3, 4]].\n\n",
3961 3962 3963
 "inputs" : [ 
 { 
   "name" : "X",
3964
   "comment" : "The input of pad op. The input should be a k-D tensor(k > 0 and k < 7).",
3965 3966 3967
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
3968 3969
   "name" : "Y",
   "comment" : "The input used as reference for cropping, which is of the same dimensions as X.",
3970 3971 3972 3973 3974
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
3975 3976
   "name" : "Out",
   "comment" : "The output of crop op, which is of the same dimensions as X.",
3977 3978 3979 3980 3981
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
3982 3983 3984 3985 3986 3987 3988 3989
   "name" : "offsets",
   "type" : "int array",
   "comment" : "A list<int> describing offsets to be cropped. The size of offsets list should be the same as the dimension size of input X.",
   "generated" : 0
 }, { 
   "name" : "shape",
   "type" : "int array",
   "comment" : "A list<int> describing the shape of output. The size of shape list should be the same as the dimension size of input X.",
3990 3991 3992
   "generated" : 0
 } ] 
},{
3993 3994
 "type" : "brelu",
 "comment" : "\nBRelu Activation Operator.\n\n$y = \\max(\\min(x, t_{min}), t_{max})$\n\n",
3995 3996
 "inputs" : [ 
 { 
3997 3998
   "name" : "X",
   "comment" : "Input of BRelu operator",
3999 4000 4001 4002 4003
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4004 4005
   "name" : "Y",
   "comment" : "Output of BRelu operator",
4006 4007 4008 4009 4010
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
4011
   "name" : "t_min",
4012
   "type" : "float",
4013
   "comment" : "The min marginal value of BRelu",
4014 4015
   "generated" : 0
 }, { 
4016 4017 4018
   "name" : "t_max",
   "type" : "float",
   "comment" : "The max marginal value of BRelu",
4019 4020 4021
   "generated" : 0
 } ] 
},{
4022 4023
 "type" : "crf_decoding",
 "comment" : "\nThe crf_decoding operator reads the emission feature weights and the transition\nfeature weights learned by the linear_chain_crf operator. It implements the\nViterbi algorithm which is a dynamic programming algorithm for finding the most\nlikely sequence of hidden states, called the Viterbi path, that results in a\nsequence of observed tags.\n\nThe output of this operator changes according to whether Input(Label) is given:\n\n1. Input(Label) is given:\n\nThis happens in training. This operator is used to co-work with the chunk_eval\noperator.\n\nWhen Input(Label) is given, the crf_decoding operator returns a row vector\nwith shape [N x 1] whose values are fixed to be 0, indicating an incorrect\nprediction, or 1 indicating a tag is correctly predicted. Such an output is the\ninput to chunk_eval operator.\n\n2. Input(Label) is not given:\n\nThis is the standard decoding process.\n\nThe crf_decoding operator returns a row vector with shape [N x 1] whose values\nrange from 0 to maximum tag number - 1. Each element indicates an index of a\npredicted tag.\n",
4024 4025
 "inputs" : [ 
 { 
4026 4027
   "name" : "Emission",
   "comment" : "(LoDTensor, default: LoDTensor<float>). A LoDTensor with shape [N x D] where N is the size of the mini-batch and D is the total tag number. This input is the unscaled emission weight matrix of the linear_chain_crf operator.",
4028 4029 4030
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4031 4032
   "name" : "Transition",
   "comment" : "(Tensor, default: Tensor<float>). A Tensor with shape [(D + 2) x D]. This input is the transition weights learned by the linear_chain_crf operator, denoted as w. The 1st row of w are transition weights for the start mask. The 2nd row of w are transition weights for the end mask. Transition weights between other tags begin from the 3rd row of w. See more details in comments of the linear_chain_crf operator.",
4033 4034 4035
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4036 4037
   "name" : "Label",
   "comment" : "(LoDTensor,  LoDTensor<int64_t>). The ground truth with shape [N x 1]. This input is optional. See more details in the operator's comments.",
4038 4039 4040 4041 4042
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4043 4044
   "name" : "ViterbiPath",
   "comment" : "(LoDTensor, LoDTensor<int64_t>). The decoding results. What to return changes depending on whether the Input(Label) (the ground truth) is given. See more details in the operator's comment.",
4045 4046 4047 4048 4049
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4050 4051
 "type" : "cos_sim",
 "comment" : "\nCosine Similarity Operator.\n\n$Out = X^T * Y / (\\sqrt{X^T * X} * \\sqrt{Y^T * Y})$\n\nThe input X and Y must have the same shape, except that the 1st dimension\nof input Y could be just 1 (different from input X), which will be\nbroadcasted to match the shape of input X before computing their cosine\nsimilarity.\n\nBoth the input X and Y can carry the LoD (Level of Details) information,\nor not. But the output only shares the LoD information with input X.\n\n",
4052 4053 4054
 "inputs" : [ 
 { 
   "name" : "X",
4055
   "comment" : "The 1st input of cos_sim op.",
4056 4057 4058 4059
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
4060
   "comment" : "The 2nd input of cos_sim op.",
4061 4062 4063 4064 4065 4066
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4067
   "comment" : "The output of cos_sim op.",
4068 4069
   "duplicable" : 0,
   "intermediate" : 0
4070 4071 4072 4073 4074 4075 4076 4077 4078 4079
 }, { 
   "name" : "XNorm",
   "comment" : "Norm of the first input, reduced along the 1st dimension.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "YNorm",
   "comment" : "Norm of the second input, reduced along the 1st dimension.",
   "duplicable" : 0,
   "intermediate" : 1
4080 4081 4082
 } ], 
 "attrs" : [  ] 
},{
4083 4084
 "type" : "conv3d_transpose_cudnn",
 "comment" : "\nConvolution3D Transpose Operator.\n\nThe convolution transpose operation calculates the output based on the input, filter\nand strides, paddings, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and output(Output) are in NCDHW format. Where N is batch size, C is the\nnumber of channels, D is the depth of the feature, H is the height of the feature,\nand W is the width of the feature.\nFilter(Input) is in MCDHW format. Where M is the number of input feature channels,\nC is the number of output feature channels, D is the depth of the filter,H is the\nheight of the filter, and W is the width of the filter.\nParameters(strides, paddings) are three elements. These three elements represent\ndepth, height and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:   \n  Input:\n       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{in}, C_{out}, D_f, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$\n  Where\n  $$\n       D_{out} = (D_{in} - 1) * strides[0] - 2 * paddings[0] + D_f \\\\\n       H_{out} = (H_{in} - 1) * strides[1] - 2 * paddings[1] + H_f \\\\\n       W_{out} = (W_{in} - 1) * strides[2] - 2 * paddings[2] + W_f\n  $$\n",
4085 4086
 "inputs" : [ 
 { 
4087 4088
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution transpose operator.The format of input tensor is NCDHW. Where N is batch size, C is the number of channels, D is the depth of the feature, H is the height of the feature, and W is the width of the feature.",
4089 4090 4091
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4092 4093
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution transpose operator.The format of the filter tensor is MCDHW, where M is the number of input feature channels, C is the number of output feature channels, D is the depth of the filter, H is the height of the filter, and W is the width of the filter.We enforce groups number == 1 and padding == 0 in the convolution3d transpose scenario.",
4094 4095 4096 4097 4098
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4099 4100
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution transpose operator.The format of output tensor is also NCDHW.Where N is batch size, C is the number of channels, D is the depth of the feature, H is the height of the feature, and W is the width of the feature.",
4101 4102 4103
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125
 "attrs" : [ 
 { 
   "name" : "strides",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1, 1}), the strides{d_stride, h_stride, w_stride} of convolution transpose operator.",
   "generated" : 0
 }, { 
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int> default:{0, 0, 0}), paddings(d_pad, h_pad, w_pad) of convolution transpose operator.",
   "generated" : 0
 }, { 
   "name" : "dilations",
   "type" : "int array",
   "comment" : "dilations of convolution operator.",
   "generated" : 0
 }, { 
   "name" : "workspace_size_MB",
   "type" : "int",
   "comment" : "workspace size for cudnn, in MB, workspace is a section of GPU memory which will be allocated/freed each time the operator runs, larger workspace size can increase performance but also requires better hardward. This size should be carefully setted.",
   "generated" : 0
 } ] 
4126
},{
4127 4128
 "type" : "conv_shift",
 "comment" : "\nConvShift Operator.\n\nA layer for circular convolution of two vectors,\nas used in the Neural Turing Machine: https://arxiv.org/abs/1410.5401\n\nThe equation is:\n\n$$Out[i] = \\sum_{j=-(N-1)/2}^{(N-1)/2} X_{i+j} * Y_{j}$$\n\nwhere X's index is computed modulo M, and Y's index is computed modulo N.\n\nBoth inputs X and Y can carry LoD (Level of Details) information.\nHowever, the output only shares the LoD information with input X.\n\n",
4129 4130 4131
 "inputs" : [ 
 { 
   "name" : "X",
4132
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor with shape B x M, where B is the batch size and M is the data dimension.",
4133 4134 4135 4136
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
4137
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor with shape B x N, where B is the batch size and N is the data dimension. N must be odd.",
4138 4139 4140 4141 4142 4143
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4144
   "comment" : "(Tensor, default Tensor<float>), a 2-D tensor with shape B x M, i.e., the same shape as X.",
4145 4146 4147 4148 4149
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4150 4151
 "type" : "conv2d_cudnn",
 "comment" : "\nConvolution Operator.\n\nThe convolution operation calculates the output based on the input, filter\nand strides, paddings, dilations, groups parameters. The size of each dimension of the\nparameters is checked in the infer-shape.\nInput(Input) and Output(Output) are in NCHW format. Where N is batch\nsize, C is the number of channels, H is the height of the feature, and W is\nthe width of the feature.\nFilters(Input) is MCHW format. Where M is the number of output image channels, C is\nthe number of input image channels, H is the height of the filter, and W\nis the width of the filter.\nParameters(strides, paddings, dilations) are two elements. These two elements represent\nheight and width, respectively.\nThe input(X) size and output(Out) size may be different.\n\nExample:\n  Input:\n       Input shape: $(N, C_{in}, H_{in}, W_{in})$\n       Filter shape: $(C_{out}, C_{in}, H_f, W_f)$\n  Output:\n       Output shape: $(N, C_{out}, H_{out}, W_{out})$\n  Where\n$$\n       H_{out}= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]}+ 1 \\\\\n       W_{out}= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]}+ 1\n$$\n",
4152 4153
 "inputs" : [ 
 { 
4154 4155
   "name" : "Input",
   "comment" : "(Tensor) The input tensor of convolution operator. The format of input tensor is NCHW, where N is batch size, C is the number of channels, H is the height of the feature, and W is the width of the feature.",
4156 4157 4158
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4159 4160
   "name" : "Filter",
   "comment" : "(Tensor) The filter tensor of convolution operator. The format of the filter tensor is MCHW, where M is the number of output image channels, C is the number of input image channels, H is the height of the filter, and W is the width of the filter. If the groups attribute is greater than 1, C equals the number of input image channels divided by the groups.",
4161 4162 4163 4164 4165
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4166 4167
   "name" : "Output",
   "comment" : "(Tensor) The output tensor of convolution operator. The format of output tensor is also NCHW.",
4168 4169 4170 4171 4172
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
4173
   "name" : "strides",
4174
   "type" : "int array",
4175
   "comment" : "(vector<int> default:{1, 1}), the strides(h_stride, w_stride) of convolution operator.",
4176 4177
   "generated" : 0
 }, { 
4178 4179 4180
   "name" : "paddings",
   "type" : "int array",
   "comment" : "(vector<int> default:{0, 0}), the paddings(h_pad, w_pad) of convolution operator.",
4181 4182
   "generated" : 0
 }, { 
4183 4184 4185
   "name" : "groups",
   "type" : "int",
   "comment" : "(int default:1), the groups number of the convolution operator. According to grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels.",
4186 4187
   "generated" : 0
 }, { 
4188 4189 4190
   "name" : "dilations",
   "type" : "int array",
   "comment" : "(vector<int> default:{1, 1}), the dilations(h_dilation, w_dilation) of convolution operator.",
4191 4192
   "generated" : 0
 }, { 
4193
   "name" : "workspace_size_MB",
4194
   "type" : "int",
4195
   "comment" : "workspace size for cudnn, in MB, workspace is a section of GPU memory which will be allocated/freed each time the operator runs, larger workspace size can increase performance but also requires better hardware. This size should be chosen carefully.",
4196 4197 4198
   "generated" : 0
 } ] 
},{
4199 4200
 "type" : "conditional_block",
 "comment" : "Conditional block operator\n\nRun the sub-block if X is not empty. Params is the other inputs and Out is the\noutputs of the sub-block.\n",
4201 4202 4203
 "inputs" : [ 
 { 
   "name" : "X",
4204 4205
   "comment" : "The conditional variable of this operator. If X is empty, the whole sub-block will not be executed.",
   "duplicable" : 1,
4206 4207
   "intermediate" : 0
 }, { 
4208 4209 4210
   "name" : "Params",
   "comment" : "The input variables of the sub-block.",
   "duplicable" : 1,
4211 4212 4213 4214 4215
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4216 4217
   "comment" : "The output variables of the sub-block.",
   "duplicable" : 1,
4218 4219
   "intermediate" : 0
 }, { 
4220 4221
   "name" : "Scope",
   "comment" : "(std::vector<Scope*>) The step scope of conditional block. To unify the conditional block, rnn and while op, the type of scope is std::vector<Scope*>",
4222
   "duplicable" : 0,
4223
   "intermediate" : 0
4224 4225 4226
 } ], 
 "attrs" : [ 
 { 
4227 4228 4229
   "name" : "block",
   "type" : "block id",
   "comment" : "The step block of conditional block operator",
4230 4231 4232
   "generated" : 0
 } ] 
},{
4233 4234
 "type" : "sum",
 "comment" : "\nSum operator.\n\nThis operators sums the input tensors. All the inputs can carry the \nLoD (Level of Details) information. However, the output only shares \nthe LoD information with the first input.\n",
4235 4236 4237
 "inputs" : [ 
 { 
   "name" : "X",
4238 4239
   "comment" : "(vector<Tensor>) The input tensors of sum operator.",
   "duplicable" : 1,
4240 4241 4242 4243
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4244 4245
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of sum operator.",
4246 4247 4248 4249 4250
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4251 4252
 "type" : "concat",
 "comment" : "\nConcat Operator.\n\nConcatenate the input tensors along dimension axis.\nExamples:\n  Input[0] = [[1,2],[3,4]]\n  Input[1] = [[5,6]]\n  axis = 0\n  Output = [[1,2],\n            [3,4],\n            [5,6]]\n\n",
4253 4254 4255
 "inputs" : [ 
 { 
   "name" : "X",
4256 4257
   "comment" : "Input tensors of concat operator.",
   "duplicable" : 1,
4258 4259 4260 4261 4262
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4263
   "comment" : "Output tensor of concat operator.",
4264 4265 4266
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
4267 4268 4269 4270 4271 4272 4273
 "attrs" : [ 
 { 
   "name" : "axis",
   "type" : "int",
   "comment" : "The axis along which the input tensors will be concatenated.",
   "generated" : 0
 } ] 
4274
},{
4275 4276
 "type" : "less_equal",
 "comment" : "less_equal Operator\n\nIt operates element-wise on X and Y, and returns the Out. Each of them is a\nN-dim tensor. X and Y could be any type.  The each element of the Out tensor is\ncalculated by Out = X <= Y\n",
4277 4278 4279
 "inputs" : [ 
 { 
   "name" : "X",
4280 4281 4282 4283 4284 4285
   "comment" : "(LoDTensor) the left hand operand of less_equal operator",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
   "comment" : "(LoDTensor) the right hand operand of less_equal operator",
4286 4287 4288 4289 4290
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4291 4292
   "name" : "Out",
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is Out = X <= Y",
4293 4294 4295 4296 4297
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4298 4299
 "type" : "equal",
 "comment" : "equal Operator\n\nIt operates element-wise on X and Y, and returns the Out. Each of them is a\nN-dim tensor. X and Y could be any type.  The each element of the Out tensor is\ncalculated by Out = X == Y\n",
4300 4301 4302
 "inputs" : [ 
 { 
   "name" : "X",
4303
   "comment" : "(LoDTensor) the left hand operand of equal operator",
4304 4305 4306 4307
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Y",
4308
   "comment" : "(LoDTensor) the right hand operand of equal operator",
4309 4310 4311 4312 4313 4314
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4315
   "comment" : "(LoDTensor) n-dim bool tensor. Each element is Out = X == Y",
4316 4317 4318 4319 4320
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4321 4322
 "type" : "gather",
 "comment" : "\nGather Operator.\n\n$Out = X[Index]$\n\nOut is obtained by gathering entries of the outer-most dimension \nof X indexed by Index and concatenate them together.\n\nExample:\n\nX = [[1, 2],\n     [3, 4],\n     [5, 6]]\n\nIndex = [[1, 2]]\n\nThen:\n\nOut = [[3, 4],\n       [5, 6]]\n\n",
4323 4324 4325
 "inputs" : [ 
 { 
   "name" : "X",
4326 4327 4328 4329 4330 4331
   "comment" : "The source input of gather op",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Index",
   "comment" : "The index input of gather op",
4332 4333 4334 4335 4336 4337
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4338
   "comment" : "The output of gather op",
4339 4340 4341 4342 4343
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4344 4345
 "type" : "clip_by_norm",
 "comment" : "\nClipByNorm Operator.\n\nThis operator limits the L2 norm of the input $X$ within $max\\_norm$.\nIf the L2 norm of $X$ is less than or equal to $max\\_norm$, $Out$ will be\nthe same as $X$. If the L2 norm of $X$ is greater than $max\\_norm$, $X$ will\nbe linearly scaled to make the L2 norm of $Out$ equal to $max\\_norm$, as\nshown in the following formula:\n\n$$\nOut = \\frac{max\\_norm * X}{norm(X)},\n$$\n\nwhere $norm(X)$ represents the L2 norm of $X$.\n",
4346 4347 4348
 "inputs" : [ 
 { 
   "name" : "X",
4349
   "comment" : "(Tensor) The input of clip_by_norm op.The number of dimensions must be between [1, 9].",
4350 4351 4352 4353 4354
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4355 4356
   "name" : "Out",
   "comment" : "(Tensor) The output of clip_by_norm op with shape as input(X)",
4357 4358 4359
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
4360 4361 4362 4363 4364 4365 4366
 "attrs" : [ 
 { 
   "name" : "max_norm",
   "type" : "float",
   "comment" : "(float) The maximum norm value.",
   "generated" : 0
 } ] 
4367
},{
4368 4369
 "type" : "chunk_eval",
 "comment" : "\nFor some basics of chunking, please refer to\n‘Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>’.\n\n\nCheckEvalOp computes the precision, recall, and F1-score of chunk detection,\nand supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes.\nHere is a NER example of labeling for these tagging schemes:\n\n \t     Li     Ming    works  at  Agricultural   Bank   of    China  in  Beijing.\n  IO:    I-PER  I-PER   O      O   I-ORG          I-ORG  I-ORG I-ORG  O   I-LOC\n  IOB:   B-PER  I-PER   O      O   B-ORG          I-ORG  I-ORG I-ORG  O   B-LOC\n  IOE:   I-PER  E-PER   O      O   I-ORG          I-ORG  I-ORG E-ORG  O   E-LOC\n  IOBES: B-PER  E-PER   O      O   I-ORG          I-ORG  I-ORG E-ORG  O   S-LOC\n\nThere are three chunk types(named entity types) including PER(person), ORG(organization)\nand LOC(LOCATION), and we can see that the labels have the form <tag type>-<chunk type>.\n\nSince the calculations actually use label ids rather than labels, extra attention\nshould be paid when mapping labels to ids to make CheckEvalOp work. The key point\nis that the listed equations are satisfied by ids.\n\n    tag_type = label % num_tag_type\n    chunk_type = label / num_tag_type\n\nwhere `num_tag_type` is the num of tag types in the tagging scheme, `num_chunk_type`\nis the num of chunk types, and `tag_type` get its value from the following table.\n\n    Scheme Begin Inside End   Single\n     plain   0     -      -     -\n     IOB     0     1      -     -\n     IOE     -     0      1     -\n     IOBES   0     1      2     3\n\nStill use NER as example, assuming the tagging scheme is IOB while chunk types are ORG,\nPER and LOC. To satisfy the above equations, the label map can be like this:\n\n    B-ORG  0\n    I-ORG  1\n    B-PER  2\n    I-PER  3\n    B-LOC  4\n    I-LOC  5\n    O      6\n\nIt’s not hard to verify the equations noting that the num of chunk types\nis 3 and the num of tag types in IOB scheme is 2. For example, the label\nid of I-LOC is 5, the tag type id of I-LOC is 1, and the chunk type id of\nI-LOC is 2, which consistent with the results from the equations.\n",
4370 4371
 "inputs" : [ 
 { 
4372 4373
   "name" : "Inference",
   "comment" : "(Tensor, default: Tensor<int64_t>). Predictions from the network.",
4374 4375 4376
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4377 4378
   "name" : "Label",
   "comment" : "(Tensor, default: Tensor<int64_t>). The true tag sequences.",
4379 4380 4381 4382 4383
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4384 4385
   "name" : "Precision",
   "comment" : "(float). The evaluated precision (called positive predictive value) of chunks on the given mini-batch.",
4386 4387
   "duplicable" : 0,
   "intermediate" : 0
4388 4389 4390
 }, { 
   "name" : "Recall",
   "comment" : "(float). The evaluated recall (true positive rate or sensitivity) of chunks on the given mini-batch.",
4391 4392 4393
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4394 4395
   "name" : "F1-Score",
   "comment" : "(float). The evaluated F1-Score on the given mini-batch.",
4396 4397
   "duplicable" : 0,
   "intermediate" : 0
4398 4399 4400 4401 4402 4403 4404
 } ], 
 "attrs" : [ 
 { 
   "name" : "num_chunk_types",
   "type" : "int",
   "comment" : "(int). The number of chunk type. See below for details.",
   "generated" : 0
4405
 }, { 
4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422
   "name" : "chunk_scheme",
   "type" : "string",
   "comment" : "(string, default IOB). The labeling scheme indicating how to encode the chunks. Must be IOB, IOE, IOBES or plain. See below for details.",
   "generated" : 0
 }, { 
   "name" : "excluded_chunk_types",
   "type" : "int array",
   "comment" : "(list<int>) A list including chunk type ids indicating chunk types that are not counted. See below for details.",
   "generated" : 0
 } ] 
},{
 "type" : "sigmoid",
 "comment" : "\nSigmoid Activation Operator\n\n$$y = \\frac{1}{1 + e^{-x}}$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Sigmoid operator",
4423 4424 4425 4426 4427
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4428 4429
   "name" : "Y",
   "comment" : "Output of Sigmoid operator",
4430 4431 4432 4433 4434
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4435 4436
 "type" : "sequence_concat",
 "comment" : "\nThe sequence_concat operator concatenates multiple LoDTensors. \nIt only supports sequence (LoD Tensor with level number is 1) \nor a nested sequence (LoD tensor with level number is 2) as its input.\n- Case1:\n  If the axis is other than 0(here, axis is 1 and level is 1),\n  each input should have the same LoD information and the LoD \n  information of the output keeps the same as the input.\n\n  LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)\n  LoD(x1) = {{0,2,4}, {0,1,2,3,4}}; Dims(x1) = (4,4,4)\n  LoD(Out) = {{0,2,4}, {0,1,2,3,4}}; Dims(Out) = (4,7,4)\n\n- Case2:\n  If the axis is 0(here, leve is 0), the inputs are concatenated along \n  time steps, the LoD information of the output need to re-compute.\n  The LoD information of level-1 should be same.\n\n  LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)\n  LoD(x1) = {{0,2,4}, {0,1,3,5,7}}; Dims(x1) = (7,3,4)\n  LoD(Out) = {{0,2,4}, {0,2,5,8,11}}; Dims(Out) = (11,3,4)\n\n- Case3:\n  If the axis is 0(here, level is 1).\n\n  LoD(x0) = {{0,2,4}, {0,1,2,3,4}}; Dims(x0) = (4,3,4)\n  LoD(x1) = {{0,3,4}, {0,1,3,5,7}}; Dims(x1) = (7,3,4)\n  LoD(Out) = {{0,5,8}, {0,1,2,3,5,7,8,9,11}}; Dims(Out) = (11,3,4)\n\n- Case4:\n  If the LoD number is 1, axis is 0, level is 0\n\n  LoD(x0) = {{0,1,2,3,4}}; Dims(x0) = (4,3,4)\n  LoD(x1) = {{0,1,3,5,7}}; Dims(x1) = (7,3,4)\n  LoD(Out) = {{0,2,5,8,11}}; Dims(Out) = (11,3,4)\n\nNOTE: The levels of all the inputs should be the same.\n    ",
4437 4438
 "inputs" : [ 
 { 
4439 4440 4441 4442 4443 4444 4445 4446 4447
   "name" : "X",
   "comment" : "(LodTensorArray) Input is a vector of LoDTensor, each of which is a variable-length sequence or nested sequence.",
   "duplicable" : 1,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDTensor), Variable-length output of sequence_concat Op.",
4448 4449
   "duplicable" : 0,
   "intermediate" : 0
4450 4451 4452 4453 4454 4455 4456
 } ], 
 "attrs" : [ 
 { 
   "name" : "axis",
   "type" : "int",
   "comment" : "(int, default 0) The axis along which the inputs will be joined. If axis is 0, the inputs will be joined with LoD index.",
   "generated" : 0
4457
 }, { 
4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469
   "name" : "level",
   "type" : "int",
   "comment" : "(int, default 0) The level at which the inputs will be joined. If the level is 0, the inputs will be joined at the nested sequence level. If the level is 1, the inputs will be joined at the sequence level. The level should be less than the level number of inputs.",
   "generated" : 0
 } ] 
},{
 "type" : "floor",
 "comment" : "\nFloor Activation Operator.\n\n$y = floor(x)$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Floor operator",
4470 4471 4472 4473 4474
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4475 4476
   "name" : "Y",
   "comment" : "Output of Floor operator",
4477 4478 4479 4480 4481
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4482 4483
 "type" : "cast",
 "comment" : "\nCast Operator.\n\nThis Operator casts the input tensor to another data type and\nreturns tha Output Tensor.\n\n",
4484 4485 4486
 "inputs" : [ 
 { 
   "name" : "X",
4487
   "comment" : "The input tensor of cast op",
4488 4489
   "duplicable" : 0,
   "intermediate" : 0
4490 4491 4492 4493 4494
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "The output tensor of cast op",
4495 4496
   "duplicable" : 0,
   "intermediate" : 0
4497 4498 4499 4500 4501 4502 4503
 } ], 
 "attrs" : [ 
 { 
   "name" : "out_dtype",
   "type" : "int",
   "comment" : "output data type",
   "generated" : 0
4504
 }, { 
4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516
   "name" : "in_dtype",
   "type" : "int",
   "comment" : "input data type",
   "generated" : 0
 } ] 
},{
 "type" : "ceil",
 "comment" : "\nCeil Activation Operator.\n\n$y = ceil(x)$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Ceil operator",
4517 4518 4519 4520 4521
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4522 4523
   "name" : "Y",
   "comment" : "Output of Ceil operator",
4524 4525 4526 4527 4528 4529
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "lrn",
4530
 "comment" : "\nLocal Response Normalization Operator.\n\nThis operator comes from the paper:\n<<ImageNet Classification with Deep Convolutional Neural Networks>>.\n\nThe original formula is:\n\n$$\nOutput(i, x, y) = Input(i, x, y) / \\left(\nk + \\alpha \\sum\\limits^{\\min(C, c + n/2)}_{j = \\max(0, c - n/2)}\n(Input(j, x, y))^2\n\\right)^{\\beta}\n$$\n\nFunction implementation:\n\nInputs and outpus are in NCHW format, while input.shape.ndims() equals 4.\nAnd dimensions 0 ~ 3 represent batch size, feature maps, rows,\nand columns, respectively.\n\nInput and Output in the formula above is for each map(i) of one image, and\nInput(i, x, y), Output(i, x, y) represents an element in an image.\n\nC is the number of feature maps of one image. n is a hyper-parameter\nconfigured when operator is initialized. The sum in the denominator\nis the sum of the same positions in the neighboring maps.\n\n",
4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input of LRN operator. It must be a 4D tenor with NCHW format.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output of LRN operator, which is also the 4D tensor with NCHW format.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "MidOut",
   "comment" : "(Tensor) Middle result of LRN operator. It's computed in forward process and also used in backward process.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "n",
   "type" : "int",
   "comment" : "(int default 5) n is the \"adjacent\" kernel that maps at the same spatial position.",
   "generated" : 0
 }, { 
   "name" : "k",
   "type" : "float",
   "comment" : "(float, default 2.0) k is the bias.",
   "generated" : 0
 }, { 
   "name" : "alpha",
   "type" : "float",
   "comment" : "(float, default 0.0001) alpha is the scale number.",
   "generated" : 0
 }, { 
   "name" : "beta",
   "type" : "float",
   "comment" : "(float, default 0.75) beta is the power number.",
   "generated" : 0
 } ] 
},{
4573 4574
 "type" : "bilinear_tensor_product",
 "comment" : "\nBilinear Tensor Product operator.\nGiven input X and Y, a 3D tensor Weight and a Bias. Each column of the\nOutput is computed by one slice $i = 1, . . . , k$ of the tensor:\n\n$$\nM =  (X W_i) * Y \\\\\nOut_i = \\sum_j {M_j} + Bias_i\n$$\n\nWhere $W_i$ is the $i$-th slice of Input(Weight);\n      $M_j$ is the $j$-th column of $M$;\n      $Out_i$ is the $i$-th column of Output(Out);\n      $Bias_i$ is a column vector, each element of it is equal to\n        the $i$-th element of $Bias$;\n\n",
4575 4576
 "inputs" : [ 
 { 
4577 4578
   "name" : "X",
   "comment" : "The first input of bilinear_tensor_product operator.",
4579 4580 4581
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4582 4583
   "name" : "Y",
   "comment" : "The second input of bilinear_tensor_product operator.",
4584 4585 4586
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4587 4588
   "name" : "Weight",
   "comment" : "The learnable parameters of bilinear_tensor_product operator.",
4589 4590
   "duplicable" : 0,
   "intermediate" : 0
4591 4592 4593
 }, { 
   "name" : "Bias",
   "comment" : "The learnable bias of bilinear_tensor_product operator.",
4594 4595 4596 4597 4598 4599
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4600
   "comment" : "The output of bilinear_tensor_product operator.",
4601 4602 4603 4604 4605
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4606 4607
 "type" : "batch_norm",
 "comment" : "\nBatch Normalization.\n\nBatch Norm has been implemented as discussed in the paper:\nhttps://arxiv.org/pdf/1502.03167.pdf\nCan be used as a normalizer function for conv2d and fully_connected operations.\nThe required data format for this layer is one of the following:\n1. NHWC `[batch, in_height, in_width, in_channels]`\n2. NCHW `[batch, in_channels, in_height, in_width]`\n\n",
4608 4609 4610
 "inputs" : [ 
 { 
   "name" : "X",
4611
   "comment" : "The input tensor",
4612 4613
   "duplicable" : 0,
   "intermediate" : 0
4614 4615 4616 4617
 }, { 
   "name" : "Scale",
   "comment" : "Scale is a 1-dimensional tensor of size C that is applied to the output",
   "duplicable" : 0,
4618 4619
   "intermediate" : 0
 }, { 
4620 4621 4622 4623
   "name" : "Bias",
   "comment" : "Bias is a 1-dimensional tensor of size C that is applied to the output",
   "duplicable" : 0,
   "intermediate" : 0
4624
 }, { 
4625 4626
   "name" : "Mean",
   "comment" : "The global mean (for training) or estimated mean (for testing)",
4627 4628 4629
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4630 4631
   "name" : "Variance",
   "comment" : "The global variance (for training) or estimated Variance (for testing)",
4632 4633 4634 4635 4636
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4637 4638
   "name" : "Y",
   "comment" : "result after normalization",
4639 4640 4641
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4642 4643
   "name" : "MeanOut",
   "comment" : "Share memory with Mean. Store the global mean when training",
4644 4645 4646
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4647 4648
   "name" : "VarianceOut",
   "comment" : "Share memory with Variance. Store the global Variance when training",
4649 4650
   "duplicable" : 0,
   "intermediate" : 0
4651 4652 4653 4654 4655 4656 4657 4658 4659 4660
 }, { 
   "name" : "SavedMean",
   "comment" : "Mean of the current mini batch, will apply to output when training",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "SavedVariance",
   "comment" : "Variance of the current mini batch, will apply to output when training",
   "duplicable" : 0,
   "intermediate" : 1
4661 4662 4663
 } ], 
 "attrs" : [ 
 { 
4664 4665 4666
   "name" : "is_test",
   "type" : "bool",
   "comment" : "",
4667 4668
   "generated" : 0
 }, { 
4669 4670 4671
   "name" : "momentum",
   "type" : "float",
   "comment" : "",
4672 4673
   "generated" : 0
 }, { 
4674 4675 4676 4677 4678 4679 4680 4681
   "name" : "epsilon",
   "type" : "float",
   "comment" : "",
   "generated" : 0
 }, { 
   "name" : "tensor_format",
   "type" : "string",
   "comment" : "",
4682 4683 4684
   "generated" : 0
 } ] 
},{
4685 4686
 "type" : "auc",
 "comment" : "\nArea Under The Curve (AUC) Operator.\n\nThis implementation computes the AUC according to forward output and label.\nIt is used very widely in binary classification evaluation. As a note:\nIf input label contains values other than 0 and 1, it will be cast\nto bool. You can find the relevant definitions here:\nhttps://en.wikipedia.org/wiki/Receiver_operating_characteristic#Area_under_the_curve\n\nThere are two types of possible curves:\n1. ROC: Receiver operating characteristic\n2. PR: Precision Recall\n",
4687 4688
 "inputs" : [ 
 { 
4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700
   "name" : "Out",
   "comment" : "A floating point 2D tensor, values are in the range [0, 1].Each row is sorted in descending order. This input should be theoutput of topk.Typically, this tensor indicates the probability of each label",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Indices",
   "comment" : "An int 2D tensor, indicating the indices of originaltensor before sorting. Typically, this tensor indicates which label the probability stands for.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Label",
   "comment" : "A 2D int tensor indicating the label of the training data.The height is batch size and width is always 1.",
4701 4702 4703 4704 4705
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4706 4707
   "name" : "AUC",
   "comment" : "A scalar representing the current area-under-the-curve.",
4708 4709 4710
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722
 "attrs" : [ 
 { 
   "name" : "curve",
   "type" : "string",
   "comment" : "Curve type, can be 'ROC' or 'PR'.",
   "generated" : 0
 }, { 
   "name" : "num_thresholds",
   "type" : "int",
   "comment" : "The number of thresholds to use when discretizing the roc curve.",
   "generated" : 0
 } ] 
4723
},{
4724 4725
 "type" : "read_from_array",
 "comment" : "\nReadFromArray Operator.\n\nRead a LoDTensor from a LoDTensor Array.\n\nAssume $T$ is LoDTensor, $i$ is the subscript of the array, and $A$ is the array. The\nequation is\n\n$$T = A[i]$$\n\n",
4726 4727 4728
 "inputs" : [ 
 { 
   "name" : "X",
4729
   "comment" : "(TensorArray) the array will be read from.",
4730 4731 4732
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4733 4734
   "name" : "I",
   "comment" : "(Tensor) the subscript index in tensor array. The number of element should be 1",
4735 4736 4737 4738 4739 4740
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4741
   "comment" : "(LoDTensor) the tensor will be read from.",
4742 4743 4744 4745 4746
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4747 4748
 "type" : "softplus",
 "comment" : "\nSoftplus Activation Operator.\n\n$y = \\ln(1 + e^{x})$\n\n",
4749 4750 4751
 "inputs" : [ 
 { 
   "name" : "X",
4752
   "comment" : "Input of Softplus operator",
4753 4754 4755 4756 4757 4758
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
4759
   "comment" : "Output of Softplus operator",
4760 4761 4762 4763 4764
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
4765 4766
 "type" : "split",
 "comment" : "\nSplit operator\n\nThis operator splits the input tensor into multiple sub-tensors.\n\nExample:\n  Input = [[1,2],\n           [3,4],\n           [5,6]]\n  sections = [2,1]\n  axis = 0\n  Output[0] = [[1,2],\n               [3,4]]\n  Output[1] = [[5,6]]\n\n    ",
4767 4768 4769
 "inputs" : [ 
 { 
   "name" : "X",
4770
   "comment" : "(Tensor) Input tensor of the split operator.",
4771 4772 4773 4774 4775 4776
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
4777
   "comment" : "(Tensor) Output tensors of the split operator.",
4778 4779 4780 4781 4782
   "duplicable" : 1,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
4783 4784 4785
   "name" : "sections",
   "type" : "int array",
   "comment" : "(vector<int>) the length of each output along the specified axis.",
4786 4787
   "generated" : 0
 }, { 
4788 4789 4790
   "name" : "num",
   "type" : "int",
   "comment" : "(int, default 0)Number of sub-tensors. This must evenly divide Input.dims()[axis]",
4791 4792
   "generated" : 0
 }, { 
4793 4794 4795
   "name" : "axis",
   "type" : "int",
   "comment" : "(int, default 0) The axis which the input will be splited on.",
4796 4797 4798
   "generated" : 0
 } ] 
},{
4799 4800
 "type" : "beam_search_decode",
 "comment" : "\nPack the result of Beam search op into SentenceIds and SentenceScores.\n",
4801 4802
 "inputs" : [ 
 { 
4803 4804
   "name" : "Ids",
   "comment" : "(LodTensorArray)score of the candidate words in each step",
4805 4806 4807
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4808 4809
   "name" : "Scores",
   "comment" : "(LodTensorArray)score of the candidate words in each step",
4810 4811 4812 4813 4814
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4815 4816
   "name" : "SentenceIds",
   "comment" : "(LodTensor)All possible result sentences of word ids",
4817 4818 4819
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4820 4821
   "name" : "SentenceScores",
   "comment" : "(LodTensor)All possible result sentences of word scores",
4822
   "duplicable" : 0,
4823
   "intermediate" : 0
4824 4825 4826
 } ], 
 "attrs" : [  ] 
},{
4827 4828
 "type" : "assign",
 "comment" : "Assign Operator\n\nOut = X,  when type in [LoDTensor/SelectedRows/LoDTensorArray]\nraise error if the type is not listed above.\n",
4829 4830
 "inputs" : [ 
 { 
4831 4832
   "name" : "X",
   "comment" : "(LoDTensor, SelectedRows or LoDTensorArray) The input variable could be LoDTensor, SelectedRows or LoDTensorArray.",
4833 4834 4835 4836 4837
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4838 4839
   "name" : "Out",
   "comment" : "(LoDTensor, SelectedRows or LoDTensorArray) The type of output is the same as input X.",
4840 4841
   "duplicable" : 0,
   "intermediate" : 0
4842 4843 4844 4845 4846 4847 4848 4849 4850
 } ], 
 "attrs" : [  ] 
},{
 "type" : "adam",
 "comment" : "\nAdam Optimizer.\n\nThis implements the Adam optimizer from Section 2 of the Adam\npaper : https://arxiv.org/abs/1412.6980.\nAdam is a first-order gradient-based optimization method based on\nadaptive estimates of lower-order moments.\n\nAdam updates:\n\n$$\nmoment\\_1\\_out = \\beta_1 * moment\\_1 + (1 - \\beta_1) * grad \\\\\nmoment\\_2_\\out = \\beta_2 * moment\\_2 + (1 - \\beta_2) * grad * grad \\\\\nlearning\\_rate = learning\\_rate *\n                  \\frac{\\sqrt{1 - \\beta_{2\\_pow}}}{1 - \\beta_{1\\_pow}} \\\\\nparam\\_out = param - learning\\_rate * \\frac{moment\\_1}{\\sqrt{moment\\_2} + \\epsilon}\n$$\n\n",
 "inputs" : [ 
 { 
   "name" : "Param",
   "comment" : "(Tensor) Input parameter",
4851 4852 4853
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4854 4855
   "name" : "Grad",
   "comment" : "(Tensor) Input gradient",
4856 4857
   "duplicable" : 0,
   "intermediate" : 0
4858 4859 4860
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor) Learning rate",
4861 4862 4863
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4864 4865
   "name" : "Moment1",
   "comment" : "(Tensor) Input first moment",
4866 4867 4868
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4869 4870
   "name" : "Moment2",
   "comment" : "(Tensor) Input second moment",
4871 4872 4873
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4874 4875
   "name" : "Beta1Pow",
   "comment" : "(Tensor) Input beta1 power accumulator",
4876 4877 4878
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4879 4880
   "name" : "Beta2Pow",
   "comment" : "(Tensor) Input beta2 power accumulator",
4881 4882 4883 4884 4885
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4886 4887
   "name" : "ParamOut",
   "comment" : "(Tensor) Output parameter",
4888 4889 4890
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4891 4892
   "name" : "Moment1Out",
   "comment" : "(Tensor) Output first moment",
4893 4894 4895
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4896 4897
   "name" : "Moment2Out",
   "comment" : "(Tensor) Output second moment",
4898 4899 4900 4901 4902
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
4903 4904 4905
   "name" : "beta1",
   "type" : "float",
   "comment" : "(float, default 0.9) Exponential decay rate for the first moment estimates.",
4906 4907
   "generated" : 0
 }, { 
4908
   "name" : "beta2",
4909
   "type" : "float",
4910
   "comment" : "(float, default 0.999) exponential decay rate for the second moment estimates.",
4911 4912 4913 4914
   "generated" : 0
 }, { 
   "name" : "epsilon",
   "type" : "float",
4915
   "comment" : "(float, default 1.0e-8) Constant for numerical stability",
4916 4917 4918
   "generated" : 0
 } ] 
},{
4919 4920
 "type" : "adadelta",
 "comment" : "\nAdadelta Optimizer.\n\nAdadelta optimizer is implemented as explained in:\nhttps://arxiv.org/abs/1212.5701\nAdadelta is a per-dimension adaptive learning rate method used\nfor gradient descent.\n\nAdadelta updates are as follows:\n\n$$\navg\\_squared\\_grad\\_out = \\rho * avg\\_squared\\_grad + (1 - \\rho) * grad * grad \\\\\nparam\\_update =  - \\sqrt{\\frac{avg\\_squared\\_update + \\epsilon}{avg\\_squared\\_grad\\_out + \\epsilon}} * grad \\\\\navg\\_squared\\_update\\_out = \\rho * avg\\_squared\\_update + (1 - \\rho) * {param\\_update}^2 \\\\\nparam\\_out = param + param\\_update\n$$\n\n",
4921 4922
 "inputs" : [ 
 { 
4923 4924
   "name" : "Param",
   "comment" : "(Tensor) Input parameter",
4925 4926 4927
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4928 4929
   "name" : "Grad",
   "comment" : "(Tensor) Input gradient",
4930 4931 4932
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4933 4934
   "name" : "AvgSquaredGrad",
   "comment" : "(Tensor) Input average of squared gradient",
4935 4936 4937
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4938 4939
   "name" : "AvgSquaredUpdate",
   "comment" : "(Tensor) Input average of squared parameter updates",
4940 4941 4942 4943 4944
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4945 4946
   "name" : "ParamOut",
   "comment" : "(Tensor) Output parameter",
4947 4948 4949
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4950 4951
   "name" : "AvgSquaredGradOut",
   "comment" : "(Tensor) Output average of squared gradient",
4952 4953 4954
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
4955 4956
   "name" : "AvgSquaredUpdateOut",
   "comment" : "(Tensor) Output average of squared parameter updates",
4957 4958 4959
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971
 "attrs" : [ 
 { 
   "name" : "rho",
   "type" : "float",
   "comment" : "(float, default 0.95) Exponential decay rate for squared gradients.",
   "generated" : 0
 }, { 
   "name" : "epsilon",
   "type" : "float",
   "comment" : "(float, default 1.0e-6) Constant for numerical stability",
   "generated" : 0
 } ] 
4972
},{
4973 4974
 "type" : "log",
 "comment" : "\nLog Activation Operator.\n\n$y = \\ln(x)$\n\nNatural logarithm of x.\n\n",
4975 4976 4977
 "inputs" : [ 
 { 
   "name" : "X",
4978
   "comment" : "Input of Log operator",
4979 4980 4981 4982 4983
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
4984 4985
   "name" : "Y",
   "comment" : "Output of Log operator",
4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "nce",
 "comment" : "\nCompute and return the noise-contrastive estimation training loss.\nSee [Noise-contrastive estimation: A new estimation principle for unnormalized statistical models](http://www.jmlr.org/proceedings/papers/v9/gutmann10a/gutmann10a.pdf).\nBy default this operator uses a uniform distribution for sampling.\n",
 "inputs" : [ 
 { 
   "name" : "Input",
   "comment" : "(Tensor) A tensor of shape [batch_size, dim].",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Label",
   "comment" : "(Tensor) A tensor of shape [batch_size, num_true_class]. 'num_true_class' is the number of target classes in each sample.The number of target classes per sample should be same. If you have a variable number of target classes, you can pad them out to a constant number by either repeating them or by padding with an otherwise unused class.)",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Weight",
   "comment" : "(Tensor) A tensor of shape [num_class, dim]. 'num_class' is the total number of class.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Bias",
   "comment" : "(Tensor) A tensor of shape [num_class, 1]. 'num_class' is the total number of class. It is a dispensable input.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "SampleWeight",
   "comment" : "(Tensor) A tensor of shape [batch_size, 1] storing a weight for each sample. And it is a dispensable input. The default value of sample is 1.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Cost",
   "comment" : "(Tensor) A tensor of shape [batch_size, 1]. Cost of samples.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "SampleLogits",
   "comment" : "An intermediate tensor of shape[batch_size, num_neg_samples + num_pos_samples].This tensor is output of forward kernel and used in backward kernel to compute grads.Given X is  the dot product of input tensor and sampled labels' weights.Then 'SampleLogits' is sigmoid(X).",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "SampleLabels",
   "comment" : "An intermediate tensor of shape[batch_size, num_neg_samples + num_pos_samples].This tensor is output of forward kernel and used in backward kernel to compute grads.",
   "duplicable" : 0,
   "intermediate" : 1
 } ], 
 "attrs" : [ 
 { 
   "name" : "num_total_classes",
   "type" : "int",
   "comment" : "Total number of classes in all samples.",
   "generated" : 0
 }, { 
   "name" : "num_neg_samples",
   "type" : "int",
   "comment" : "The number of negative classes. The default value is 10.",
   "generated" : 0
 }, { 
   "name" : "custom_neg_classes",
   "type" : "int array",
   "comment" : "This attribute only be used in unitest. Classes in this list wiil be used as negative classes for every samples. Under normal conditions, user should avoid setting this attribute.",
   "generated" : 0
 } ] 
},{
 "type" : "linear_chain_crf",
 "comment" : "\nLinearChainCRF Operator.\n\nConditional Random Field defines an undirected probabilistic graph with nodes\ndenoting random variables and edges denoting dependencies between these\nvariables. CRF learns the conditional probability $P(Y|X)$, where\n$X = (x_1, x_2, ... , x_n)$ are structured inputs and\n$Y = (y_1, y_2, ... , y_n)$ are labels for the inputs.\n\nLinear chain CRF is a special case of CRF that is useful for sequence labeling\ntask. Sequence labeling tasks do not assume a lot of conditional\nindependences among inputs. The only constraint they impose is that the input\nand output must be linear sequences. Thus, the graph of such a CRF is a simple\nchain or a line, which results in the linear chain CRF.\n\nThis operator implements the Forward-Backward algorithm for the linear chain\nCRF. Please refer to http://www.cs.columbia.edu/~mcollins/fb.pdf and\nhttp://cseweb.ucsd.edu/~elkan/250Bwinter2012/loglinearCRFs.pdf for details.\n\nEquation:\n1. Denote Input(Emission) to this operator as $x$ here.\n2. The first D values of Input(Transition) to this operator are for starting\nweights, denoted as $a$ here.\n3. The next D values of Input(Transition) of this operator are for ending\nweights, denoted as $b$ here.\n4. The remaning values of Input(Transition) are for transition weights,\ndenoted as $w$ here.\n5. Denote Input(Label) as $s$ here.\n\nThe probability of a sequence $s$ of length $L$ is defined as:\n$$P(s) = (1/Z) \\exp(a_{s_1} + b_{s_L}\n                + \\sum_{l=1}^L x_{s_l}\n                + \\sum_{l=2}^L w_{s_{l-1},s_l})$$\n\nwhere $Z$ is a normalization value so that the sum of $P(s)$ over\nall possible sequences is 1, and $x$ is the emission feature weight\nto the linear chain CRF.\n\nFinally, the linear chain CRF operator outputs the logarithm of the conditional\nlikelihood of each training sample in a mini-batch.\n\nNOTE:\n1. The feature function for a CRF is made up of the emission features and the\ntransition features. The emission feature weights are NOT computed in\nthis operator. They MUST be computed first before this operator is called.\n\n2. Because this operator performs global normalization over all possible\nsequences internally, it expects UNSCALED emission feature weights.\nPlease do not call this op with the emission feature being output of any\nnonlinear activation.\n\n3. The 2nd dimension of Input(Emission) MUST be equal to the tag number.\n\n",
 "inputs" : [ 
 { 
   "name" : "Emission",
   "comment" : "(LoDTensor, default LoDTensor<float>) A 2-D LoDTensor with shape [N x D], where N is the size of the mini-batch and D is the total tag number. The unscaled emission weight matrix for the linear chain CRF. ",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Transition",
   "comment" : "(Tensor, default Tensor<float>) A 2-D Tensor with shape [(D + 2) x D]. The learnable parameter for the linear_chain_crf operator. See more details in the operator's comments.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Label",
   "comment" : "(LoDTensor, default LoDTensor<int64_t>) A LoDTensor with shape [N x 1], where N is the total element number in a mini-batch. The ground truth.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Alpha",
   "comment" : "(Tensor, default Tensor<float>) A 2-D Tensor with shape [N x D]. The forward vectors for the entire batch. Denote it as $lpha$. $lpha$ is a memo table used to calculate the normalization factor in CRF. $lpha[k, v]$ stores the unnormalized probabilites of all possible unfinished sequences of tags that end at position $k$ with tag $v$. For each $k$, $lpha[k, v]$ is a vector of length $D$ with a component for each tag value $v$. This vector is called a forward vecotr and will also be used in backward computations.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "EmissionExps",
   "comment" : "(Tensor, default Tensor<float>) A 2-D Tensor with shape [N x D]. The exponentials of Input(Emission). This is an intermediate computational result in forward computation, and will be reused in backward computation.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "TransitionExps",
   "comment" : "(Tensor, default Tensor<float>) A 2-D Tensor with shape [(D + 2) x D]. The exponentials of Input(Transition). This is an intermediate computational result in forward computation, and will be reused in backward computation.",
   "duplicable" : 0,
   "intermediate" : 1
 }, { 
   "name" : "LogLikelihood",
   "comment" : "(Tensor, default Tensor<float>) The logarithm of the conditional likelihood of each training sample in a mini-batch. This is a 2-D tensor with shape [S x 1], where S is the sequence number in a mini-batch. Note: S is equal to the sequence number in a mini-batch. The output is no longer a LoDTensor.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "logsigmoid",
 "comment" : "\nLogsigmoid Activation Operator\n\n$$y = \\log \\frac{1}{1 + e^{-x}}$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of LogSigmoid operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of LogSigmoid operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137
},{
 "type" : "row_conv",
 "comment" : "\nRow-convolution Operator.\n\nThe row convolution is called lookahead convolution.  This operator was \nintroduced in the following paper for DeepSpeech2:\nhttp://www.cs.cmu.edu/~dyogatam/papers/wang+etal.iclrworkshop2016.pdf \n\nThe main motivation is that a bidirectional RNN, useful in DeepSpeech \nlike speech models, learns representation for a sequence by performing a \nforward and a backward pass through the entire sequence. However, unlike \nunidirectional RNNs, bidirectional RNNs are challenging to deploy in an online\nand low-latency setting. The lookahead convolution incorporates information \nfrom future subsequences in a computationally efficient manner to improve \nunidirectional recurrent neural networks. The row convolution operator is \ndifferent from the 1D sequence convolution, and is computed as follows:\n\nGiven an input sequence $in$ of length $t$ and input dimension $d$, \nand a filter ($W$) of size $context \\times d$, \nthe output sequence is convolved as:\n\n$$\nout_{i, :} = \\sum_{j=i}^{i + context} in_{j,:} \\dot W_{i-j, :}\n$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(LoDTensor), the input(X) is a LodTensor, which supports variable time-length input sequences. The underlying tensor in this LoDTensor is a matrix with shape (T x N), where T is the total time steps in this mini-batch and N is the input data dimension.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Filter",
   "comment" : "(Tensor), the input(Filter) is a learnable parameter. It is a 2-D tensor with shape (future_context x N), where, future_context is the future context length and N is the data dimension.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(LoDTensor), the output(Out) is a LodTensor, which supports variable time-length input sequences. The underlying tensor in this LodTensor is a matrix with shape T x N, i.e., the same shape as X.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310
},{
 "type" : "exp",
 "comment" : "\nExp Activation Operator.\n\n$y = e^x$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Exp operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of Exp operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
},{
 "type" : "soft_relu",
 "comment" : "\nSoftRelu Activation Operator.\n\n$y = \\ln(1 + \\exp(\\max(\\min(x, threshold), threshold))$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of SoftRelu operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of SoftRelu operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "threshold",
   "type" : "float",
   "comment" : "The threshold value of SoftRelu",
   "generated" : 0
 } ] 
},{
 "type" : "softshrink",
 "comment" : "\nSoftshrink Activation Operator.\n\n$$\ny = \\begin{cases} \n    x - \\lambda, \\text{if } x > \\lambda \\\\\n    x + \\lambda, \\text{if } x < -\\lambda \\\\\n    0,  \\text{otherwise}\n    \\end{cases}\n$$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Softshrink operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of Softshrink operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "lambda",
   "type" : "float",
   "comment" : "non-negative offset",
   "generated" : 0
 } ] 
},{
 "type" : "maxout",
 "comment" : "\nMaxOut Operator.\n\nAssumed the input shape is (N, Ci, H, W).\nThe output shape is (N, Co, H, W).\nThen $Co = Ci / groups$ and the operator formula is as follows:\n\n$$\ny_{si+j} = \\max_k x_{gsi + sk + j} \\\\\ng = groups \\\\\ns = \\frac{input.size}{num\\_channels} \\\\\n0 \\le i < \\frac{num\\_channels}{groups} \\\\\n0 \\le j < s \\\\\n0 \\le k < groups\n$$\n\nPlease refer to Paper:\n  - Maxout Networks: http://www.jmlr.org/proceedings/papers/v28/goodfellow13.pdf\n  - Multi-digit Number Recognition from Street View \\\n    Imagery using Deep Convolutional Neural Networks: \\\n    https://arxiv.org/pdf/1312.6082v4.pdf\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "(Tensor) The input tensor of maxout operator. The format of input tensor is NCHW. Where N is batch size, C is the number of channels, H and W is the height and width of feature.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Out",
   "comment" : "(Tensor) The output tensor of maxout operator.The format of output tensor is also NCHW.Where N is batch size, C is the number of channels, H and W is the height and width of feature.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "groups",
   "type" : "int",
   "comment" : "\"Specifies how many groups the input tensor will be split\"\n        \"in the channel dimension. And the number of output channel is \"\n        \"the number of channels divided by groups..\"\n        ",
   "generated" : 0
 } ] 
},{
 "type" : "ftrl",
 "comment" : "\nFTRL (Follow The Regularized Leader) Operator.\n\nOptimizer that implements the FTRL algorithm:\n\n$$\nnew\\_accum = squared\\_accum + grad^2 \\\\\nif (lr\\_power == -0.5) {\n   linear\\_accum += grad - (\\surd(new\\_accum) - \\surd(squared\\_accum)) /\n                   (learning\\_rate * param) \\\\\n} else {\n   linear\\_accum += grad -\n                  (new\\_accum^{-lr\\_power} - accum^{-lr\\_power}) /\n                  (learning\\_rate * param) \\\\\n}\n\nx = (l1 * sign(linear\\_accum) - linear\\_accum)\nif (lr\\_power == -0.5) {\n   y = \\frac{\\surd(new\\_accum)}{learning\\_rate} + (2 * l2) \\\\\n   pre\\_shrink = \\frac{x}{y} \\\\\n   param = (abs(linear\\_accum) > l1).select(pre\\_shrink, 0.0) \\\\\n} else {\n   y = \\frac{new\\_accum^{-lr\\_power}}{learning\\_rate} + (2 * l2) \\\\\n   pre\\_shrink = \\frac{x}{y} \\\\\n   param = (abs(linear\\_accum) > l1).select(pre\\_shrink, 0.0) \\\\\n}\nsquared\\_accum += grad^2;\n$$\n\nThe paper that proposed Follow The Regularized Leader (FTRL):\n(https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf)\n\n",
 "inputs" : [ 
 { 
   "name" : "Param",
   "comment" : "(Tensor, default Tensor<float>) Input parameter value that has to be updated.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "SquaredAccumulator",
   "comment" : "(Tensor, default Tensor<float>) Accumulator that accumulates squared gradients.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LinearAccumulator",
   "comment" : "(Tensor, default Tensor<float>) Accumulator that accumulates linear gradients.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "Grad",
   "comment" : "(Tensor, default Tensor<float>) Input gradient of the parameter.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LearningRate",
   "comment" : "(Tensor, default Tensor<float>) The learning rate should be a tensor of size 1.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "ParamOut",
   "comment" : "(Tensor) Output updated parameter value.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "SquaredAccumOut",
   "comment" : "(Tensor) Output accumulated squared gradients.",
   "duplicable" : 0,
   "intermediate" : 0
 }, { 
   "name" : "LinearAccumOut",
   "comment" : "(Tensor) Output accumulated linear gradients.",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [ 
 { 
   "name" : "l1",
   "type" : "float",
   "comment" : "(float, default 0.0) L1 regularization strength.",
   "generated" : 0
 }, { 
   "name" : "l2",
   "type" : "float",
   "comment" : "(float, default 0.0) L2 regularization strength.",
   "generated" : 0
 }, { 
   "name" : "lr_power",
   "type" : "float",
   "comment" : "(float, default -0.5f) Learning Rate Power.",
   "generated" : 0
 } ] 
},{
 "type" : "round",
 "comment" : "\nRound Activation Operator.\n\n$y = [x]$\n\n",
 "inputs" : [ 
 { 
   "name" : "X",
   "comment" : "Input of Round operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "outputs" : [ 
 { 
   "name" : "Y",
   "comment" : "Output of Round operator",
   "duplicable" : 0,
   "intermediate" : 0
 } ], 
 "attrs" : [  ] 
}]