multiary.cc 105.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

15
#include "paddle/phi/infermeta/multiary.h"
16

17
#include <vector>
18

H
hong 已提交
19
#include "paddle/phi/common/layout.h"
20
#include "paddle/phi/common/scalar.h"
H
hong 已提交
21
#include "paddle/phi/core/infermeta_utils.h"
22
#include "paddle/phi/core/meta_tensor.h"
23
#include "paddle/phi/kernels/funcs/common_shape.h"
24 25
#include "paddle/phi/kernels/funcs/concat_funcs.h"
namespace phi {
26

27 28
std::vector<DDim> GetMetaTensorsDim(
    const std::vector<const MetaTensor*>& tensors) {
29 30 31 32 33 34 35 36
  std::vector<DDim> dims;
  dims.reserve(tensors.size());
  for (const MetaTensor* tensor : tensors) {
    dims.emplace_back(tensor->dims());
  }
  return dims;
}

F
From00 已提交
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
void AdadeltaInferMeta(const MetaTensor& param,
                       const MetaTensor& grad,
                       const MetaTensor& avg_squared_grad,
                       const MetaTensor& avg_squared_update,
                       float rho,
                       float epsilon,
                       MetaTensor* param_out,
                       MetaTensor* avg_squared_grad_out,
                       MetaTensor* avg_squared_update_out) {
  auto param_dims = param.dims();
  PADDLE_ENFORCE_EQ(
      param_dims,
      grad.dims(),
      errors::InvalidArgument(
          "Param and grad input of AdadeltaOp should have same dimension."));
  PADDLE_ENFORCE_EQ(
      param_dims,
      avg_squared_grad.dims(),
      errors::InvalidArgument("Param and AvgSquaredGrad input of AdadeltaOp "
                              "should have same dimension"));
  PADDLE_ENFORCE_EQ(
      param_dims,
      avg_squared_update.dims(),
      errors::InvalidArgument("Param and AvgSquaredUpdate input of AdadeltaOp "
                              "should have same dimension"));

  param_out->set_dims(param_dims);
  param_out->set_dtype(param.dtype());

  avg_squared_grad_out->set_dims(param_dims);
  avg_squared_grad_out->set_dtype(avg_squared_grad.dtype());

  avg_squared_update_out->set_dims(param_dims);
  avg_squared_update_out->set_dtype(avg_squared_update.dtype());
}

H
hong 已提交
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
void AdagradInferMeta(const MetaTensor& param,
                      const MetaTensor& grad,
                      const MetaTensor& moment,
                      const MetaTensor& learning_rate,
                      float epsilon,
                      MetaTensor* param_out,
                      MetaTensor* moment_out) {
  auto lr_dims = learning_rate.dims();
  PADDLE_ENFORCE_EQ(
      phi::product(lr_dims),
      1,
      phi::errors::InvalidArgument("LearningRate should have one element"));
  auto param_dims = param.dims();

  PADDLE_ENFORCE_EQ(
      param_dims,
      moment.dims(),
      phi::errors::InvalidArgument("Param and Moment input of AdagradOp "
                                   "should have the same dimension."));

  param_out->set_dims(param_dims);
  param_out->set_dtype(param.dtype());
  moment_out->set_dims(param_dims);
  moment_out->set_dtype(moment.dtype());
}

99 100 101 102 103 104 105
void AdamInferMeta(const MetaTensor& param,
                   const MetaTensor& grad,
                   const MetaTensor& learning_rate,
                   const MetaTensor& moment1,
                   const MetaTensor& moment2,
                   const MetaTensor& beta1_pow,
                   const MetaTensor& beta2_pow,
106 107
                   const MetaTensor& master_param,
                   const MetaTensor& skip_update,
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
                   const Scalar& beta1,
                   const Scalar& beta2,
                   const Scalar& epsilon,
                   bool lazy_mode,
                   int64_t min_row_size_to_use_multithread,
                   bool multi_precision,
                   bool use_global_beta_pow,
                   MetaTensor* param_out,
                   MetaTensor* moment1_out,
                   MetaTensor* moment2_out,
                   MetaTensor* beta1_pow_out,
                   MetaTensor* beta2_pow_out,
                   MetaTensor* master_param_outs) {
  auto lr_dims = learning_rate.dims();
  PADDLE_ENFORCE_EQ(
      phi::product(lr_dims),
      1,
      errors::InvalidArgument(
          "The number of LearningRate shall be 1, but received %d. Maybe "
          "the Input variable LearningRate has not "
          "been initialized. You may need to confirm "
          "if you put exe.run(startup_program) "
          "after optimizer.minimize function.",
          phi::product(lr_dims)));
  auto beta1_pow_dims = beta1_pow.dims();
  VLOG(3) << "dims of Beta1Pow : [" << beta1_pow_dims << "]";
  PADDLE_ENFORCE_GE(phi::product(beta1_pow_dims),
                    1,
                    errors::InvalidArgument(
                        "The size of Beta1 power accumulator should be greater "
                        "than 0, but received %d.",
                        phi::product(beta1_pow_dims)));
  auto beta2_pow_dims = beta2_pow.dims();
  VLOG(3) << "dims of Beta2Pow : [" << beta2_pow_dims << "]";
  PADDLE_ENFORCE_GE(phi::product(beta2_pow_dims),
                    1,
                    errors::InvalidArgument(
                        "The size of Beta2 power accumulator should be greater "
                        "than 0, but received %d.",
                        phi::product(beta2_pow_dims)));

  auto param_dims = param.dims();
  PADDLE_ENFORCE_EQ(
      param_dims,
      moment1.dims(),
      errors::InvalidArgument(
          "Param and Moment1 input of AdamOp should have same dimension. But "
          "received Param dims: [%s], Moment1 dims: [%s].",
          param_dims,
          moment1.dims()));
  PADDLE_ENFORCE_EQ(
      param_dims,
      moment2.dims(),
      errors::InvalidArgument(
          "Param and Moment2 input of AdamOp should have same dimension. But "
          "received Param dims: [%s], Moment2 dims: [%s].",
          param_dims,
          moment2.dims()));

  param_out->set_dims(param_dims);
  param_out->set_dtype(param.dtype());

  moment1_out->set_dims(param_dims);
  moment1_out->set_dtype(moment1.dtype());
  moment2_out->set_dims(param_dims);
  moment2_out->set_dtype(moment2.dtype());

  beta1_pow_out->set_dims(beta1_pow_dims);
  beta1_pow_out->set_dtype(beta1_pow.dtype());
  beta2_pow_out->set_dims(beta2_pow_dims);
  beta2_pow_out->set_dtype(beta2_pow.dtype());
}

F
From00 已提交
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
void AdamaxInferMeta(const MetaTensor& param,
                     const MetaTensor& grad,
                     const MetaTensor& learning_rate,
                     const MetaTensor& moment,
                     const MetaTensor& inf_norm,
                     const MetaTensor& beta1_pow,
                     float beta1,
                     float beta2,
                     float epsilon,
                     MetaTensor* param_out,
                     MetaTensor* moment_out,
                     MetaTensor* inf_norm_out) {
  auto lr_dims = learning_rate.dims();
  PADDLE_ENFORCE_NE(
      product(lr_dims),
      0,
      errors::InvalidArgument("Maybe the Input variable LearningRate has not "
                              "been initialized. You may need to confirm "
                              "if you put exe.run(startup_program) "
                              "after optimizer.minimize function."));
  PADDLE_ENFORCE_EQ(
      product(lr_dims),
      1,
      errors::InvalidArgument("Learning rate should have 1 dimension"));
  auto beta1_pow_dims = beta1_pow.dims();
  PADDLE_ENFORCE_EQ(product(beta1_pow_dims),
                    1,
                    errors::InvalidArgument(
                        "Beta1 power accumulator should have 1 dimension"));
  auto param_dims = param.dims();
  PADDLE_ENFORCE_EQ(
      param_dims,
      grad.dims(),
      errors::InvalidArgument(
          "Param and Grad input of AdamaxOp should have same dimension"));
  PADDLE_ENFORCE_EQ(
      param_dims,
      moment.dims(),
      errors::InvalidArgument(
          "Param and Moment input of AdamaxOp should have same dimension"));
  PADDLE_ENFORCE_EQ(
      param_dims,
      inf_norm.dims(),
      errors::InvalidArgument(
          "Param and InfNorm input of AdamaxOp should have same dimension"));

  param_out->set_dims(param_dims);
  param_out->set_dtype(param.dtype());

  moment_out->set_dims(param_dims);
  moment_out->set_dtype(moment.dtype());

  inf_norm_out->set_dims(param_dims);
  inf_norm_out->set_dtype(inf_norm.dtype());
}

237 238 239 240 241 242 243
void AdamwInferMeta(const MetaTensor& param,
                    const MetaTensor& grad,
                    const MetaTensor& learning_rate,
                    const MetaTensor& moment1,
                    const MetaTensor& moment2,
                    const MetaTensor& beta1_pow,
                    const MetaTensor& beta2_pow,
244 245
                    const MetaTensor& master_param,
                    const MetaTensor& skip_update,
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
                    const Scalar& beta1,
                    const Scalar& beta2,
                    const Scalar& epsilon,
                    float lr_ratio,
                    float coeff,
                    bool with_decay,
                    bool lazy_mode,
                    int64_t min_row_size_to_use_multithread,
                    bool multi_precision,
                    bool use_global_beta_pow,
                    MetaTensor* param_out,
                    MetaTensor* moment1_out,
                    MetaTensor* moment2_out,
                    MetaTensor* beta1_pow_out,
                    MetaTensor* beta2_pow_out,
                    MetaTensor* master_param_outs) {
  AdamInferMeta(param,
                grad,
                learning_rate,
                moment1,
                moment2,
                beta1_pow,
                beta2_pow,
                master_param,
                skip_update,
                beta1,
                beta2,
                epsilon,
                lazy_mode,
                min_row_size_to_use_multithread,
                multi_precision,
                use_global_beta_pow,
                param_out,
                moment1_out,
                moment2_out,
                beta1_pow_out,
                beta2_pow_out,
                master_param_outs);
}

286
void AddNInferMeta(const std::vector<const MetaTensor*>& x,
287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
                   MetaTensor* out,
                   MetaConfig config) {
  auto N = x.size();
  PADDLE_ENFORCE_GT(
      N,
      0,
      phi::errors::InvalidArgument(
          "The input tensor X's dimensions of SumOp "
          "should be larger than 0. But received X's dimensions %d.",
          N));
  if (N == 1) {
    VLOG(3) << "Warning: SumOp have only one input, may waste memory";
  }

  phi::DDim in_dim({0});
  for (size_t i = 0; i < x.size(); ++i) {
    auto x_dim = x[i]->dims();
Y
YuanRisheng 已提交
304 305 306 307
    // x_dim.size() == 1 means the real dim of selected rows is [0]
    if (x[i]->is_selected_rows() && x_dim.size() == 1) {
      continue;
    }
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
    if (phi::product(x_dim) == 0) {
      continue;
    }
    if (phi::product(in_dim) == 0) {
      in_dim = x_dim;
    } else {
      if (config.is_runtime) {
        PADDLE_ENFORCE_EQ(in_dim,
                          x_dim,
                          phi::errors::InvalidArgument(
                              "The input tensor X of SumOp must"
                              " have same shape. But received X[0]'s shape = "
                              "[%s], X[%d]'s shape = [%s].",
                              in_dim,
                              i,
                              x_dim));
      } else {
        PADDLE_ENFORCE_EQ(
            in_dim.size(),
            x_dim.size(),
            phi::errors::InvalidArgument(
                "The input tensor X of SumOp must have same "
                "dimensions. But received X[0]'s dimensions = %d, X[0]'s "
                "shape = "
                "[%s], X[%d]'s dimensions = %d, X[%d]'s shape = [%s].",
                in_dim.size(),
                in_dim,
                i,
                x_dim.size(),
                i,
                x_dim));
        // if in_dim or x_dim has -1, not check equal
        for (int j = 0; j < x_dim.size(); ++j) {
          if (x_dim[j] == -1 || in_dim[j] == -1) {
            continue;
          }
          PADDLE_ENFORCE_EQ(
              in_dim[j],
              x_dim[j],
              phi::errors::InvalidArgument(
                  "The input tensor X of SumOp must have same shape "
                  "if not -1."
                  "But received X[0]'s shape = [%s], X[%d]'s shape = [%s].",
                  in_dim,
                  i,
                  x_dim));
        }
      }
    }
  }
  out->set_dims(in_dim);
  out->share_lod(*x[0]);
}

Y
YuanRisheng 已提交
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
// TODO(YuanRisheng) This InferMeta is used in Fluid
//                   and will be deleted in the future.
void AddNTensorArrayInferMeta(const std::vector<const MetaTensor*>& x,
                              MetaTensor* out,
                              MetaConfig config) {
  int64_t max_length = 0;
  bool has_tensor_array = false;
  for (auto input : x) {
    if (input->is_tensor_array()) {
      has_tensor_array = true;
      // if input is lod_tensor_array, dims() will return its size (one element)
      max_length =
          input->dims()[0] > max_length ? input->dims()[0] : max_length;
    }
  }

  if (has_tensor_array) {
    if (out->is_tensor_array()) {
      out->set_dims(make_ddim({max_length}));
    }
  } else {
    AddNInferMeta(x, out, config);
  }
}

387 388 389 390
void AucInferMeta(const MetaTensor& input,
                  const MetaTensor& label,
                  const MetaTensor& stat_pos,
                  const MetaTensor& stat_neg,
391
                  const MetaTensor& ins_tag_weight,
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
                  const std::string& curve,
                  int num_thresholds,
                  int slide_steps,
                  MetaTensor* auc,
                  MetaTensor* stat_pos_out,
                  MetaTensor* stat_neg_out,
                  MetaConfig config) {
  auto predict_dims = input.dims();
  auto label_dims = label.dims();
  PADDLE_ENFORCE_GE(
      predict_dims.size(),
      2,
      phi::errors::InvalidArgument(
          "The Input(Predict) has not been initialized properly. The "
          "shape of Input(Predict) = [%s], the shape size must be "
          "greater_equal 2.",
          predict_dims));
  auto predict_width = predict_dims[1];
  PADDLE_ENFORCE_NE(
      phi::product(predict_dims),
      0,
      phi::errors::InvalidArgument(
          "The Input(Predict) has not been initialized properly. The "
          "shape of Input(Predict) = [%s], the shape can not involes 0.",
          predict_dims));
  PADDLE_ENFORCE_NE(
      phi::product(label_dims),
      0,
      phi::errors::InvalidArgument(
          "The Input(Label) has not been initialized properly. The "
          "shape of Input(Label) = [%s], the shape can not involes 0.",
          label_dims));
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
  if (config.is_runtime) {
    PADDLE_ENFORCE_LE(
        predict_width,
        2,
        phi::errors::InvalidArgument("Only support binary classification,"
                                     "prediction dims[1] should be 1 or 2"));
  }
  auto predict_height = input.dims()[0];
  auto label_height = label.dims()[0];

  if (config.is_runtime) {
    PADDLE_ENFORCE_EQ(
        predict_height,
        label_height,
        phi::errors::InvalidArgument("Out and Label should have same height."));
  }

  int num_pred_buckets = num_thresholds + 1;

  PADDLE_ENFORCE_GE(
      num_pred_buckets,
      1,
      phi::errors::InvalidArgument("num_thresholds must larger than 1"));
  PADDLE_ENFORCE_GE(
      slide_steps,
      0,
      phi::errors::InvalidArgument("slide_steps must be natural number"));

  auc->set_dims({1});
  auc->set_dtype(DataType::INT64);

  if (slide_steps) {
    stat_pos_out->set_dims({(1 + slide_steps) * num_pred_buckets + 1});
    stat_pos_out->set_dtype(DataType::INT64);
    stat_neg_out->set_dims({(1 + slide_steps) * num_pred_buckets + 1});
    stat_neg_out->set_dtype(DataType::INT64);
  } else {
    stat_pos_out->set_dims({1, num_pred_buckets});
    stat_pos_out->set_dtype(DataType::INT64);
    stat_neg_out->set_dims({1, num_pred_buckets});
    stat_neg_out->set_dtype(DataType::INT64);
  }
}

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
void AverageAccumulatesInferMeta(const MetaTensor& param,
                                 const MetaTensor& in_sum_1,
                                 const MetaTensor& in_sum_2,
                                 const MetaTensor& in_sum_3,
                                 const MetaTensor& in_num_accumulates,
                                 const MetaTensor& in_old_num_accumulates,
                                 const MetaTensor& in_num_updates,
                                 float average_window,
                                 int64_t max_average_window,
                                 int64_t min_average_window,
                                 MetaTensor* out_sum_1,
                                 MetaTensor* out_sum_2,
                                 MetaTensor* out_sum_3,
                                 MetaTensor* out_num_accumulates,
                                 MetaTensor* out_old_num_accumulates,
                                 MetaTensor* out_num_updates) {
  // auto in_dim = param.dims;
  PADDLE_ENFORCE_NE(
      out_sum_1,
      nullptr,
      errors::NotFound(
          "Output(out_sum_1) of AverageAccumulates should not be null."));
  PADDLE_ENFORCE_NE(
      out_sum_2,
      nullptr,
      errors::NotFound(
          "Output(out_sum_2) of AverageAccumulates should not be null."));
  PADDLE_ENFORCE_NE(
      out_sum_3,
      nullptr,
      errors::NotFound(
          "Output(out_sum_3) of AverageAccumulates should not be null."));
  PADDLE_ENFORCE_NE(out_num_accumulates,
                    nullptr,
                    errors::NotFound("Output(out_num_accumulates) of "
                                     "AverageAccumulates should not be null."));

  PADDLE_ENFORCE_NE(out_old_num_accumulates,
                    nullptr,
                    errors::NotFound("Output(out_old_num_accumulates) of "
                                     "AverageAccumulates should not be null."));

  PADDLE_ENFORCE_NE(
      out_num_updates,
      nullptr,
      errors::NotFound(
          "Output(out_num_updates) of AverageAccumulates should not be null."));

  out_sum_1->set_dims(in_sum_1.dims());
  out_sum_1->set_dtype(in_sum_1.dtype());
  out_sum_2->set_dims(in_sum_2.dims());
  out_sum_2->set_dtype(in_sum_2.dtype());
  out_sum_3->set_dims(in_sum_3.dims());
  out_sum_3->set_dtype(in_sum_3.dtype());
  out_num_accumulates->set_dims({1});
  out_num_accumulates->set_dtype(in_num_accumulates.dtype());
  out_old_num_accumulates->set_dims({1});
  out_old_num_accumulates->set_dtype(in_old_num_accumulates.dtype());
  out_num_updates->set_dims({1});
  out_num_updates->set_dtype(in_num_updates.dtype());
}

H
hong 已提交
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
void BatchNormInferMeta(const MetaTensor& x,
                        const MetaTensor& scale,
                        const MetaTensor& bias,
                        const MetaTensor& mean,
                        const MetaTensor& variance,
                        float momentum,
                        float epsilon,
                        const std::string& data_layout_str,
                        bool is_test,
                        bool use_global_stats,
                        bool trainable_statistics,
                        bool fuse_with_relu,
                        MetaTensor* y,
                        MetaTensor* mean_out,
                        MetaTensor* variance_out,
                        MetaTensor* saved_mean,
                        MetaTensor* saved_variance,
                        MetaTensor* reserve_space,
                        MetaConfig config) {
  const auto x_dims = x.dims();
  for (int i = 0; i < x_dims.size(); i++) {
    PADDLE_ENFORCE_EQ(
        (x_dims[i] == -1) || (x_dims[i] > 0),
        true,
        phi::errors::InvalidArgument(
            "Each dimension of input tensor is expected to be -1 or a "
557
            "positive number, but received %d. Input's shape is [%s].",
H
hong 已提交
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
            x_dims[i],
            x_dims));
  }

  const DataLayout data_layout =
      paddle::framework::StringToDataLayout(data_layout_str);

  PADDLE_ENFORCE_GE(
      x_dims.size(),
      2,
      phi::errors::InvalidArgument(
          "ShapeError: the dimension of input "
          "X must greater than or equal to 2. But received: the shape of input "
          "X = [%s], the dimension of input X =[%d]",
          x_dims,
          x_dims.size()));
  PADDLE_ENFORCE_LE(
      x_dims.size(),
      5,
      phi::errors::InvalidArgument(
          "ShapeError: the dimension of input X "
          "must smaller than or equal to 5. But received: the shape of input X "
          "= [%s], the dimension of input X = [%d]",
          x_dims,
          x_dims.size()));

  const int64_t C = ((config.is_run_mkldnn_kernel == true) ||
                             (data_layout == DataLayout::kNCHW)
                         ? x_dims[1]
                         : x_dims[x_dims.size() - 1]);
  auto scale_dim = scale.dims();
  auto bias_dim = bias.dims();

  PADDLE_ENFORCE_EQ(
      scale_dim.size(),
      1UL,
      phi::errors::InvalidArgument(
          "ShapeError: the dimension of scale must equal to 1."
          "But received: the shape of scale is [%s], the dimension "
          "of scale is [%d]",
          scale_dim,
          scale_dim.size()));
  PADDLE_ENFORCE_EQ(bias_dim.size(),
                    1UL,
                    phi::errors::InvalidArgument(
                        "ShapeError: the dimension of bias must equal to 1."
                        "But received: the shape of bias is [%s],the dimension "
                        "of bias is [%d]",
                        bias_dim,
                        bias_dim.size()));

  bool check = true;
  if ((!config.is_runtime) &&
      (phi::product(scale_dim) <= 0 || phi::product(bias_dim) <= 0)) {
    check = false;
  }

  if (check) {
    PADDLE_ENFORCE_EQ(scale_dim[0],
                      C,
                      phi::errors::InvalidArgument(
                          "ShapeError: the shape of scale must equal to [%d]"
                          "But received: the shape of scale is [%d]",
                          C,
                          scale_dim[0]));
    PADDLE_ENFORCE_EQ(bias_dim[0],
                      C,
                      phi::errors::InvalidArgument(
                          "ShapeError: the shape of bias must equal to [%d]"
                          "But received: the shape of bias is [%d]",
                          C,
                          bias_dim[0]));
  }
  y->set_dims(x_dims);
  mean_out->set_dims({C});
  variance_out->set_dims({C});
634 635 636 637 638 639
  if (saved_mean) {
    saved_mean->set_dims({C});
  }
  if (saved_variance) {
    saved_variance->set_dims({C});
  }
H
hong 已提交
640 641 642
  y->share_lod(x);
}

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
void BatchNormInferInferMeta(const MetaTensor& x,
                             const MetaTensor& scale,
                             const MetaTensor& bias,
                             const MetaTensor& mean,
                             const MetaTensor& variance,
                             float momentum,
                             float epsilon,
                             const std::string& data_layout,
                             MetaTensor* y,
                             MetaTensor* mean_out,
                             MetaTensor* variance_out,
                             MetaConfig config) {
  BatchNormInferMeta(x,
                     scale,
                     bias,
                     mean,
                     variance,
                     momentum,
                     epsilon,
                     data_layout,
                     /*is_test=*/true,
                     /*use_global_stats=*/false,
                     /*trainable_statistics=*/false,
                     /*fuse_with_relu=*/false,
                     y,
                     mean_out,
                     variance_out,
                     /*saved_mean=*/nullptr,
                     /*saved_variance=*/nullptr,
                     /*reserve_space=*/nullptr,
                     config);
}

676 677 678
void BilinearTensorProductInferMeta(const MetaTensor& x,
                                    const MetaTensor& y,
                                    const MetaTensor& weight,
679
                                    const MetaTensor& bias,
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
                                    MetaTensor* out,
                                    MetaConfig config) {
  auto x_dims = x.dims();
  auto y_dims = y.dims();
  auto weight_dims = weight.dims();

  PADDLE_ENFORCE_EQ(
      x_dims.size(),
      2UL,
      errors::InvalidArgument("The input(X) must be a 2D Tensor."));
  PADDLE_ENFORCE_EQ(
      y_dims.size(),
      2UL,
      errors::InvalidArgument("The input(Y) must be a 2D Tensor."));
  PADDLE_ENFORCE_EQ(
      weight_dims.size(),
      3UL,
      errors::InvalidArgument(
          "Expected the input(Weight) is a 3D tensor. But received %dD tensor.",
          weight_dims.size()));
  if (config.is_runtime || (x_dims[0] > 0 && y_dims[0] > 0)) {
    PADDLE_ENFORCE_EQ(x_dims[0],
                      y_dims[0],
                      errors::InvalidArgument(
                          "The first dimension(batch_size) of input(X) must be "
                          "equal to the first dimension of the input(Y)."));
  }
  PADDLE_ENFORCE_EQ(x_dims[1],
                    weight_dims[1],
                    errors::InvalidArgument(
                        "The second dimension of input(X) must be equal to "
                        "the second dimension of the input(Weight)."));
  PADDLE_ENFORCE_EQ(y_dims[1],
                    weight_dims[2],
                    errors::InvalidArgument(
                        "The second dimension of input(Y) must be equal to "
                        "the third dimension of the input(Weight)."));

718 719
  if (bias) {
    auto bias_dims = bias.dims();
720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741
    PADDLE_ENFORCE_EQ(bias_dims.size(),
                      2UL,
                      errors::InvalidArgument(
                          "The Input(Bias) must be a 2-D tensor with "
                          "the 2nd dimension fixed to 1 (a row vector)."));
    PADDLE_ENFORCE_EQ(bias_dims[0],
                      1UL,
                      errors::InvalidArgument(
                          "The Input(Bias) must be a 2-D tensor with "
                          "the 2nd dimension fixed to 1 (a row vector)."));
    PADDLE_ENFORCE_EQ(bias_dims[1],
                      weight_dims[0],
                      errors::InvalidArgument(
                          "The second dimension of input(Bias) must be equal "
                          "to the first dimension of the input(Weight)."));
  }

  out->set_dims({x_dims[0], weight_dims[0]});
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

742
void BroadcastTensorsInferMeta(const std::vector<const MetaTensor*>& x,
743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795
                               std::vector<MetaTensor*> out) {
  int target_rank = 0;
  const auto& input_dims = GetMetaTensorsDim(x);

  // 1. Find Output rank = max(Inputs rank)
  for (const auto& input_ddim : input_dims) {
    target_rank = std::max(target_rank, input_ddim.size());
  }

  PADDLE_ENFORCE_GT(target_rank,
                    0,
                    errors::InvalidArgument("BroadcastTensorsOp requires at "
                                            "least one input tensor to have "
                                            "rank greater than zero"));

  std::vector<int64_t> target_dims(target_rank, 0);
  // 2. Output dim(axis=x) = max(Inputs dim(axis=x))
  for (int index = 0; index < target_rank; index++) {
    // Loop axes in reverse order,
    // For each axis, take the maximum as target size
    // Fill size = 1 if shape vector exhausts
    int target_dim_size = 1;
    for (const auto& input_ddim : input_dims) {
      // Reversed order
      int axis = static_cast<int>(input_ddim.size()) - index - 1;
      int dim_size = 1;
      if (axis >= 0) {
        dim_size = input_ddim[axis];
      }

      if (target_dim_size != 1 && dim_size != 1 &&
          target_dim_size != dim_size) {
        PADDLE_THROW(errors::InvalidArgument(
            "BroadcastTensorsOp inputs does not satisfy bcast semantics, "
            "please check axis = %d in reverse order",
            index));
      }

      // We performed bcast semantics check at python level
      // So input tensors should all have legal shape
      target_dim_size = std::max(target_dim_size, dim_size);
    }
    target_dims[target_rank - index - 1] = target_dim_size;
  }

  // 3. Set Output Dim
  for (size_t i = 0; i < out.size(); i++) {
    out[i]->set_dims(phi::make_ddim(target_dims));
    out[i]->share_lod(*(x[i]));
    out[i]->set_dtype(x[i]->dtype());
  }
}

796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816
void CheckFiniteAndUnscaleInferMeta(const std::vector<const MetaTensor*>& xs,
                                    const MetaTensor& scale,
                                    std::vector<MetaTensor*> outs,
                                    MetaTensor* found_infinite) {
  PADDLE_ENFORCE_EQ(
      xs.size(),
      outs.size(),
      phi::errors::InvalidArgument(
          "The input(X) and output(Out) should have same size in "
          "Operator(check_finite_and_unscale), size of input(X) is %d "
          "and size of output(Out) is %d.",
          xs.size(),
          outs.size()));
  for (size_t i = 0; i < xs.size(); ++i) {
    outs[i]->set_dims(xs[i]->dims());
    outs[i]->set_dtype(xs[i]->dtype());
  }
  found_infinite->set_dims({1});
  found_infinite->set_dtype(DataType::BOOL);
}

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
void CoalesceTensorInferMeta(const std::vector<const MetaTensor*>& input,
                             DataType dtype,
                             bool copy_data,
                             bool set_constant,
                             bool persist_output,
                             float constant,
                             bool use_align,
                             int align_size,
                             int size_of_dtype,
                             const std::vector<int64_t>& concated_shapes,
                             const std::vector<int64_t>& concated_ranks,
                             std::vector<MetaTensor*> output,
                             MetaTensor* fused_output,
                             MetaConfig config) {
  if (config.is_runtime) {
    return;
  }
  if (size_of_dtype == -1) {
    size_of_dtype = paddle::experimental::SizeOf(dtype);
  }

  auto alignment = [](size_t size, size_t align_size) {
    size_t remaining = size % align_size;
    auto aligned_size = remaining == 0 ? size : size + (align_size - remaining);
    VLOG(4) << remaining << " " << size << " " << align_size << " "
            << aligned_size;
    return aligned_size;
  };
  VLOG(4) << "align_size: " << align_size;
  if (use_align && align_size > 0) {
    int64_t numel = 0;

    for (size_t i = 0; i < input.size(); ++i) {
      const auto& dim = input[i]->dims();
      auto size = phi::product(dim);
      auto len = use_align
                     ? alignment(static_cast<size_t>(size) * size_of_dtype,
                                 align_size) /
                           size_of_dtype
                     : static_cast<size_t>(size);
      numel += len;
    }
    if (fused_output) {
      fused_output->set_dims(phi::make_ddim({numel}));
      fused_output->set_dtype(dtype);
      VLOG(4) << "fused_output size:" << phi::make_ddim({numel});
    }
  }
}

867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884
void CheckMemoryContinueInferMeta(const std::vector<const MetaTensor*>& input,
                                  MetaTensor* output,
                                  std::vector<MetaTensor*> xout,
                                  MetaConfig config) {
  if (config.is_runtime) {
    return;
  }
  int64_t numel = 0;
  for (size_t i = 0; i < input.size(); ++i) {
    const auto& dim = input[i]->dims();
    auto size = phi::product(dim);
    auto len = size * paddle::experimental::SizeOf(input[i]->dtype());
    numel += len;
  }
  output->set_dims(phi::make_ddim({numel}));
  output->set_dtype(phi::DataType::INT8);
}

885
void ConcatInferMeta(const std::vector<const MetaTensor*>& x,
886 887 888 889 890
                     const Scalar& axis_scalar,
                     MetaTensor* out,
                     MetaConfig config) {
  PADDLE_ENFORCE_GE(x.size(),
                    0UL,
891
                    phi::errors::InvalidArgument(
892 893
                        "The size of input meta vector should be greater"
                        "than 0."));
894 895 896 897 898 899 900 901 902
  if (axis_scalar.FromTensor()) {
    auto out_dims =
        phi::make_ddim(std::vector<int>(x.at(0)->dims().size(), -1));
    out->set_dims(out_dims);
    out->set_dtype(x.at(0)->dtype());
    out->set_layout(x.at(0)->layout());
    out->share_lod(*x.at(0));
    return;
  }
903 904 905

  int axis = axis_scalar.to<int>();
  // 1. calculate axis
906
  int rank = x.at(0)->dims().size();
907 908 909
  PADDLE_ENFORCE_EQ(
      axis >= -rank && axis < rank,
      true,
910
      phi::errors::InvalidArgument(
911 912 913 914 915 916 917 918 919
          "The axis is expected to be in range of [%d, %d), but got %d",
          -rank,
          rank,
          axis));
  if (axis < 0) {
    axis = axis + rank;
  }

  // 2. calculate out dims
920
  std::vector<phi::DDim> x_dims;
921 922 923
  x_dims.reserve(x.size());
  for (const auto* x_t : x) {
    x_dims.emplace_back(x_t->dims());
924
  }
925 926
  phi::DDim out_dim =
      phi::funcs::ComputeAndCheckShape(config.is_runtime, x_dims, axis);
927

928
  out->set_dims(out_dim);
929 930 931
  out->set_dtype(x.at(0)->dtype());
  out->set_layout(x.at(0)->layout());
  out->share_lod(*x.at(0));
932 933
}

934 935 936 937 938 939 940 941 942
inline int ConvOutputSize(
    int input_size, int filter_size, int dilation, int padding, int stride) {
  const int dkernel = dilation * (filter_size - 1) + 1;
  int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
  PADDLE_ENFORCE_GT(
      output_size,
      0,
      phi::errors::InvalidArgument(
          "The output's size is expected to be greater than 0. "
943
          "But received: output's size is %d. The output's size is computed by "
944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959
          "((input_size + 2 * padding - (dilation * (filter_size - 1) + 1)) / "
          "stride + 1), where input_size is %d, padding is %d, "
          "filter_size is %d, dilation is %d, stride is %d.",
          output_size,
          input_size,
          padding,
          filter_size,
          dilation,
          stride));

  return output_size;
}

void DeformableConvInferMeta(const MetaTensor& x,
                             const MetaTensor& offset,
                             const MetaTensor& filter,
960
                             const MetaTensor& mask,
961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105
                             const std::vector<int>& strides,
                             const std::vector<int>& paddings,
                             const std::vector<int>& dilations,
                             int deformable_groups,
                             int groups,
                             int im2col_step,
                             MetaTensor* out,
                             MetaConfig config) {
  auto in_dims = x.dims();
  auto offset_dims = offset.dims();
  auto filter_dims = filter.dims();

  PADDLE_ENFORCE_EQ(
      in_dims.size(),
      4,
      phi::errors::InvalidArgument("Conv input should be 4-D tensor, get %u",
                                   in_dims.size()));
  PADDLE_ENFORCE_EQ(in_dims.size(),
                    filter_dims.size(),
                    phi::errors::InvalidArgument(
                        "Conv input dimension and filter dimension should be "
                        "the same. The difference is [%d]: [%d]",
                        in_dims.size(),
                        filter_dims.size()));
  PADDLE_ENFORCE_EQ(in_dims.size() - strides.size(),
                    2U,
                    phi::errors::InvalidArgument(
                        "Conv input dimension and strides "
                        "dimension should be consistent. But received input "
                        "dimension:[%d], strides dimension:[%d]",
                        in_dims.size(),
                        strides.size()));
  PADDLE_ENFORCE_EQ(paddings.size(),
                    strides.size(),
                    phi::errors::InvalidArgument(
                        "Conv paddings dimension and Conv strides dimension "
                        "should be the same. The difference is [%d]: [%d]",
                        paddings.size(),
                        strides.size()));

  PADDLE_ENFORCE_EQ(
      in_dims[1],
      filter_dims[1] * groups,
      phi::errors::InvalidArgument(
          "The number of input channels should be equal to filter "
          "channels * groups. The difference is [%d]: [%d]",
          in_dims[1],
          filter_dims[1] * groups));
  PADDLE_ENFORCE_EQ(
      filter_dims[0] % groups,
      0,
      phi::errors::InvalidArgument(
          "The number of output channels should be divided by groups. But "
          "received output channels:[%d], groups:[%d]",
          filter_dims[0],
          groups));
  PADDLE_ENFORCE_EQ(
      filter_dims[0] % deformable_groups,
      0,
      phi::errors::InvalidArgument(
          "The number of output channels should be "
          "divided by deformable groups. The difference is [%d]: [%d]",
          filter_dims[0] % groups,
          0));

  if (in_dims[0] > im2col_step) {
    PADDLE_ENFORCE_EQ(
        in_dims[0] % im2col_step,
        0U,
        phi::errors::InvalidArgument(
            "Input batchsize must be smaller than or divide im2col_step. But "
            "received Input batchsize:[%d], im2col_step:[%d]",
            in_dims[0],
            im2col_step));
  }

  for (size_t i = 0; i < strides.size(); ++i) {
    PADDLE_ENFORCE_GT(
        strides[i],
        0U,
        phi::errors::InvalidArgument("stride %d size incorrect", i));
  }
  for (size_t i = 0; i < dilations.size(); ++i) {
    PADDLE_ENFORCE_GT(
        dilations[i],
        0U,
        phi::errors::InvalidArgument("dilation %d size incorrect", i));
  }

  std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
  for (size_t i = 0; i < strides.size(); ++i) {
    if (!config.is_runtime &&
        (in_dims[i + 2] <= 0 || filter_dims[i + 2] <= 0)) {
      output_shape.push_back(-1);
    } else {
      output_shape.push_back(ConvOutputSize(in_dims[i + 2],
                                            filter_dims[i + 2],
                                            dilations[i],
                                            paddings[i],
                                            strides[i]));
    }
  }

  PADDLE_ENFORCE_EQ(
      output_shape[1] % deformable_groups,
      0U,
      phi::errors::InvalidArgument(
          "output num_filter must divide deformable group size. But received "
          "output num_filter:[%d], deformable group size:[%d]",
          output_shape[1],
          deformable_groups));

  if (config.is_runtime) {
    PADDLE_ENFORCE_EQ(output_shape[2],
                      offset_dims[2],
                      phi::errors::InvalidArgument(
                          "output height must equal to offset map height. "
                          "The difference is [%d]: [%d]",
                          output_shape[2],
                          offset_dims[2]));
    PADDLE_ENFORCE_EQ(output_shape[3],
                      offset_dims[3],
                      phi::errors::InvalidArgument(
                          "output width must equal to offset map width. The "
                          "difference is [%d]: [%d]",
                          output_shape[3],
                          offset_dims[3]));

    PADDLE_ENFORCE_EQ(offset_dims[1] % (filter_dims[2] * filter_dims[3]),
                      0U,
                      phi::errors::InvalidArgument(
                          "offset filter must divide deformable group size. "
                          "But received [%d]: [%d]",
                          offset_dims[1],
                          filter_dims[2] * filter_dims[3]));
    PADDLE_ENFORCE_EQ(
        offset_dims[1] / (2 * filter_dims[2] * filter_dims[3]),
        deformable_groups,
        phi::errors::InvalidArgument(
            "offset filter must divide deformable group size. But received "
            "[%d]: [%d]",
            offset_dims[1] / (2 * filter_dims[2] * filter_dims[3]),
            deformable_groups));

    if (mask) {
1106
      auto mask_dims = mask.dims();
1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142
      PADDLE_ENFORCE_EQ(output_shape[2],
                        mask_dims[2],
                        phi::errors::InvalidArgument(
                            "output height must equal to mask map height. The "
                            "difference is [%d] vs [%d]",
                            output_shape[2],
                            mask_dims[2]));
      PADDLE_ENFORCE_EQ(output_shape[3],
                        mask_dims[3],
                        phi::errors::InvalidArgument(
                            "output width must equal to mask map width. The "
                            "difference is [%d] vs [%d]",
                            output_shape[3],
                            mask_dims[3]));

      PADDLE_ENFORCE_EQ(mask_dims[1] % (filter_dims[2] * filter_dims[3]),
                        0U,
                        phi::errors::InvalidArgument(
                            "mask filter must divide deformable group size. "
                            "But received [%d]: [%d]",
                            mask_dims[1],
                            filter_dims[2] * filter_dims[3]));
      PADDLE_ENFORCE_EQ(mask_dims[1] / (filter_dims[2] * filter_dims[3]),
                        deformable_groups,
                        phi::errors::InvalidArgument(
                            "mask filter must divide deformable group size. "
                            "But received [%d]: [%d]",
                            mask_dims[1] / (filter_dims[2] * filter_dims[3]),
                            deformable_groups));
    }
  }

  out->set_dims(phi::make_ddim(output_shape));
  out->set_dtype(x.dtype());
}

Z
zhiboniu 已提交
1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211
void EditDistanceInferMeta(const MetaTensor& hyps,
                           const MetaTensor& refs,
                           const MetaTensor& hypslength,
                           const MetaTensor& refslength,
                           bool normalized,
                           MetaTensor* sequencenum,
                           MetaTensor* out) {
  auto hyp_dims = hyps.dims();
  auto ref_dims = refs.dims();

  if (hypslength && refslength) {
    auto hyp_length_dims = hypslength.dims();
    auto ref_length_dims = refslength.dims();

    PADDLE_ENFORCE_EQ(
        hyp_dims.size() == 2 && ref_dims.size() == 2 &&
            hyp_dims[0] == ref_dims[0],
        true,
        errors::InvalidArgument(
            "Input(hyps) and Input(refs) must be 2-D Tensors with "
            "identical first dimension. But received Input(Hyps): "
            "input rank %u, input shape [%s]; received Input(Refs): "
            "input rank %u, input shape [%s]",
            hyp_dims.size(),
            hyp_dims,
            ref_dims.size(),
            ref_dims));
    PADDLE_ENFORCE_EQ(
        hyp_length_dims[0] == ref_length_dims[0] &&
            hyp_length_dims[0] == hyp_dims[0],
        true,
        errors::InvalidArgument(
            "Input(hypslength), Input(refslength) and Input(hyps) "
            "should have identical first dimension. But received "
            "Input(hypslength): input rank %u, input shape [%s]; "
            "received Input(refslength): input rank %u, input shape "
            "[%s]; received Input(hyps): input rank %u, input shape "
            "[%s].",
            hyp_length_dims.size(),
            hyp_length_dims,
            ref_length_dims.size(),
            ref_length_dims,
            hyp_dims.size(),
            hyp_dims));
  } else {
    PADDLE_ENFORCE_EQ(
        hyp_dims.size() == 2 && hyp_dims[1] == 1,
        true,
        errors::InvalidArgument(
            "Input(Hyps) must be a 2-D LoDTensor with the 2nd dimension "
            "equal to 1. But received: input rank %u, input shape [%s].",
            hyp_dims.size(),
            hyp_dims));
    PADDLE_ENFORCE_EQ(
        ref_dims.size() == 2 && ref_dims[1] == 1,
        true,
        errors::InvalidArgument(
            "Input(Refs) must be a 2-D LoDTensor with the 2nd dimension "
            "equal to 1. But received: input rank %u, input shape [%s].",
            ref_dims.size(),
            ref_dims));
  }

  out->set_dims(refs.dims());
  out->set_dtype(DataType::FLOAT32);
  sequencenum->set_dims(phi::make_ddim({1}));
  sequencenum->set_dtype(DataType::FLOAT32);
}

Z
zhiboniu 已提交
1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229
void GenerateProposalsV2InferMeta(const MetaTensor& scores,
                                  const MetaTensor& bbox_deltas,
                                  const MetaTensor& im_shape,
                                  const MetaTensor& anchors,
                                  const MetaTensor& variances,
                                  int pre_nms_top_n,
                                  int post_nms_top_n,
                                  float nms_thresh,
                                  float min_size,
                                  float eta,
                                  bool pixel_offset,
                                  MetaTensor* rpn_rois,
                                  MetaTensor* rpn_roi_probs,
                                  MetaTensor* rpn_rois_num) {
  rpn_rois->set_dims(phi::make_ddim({-1, 4}));
  rpn_roi_probs->set_dims(phi::make_ddim({-1, 1}));
}

1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325
void GraphReindexInferMeta(const MetaTensor& x,
                           const MetaTensor& neighbors,
                           const MetaTensor& count,
                           const MetaTensor& hashtable_value,
                           const MetaTensor& hashtable_index,
                           bool flag_buffer_hashtable,
                           MetaTensor* reindex_src,
                           MetaTensor* reindex_dst,
                           MetaTensor* out_nodes) {
  auto GraphReindexShapeCheck = [](const phi::DDim& dims,
                                   std::string tensor_name) {
    if (dims.size() == 2) {
      PADDLE_ENFORCE_EQ(
          dims[1],
          1,
          phi::errors::InvalidArgument("The last dim of %s should be 1 when it "
                                       "is 2D, but we get %d",
                                       tensor_name,
                                       dims[1]));
    } else {
      PADDLE_ENFORCE_EQ(
          dims.size(),
          1,
          phi::errors::InvalidArgument(
              "The %s should be 1D, when it is not 2D, but we get %d",
              tensor_name,
              dims.size()));
    }
  };

  GraphReindexShapeCheck(x.dims(), "X");
  GraphReindexShapeCheck(neighbors.dims(), "Neighbors");
  GraphReindexShapeCheck(count.dims(), "Count");
  if (flag_buffer_hashtable) {
    GraphReindexShapeCheck(hashtable_value.dims(), "HashTable_Value");
    GraphReindexShapeCheck(hashtable_index.dims(), "HashTable_Index");
  }

  reindex_src->set_dims({-1});
  reindex_src->set_dtype(neighbors.dtype());
  reindex_dst->set_dims({-1});
  reindex_dst->set_dtype(neighbors.dtype());
  out_nodes->set_dims({-1});
  out_nodes->set_dtype(x.dtype());
}

void GraphSampleNeighborsInferMeta(const MetaTensor& row,
                                   const MetaTensor& col_ptr,
                                   const MetaTensor& x,
                                   const MetaTensor& eids,
                                   const MetaTensor& perm_buffer,
                                   int sample_size,
                                   bool return_eids,
                                   bool flag_perm_buffer,
                                   MetaTensor* out,
                                   MetaTensor* out_count,
                                   MetaTensor* out_eids) {
  // GSN: GraphSampleNeighbors
  auto GSNShapeCheck = [](const phi::DDim& dims, std::string tensor_name) {
    if (dims.size() == 2) {
      PADDLE_ENFORCE_EQ(
          dims[1],
          1,
          phi::errors::InvalidArgument("The last dim of %s should be 1 when it "
                                       "is 2D, but we get %d",
                                       tensor_name,
                                       dims[1]));
    } else {
      PADDLE_ENFORCE_EQ(
          dims.size(),
          1,
          phi::errors::InvalidArgument(
              "The %s should be 1D, when it is not 2D, but we get %d",
              tensor_name,
              dims.size()));
    }
  };

  GSNShapeCheck(row.dims(), "Row");
  GSNShapeCheck(col_ptr.dims(), "Col_Ptr");
  GSNShapeCheck(x.dims(), "X");
  if (return_eids) {
    GSNShapeCheck(eids.dims(), "Eids");
    out_eids->set_dims({-1});
    out_eids->set_dtype(row.dtype());
  }
  if (flag_perm_buffer) {
    GSNShapeCheck(perm_buffer.dims(), "Perm_Buffer");
  }

  out->set_dims({-1});
  out->set_dtype(row.dtype());
  out_count->set_dims({-1});
  out_count->set_dtype(DataType::INT32);
}

1326 1327 1328
void HierarchicalSigmoidInferMeta(const MetaTensor& x,
                                  const MetaTensor& w,
                                  const MetaTensor& label,
1329 1330 1331
                                  const MetaTensor& path,
                                  const MetaTensor& code,
                                  const MetaTensor& bias,
1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359
                                  int num_classes,
                                  bool remote_prefetch,
                                  int trainer_id,
                                  const std::vector<int64_t>& height_sections,
                                  const std::vector<std::string>& epmap,
                                  const std::vector<std::string>& table_names,
                                  bool is_sparse,
                                  MetaTensor* out,
                                  MetaTensor* pre_out,
                                  MetaTensor* w_out) {
  const int64_t input_dims = x.dims()[0];
  const int64_t label_dims = label.dims()[0];
  PADDLE_ENFORCE_EQ(input_dims,
                    label_dims,
                    phi::errors::InvalidArgument(
                        "The first dimension of "
                        "input and label is expected to be the same. "
                        "But received input's first dimension is %d; "
                        "label's first dimension is %d.",
                        input_dims,
                        label_dims));

  std::vector<int64_t> output_shape({input_dims, 1});
  out->set_dims(phi::make_ddim(output_shape));
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

1360 1361
static void Interpolate1DInferShapeCheck(
    const MetaTensor& x,
1362 1363 1364
    const MetaTensor& out_size,
    const paddle::optional<std::vector<const MetaTensor*>>& size_tensor,
    const MetaTensor& scale_tensor,
1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418
    const std::string& data_layout_str,
    int out_d,
    int out_h,
    int out_w,
    const std::vector<float>& scale,
    const std::string& interp_method,
    bool align_corners,
    int align_mode,
    MetaTensor* output,
    MetaConfig config) {
  auto dim_x = x.dims();

  PADDLE_ENFORCE_EQ("linear",
                    interp_method,
                    phi::errors::InvalidArgument(
                        "Interpolation method can only be \"linear\" when"
                        "Input(X) dimension is 3, but got method = %s .",
                        interp_method));
  const DataLayout data_layout =
      paddle::framework::StringToDataLayout(data_layout_str);
  for (int i = 0; i < dim_x.size(); ++i) {
    PADDLE_ENFORCE_NE(
        dim_x[i],
        0,
        phi::errors::InvalidArgument("The shape of input(x) should be larged "
                                     "than 0, bug received shape[%d] is %d ",
                                     i,
                                     dim_x[i]));
  }
  if (size_tensor && size_tensor->size() > 0) {
    // top prority size
    auto inputs_name = size_tensor.get();
    PADDLE_ENFORCE_EQ(
        inputs_name.size(),
        1,
        phi::errors::InvalidArgument(
            "Input(SizeTensor)'size of Op(interpolate) must be 1. "
            "Attr(out_shape)'s length must be 1 for 3-D input tensor, but got "
            "size = %d .",
            inputs_name.size()));
    phi::DDim dim_out;
    if (data_layout == DataLayout::kNCHW) {
      dim_out = {dim_x[0], dim_x[1], out_w};
    } else {
      dim_out = {dim_x[0], out_w, dim_x[2]};
    }
    output->set_dims(dim_out);
    output->set_dtype(x.dtype());

    return;
  }

  int out_w_tmp;
  if (scale_tensor) {
1419
    auto scale_tensor_dim = scale_tensor.dims();
1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456
    PADDLE_ENFORCE_EQ(
        scale_tensor_dim.size(),
        1,
        phi::errors::InvalidArgument(
            "Scale's dimension size must be 1, but got dimension = %d .",
            scale_tensor_dim.size()));
    PADDLE_ENFORCE_EQ(scale_tensor_dim[0],
                      1,
                      phi::errors::InvalidArgument(
                          "Scale's shape must be 1, but got shape = %d .",
                          scale_tensor_dim[0]));
    out_w_tmp = -1;
  } else {
    if (scale.size() > 0) {
      float scale_w = -1;
      scale_w = scale[0];
      PADDLE_ENFORCE_EQ(
          scale_w > 0,
          true,
          phi::errors::InvalidArgument(
              "The scale_w in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
      if (scale_w > 0.) {
        // round down
        out_w_tmp = (data_layout == DataLayout::kNCHW
                         ? static_cast<int>(dim_x[2] * scale_w)
                         : static_cast<int>(dim_x[1] * scale_w));
        // protect when input shape is -1
        out_w_tmp = out_w_tmp > 0 ? out_w_tmp : -1;
      }
    } else {
      out_w_tmp = out_w;
    }
  }

  if (out_size && config.is_runtime) {
1457
    auto out_size_dim = out_size.dims();
1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488
    PADDLE_ENFORCE_EQ(
        out_size_dim.size(),
        1,
        phi::errors::InvalidArgument(
            "OutSize's dimension size must be 1, but got dimention = %d .",
            out_size_dim.size()));
    PADDLE_ENFORCE_EQ(
        out_size_dim[0],
        1,
        phi::errors::InvalidArgument(
            "OutSize's 0-th dimension's value must be 1, but got value = %d .",
            out_size_dim[0]));

    // dims will be seted in kernel
    output->set_dtype(x.dtype());
    output->share_lod(x);
    return;
  }

  phi::DDim dim_out;
  if (data_layout == DataLayout::kNCHW) {
    dim_out = {dim_x[0], dim_x[1], out_w_tmp};
  } else {
    dim_out = {dim_x[0], out_w_tmp, dim_x[2]};
  }
  output->set_dims(dim_out);
  output->set_dtype(x.dtype());
}

static void Interpolate2DInferShapeCheck(
    const MetaTensor& x,
1489 1490 1491
    const MetaTensor& out_size,
    const paddle::optional<std::vector<const MetaTensor*>>& size_tensor,
    const MetaTensor& scale_tensor,
1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548
    const std::string& data_layout_str,
    int out_d,
    int out_h,
    int out_w,
    const std::vector<float>& scale,
    const std::string& interp_method,
    bool align_corners,
    int align_mode,
    MetaTensor* output,
    MetaConfig config) {
  auto dim_x = x.dims();

  PADDLE_ENFORCE(
      "bilinear" == interp_method || "nearest" == interp_method ||
          "bicubic" == interp_method,
      phi::errors::InvalidArgument(
          "Interpolation method can only be \"bilinear\" or \"nearest\" when "
          "Input(X) dimension is 4, but got method = %s.",
          interp_method));
  const DataLayout data_layout =
      paddle::framework::StringToDataLayout(data_layout_str);

  for (int i = 0; i < dim_x.size(); ++i) {
    PADDLE_ENFORCE_NE(
        dim_x[i],
        0,
        phi::errors::InvalidArgument("The shape of input(x) should be larged "
                                     "than 0, bug received shape[%d] is %d ",
                                     i,
                                     dim_x[i]));
  }

  if (size_tensor && size_tensor->size()) {
    // top prority size
    auto inputs_name = size_tensor.get();
    PADDLE_ENFORCE_EQ(
        inputs_name.size(),
        2,
        phi::errors::InvalidArgument(
            "Input(SizeTensor)'size of Op(interpolate) must be 2. "
            "Attr(out_shape)'s length must be 2 for 4-D input "
            "tensor, but got size = %d .",
            inputs_name.size()));
    phi::DDim dim_out;
    if (data_layout == DataLayout::kNCHW) {
      dim_out = {dim_x[0], dim_x[1], out_h, out_w};
    } else {
      dim_out = {dim_x[0], out_h, out_w, dim_x[3]};
    }
    output->set_dims(dim_out);
    output->set_dtype(x.dtype());

    return;
  }

  int out_h_tmp, out_w_tmp;
  if (scale_tensor) {
1549
    auto scale_tensor_dim = scale_tensor.dims();
1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601
    PADDLE_ENFORCE_EQ(
        scale_tensor_dim.size(),
        1,
        phi::errors::InvalidArgument(
            "Scale's dimension size must be 1, but got dimension = %d .",
            scale_tensor_dim.size()));
    PADDLE_ENFORCE_EQ(scale_tensor_dim[0] == 2 || scale_tensor_dim[0] == 1,
                      true,
                      phi::errors::InvalidArgument(
                          "Scale's shape must be 2 or 1, but got shape = %d .",
                          scale_tensor_dim[0]));
    out_h_tmp = -1;
    out_w_tmp = -1;
  } else {
    if (scale.size() > 0) {
      float scale_h = -1;
      float scale_w = -1;
      scale_h = scale[0];
      scale_w = scale[1];
      PADDLE_ENFORCE_EQ(
          scale_w > 0,
          true,
          phi::errors::InvalidArgument(
              "The scale_w in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
      PADDLE_ENFORCE_EQ(
          scale_h > 0,
          true,
          phi::errors::InvalidArgument(
              "The scale_h in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_h));
      if (scale_h > 0. && scale_w > 0.) {
        // round down
        out_h_tmp = (data_layout == DataLayout::kNCHW
                         ? static_cast<int>(dim_x[2] * scale_h)
                         : static_cast<int>(dim_x[1] * scale_h));
        out_w_tmp = (data_layout == DataLayout::kNCHW
                         ? static_cast<int>(dim_x[3] * scale_w)
                         : static_cast<int>(dim_x[2] * scale_w));
        // protect when input shape is -1
        out_h_tmp = out_h_tmp > 0 ? out_h_tmp : -1;
        out_w_tmp = out_w_tmp > 0 ? out_w_tmp : -1;
      }
    } else {
      out_h_tmp = out_h;
      out_w_tmp = out_w;
    }
  }

  if (out_size && config.is_runtime) {
1602
    auto out_size_dim = out_size.dims();
1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633
    PADDLE_ENFORCE_EQ(
        out_size_dim.size(),
        1,
        phi::errors::InvalidArgument(
            "OutSize's dimension size must be 1, but got dimension = %d .",
            out_size_dim.size()));
    PADDLE_ENFORCE_EQ(
        out_size_dim[0],
        2,
        phi::errors::InvalidArgument(
            "OutSize's dim[0] must be 2, but got dimention = %d .",
            out_size_dim[0]));
    // dims will be seted in kernel
    output->set_dtype(x.dtype());
    output->share_lod(x);
    return;
  }

  phi::DDim dim_out;
  if (data_layout == DataLayout::kNCHW) {
    dim_out = {dim_x[0], dim_x[1], out_h_tmp, out_w_tmp};
  } else {
    dim_out = {dim_x[0], out_h_tmp, out_w_tmp, dim_x[3]};
  }

  output->set_dims(dim_out);
  output->set_dtype(x.dtype());
}

static void Interpolate3DInferShapeCheck(
    const MetaTensor& x,
1634 1635 1636
    const MetaTensor& out_size,
    const paddle::optional<std::vector<const MetaTensor*>>& size_tensor,
    const MetaTensor& scale_tensor,
1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691
    const std::string& data_layout_str,
    int out_d,
    int out_h,
    int out_w,
    const std::vector<float>& scale,
    const std::string& interp_method,
    bool align_corners,
    int align_mode,
    MetaTensor* output,
    MetaConfig config) {
  auto dim_x = x.dims();

  PADDLE_ENFORCE("nearest" == interp_method || "trilinear" == interp_method,
                 phi::errors::InvalidArgument(
                     "Interpolation method can only be \"trilinear\" or "
                     "\"nearest\" when Input(X) "
                     "dimension is 5, but got method = %s .",
                     interp_method));
  const DataLayout data_layout =
      paddle::framework::StringToDataLayout(data_layout_str);

  for (int i = 0; i < dim_x.size(); ++i) {
    PADDLE_ENFORCE_NE(
        dim_x[i],
        0,
        phi::errors::InvalidArgument("The shape of input(x) should be larged "
                                     "than 0, bug received shape[%d] is %d ",
                                     i,
                                     dim_x[i]));
  }

  if (size_tensor && size_tensor->size() > 0) {
    // top prority size
    auto inputs_name = size_tensor.get();
    PADDLE_ENFORCE_EQ(
        inputs_name.size(),
        3,
        phi::errors::InvalidArgument(
            "Input(SizeTensor)'s size of Op(interpolate) must be 3. "
            "Attr(out_shape)'s length must be 3 for 5-D input "
            "tensor, but got size = %d .",
            inputs_name.size()));
    phi::DDim dim_out;
    if (data_layout == DataLayout::kNCHW) {
      dim_out = {dim_x[0], dim_x[1], out_d, out_h, out_w};
    } else {
      dim_out = {dim_x[0], out_d, out_h, out_w, dim_x[4]};
    }
    output->set_dims(dim_out);
    output->set_dtype(x.dtype());
    return;
  }

  int out_d_tmp, out_h_tmp, out_w_tmp;
  if (scale_tensor) {
1692
    auto scale_tensor_dim = scale_tensor.dims();
1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759
    PADDLE_ENFORCE_EQ(
        scale_tensor_dim.size(),
        1,
        phi::errors::InvalidArgument(
            "Scale's dimension size must be 1, but got size = %d .",
            scale_tensor_dim.size()));
    PADDLE_ENFORCE_EQ(scale_tensor_dim[0] == 3 || scale_tensor_dim[0] == 1,
                      true,
                      phi::errors::InvalidArgument(
                          "Scale's shape must be 3 or 1, but got shape = %d .",
                          scale_tensor_dim[0]));
    out_d_tmp = -1;
    out_h_tmp = -1;
    out_w_tmp = -1;
  } else {
    if (scale.size() > 0) {
      float scale_d = -1;
      float scale_h = -1;
      float scale_w = -1;
      scale_d = scale[0];
      scale_h = scale[1];
      scale_w = scale[2];
      PADDLE_ENFORCE_EQ(
          scale_w > 0,
          true,
          phi::errors::InvalidArgument(
              "The scale_w in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_w));
      PADDLE_ENFORCE_EQ(
          scale_h > 0,
          true,
          phi::errors::InvalidArgument(
              "The scale_h in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_h));
      PADDLE_ENFORCE_EQ(
          scale_d > 0,
          true,
          phi::errors::InvalidArgument(
              "The scale_d in Attr(scale) of Operator(interpolate) "
              "should be greater than 0, but received value is %d.",
              scale_d));
      if (scale_d > 0. && scale_h > 0. && scale_w > 0.) {
        // round down
        out_d_tmp = (data_layout == DataLayout::kNCHW
                         ? static_cast<int>(dim_x[2] * scale_d)
                         : static_cast<int>(dim_x[1] * scale_d));
        out_h_tmp = (data_layout == DataLayout::kNCHW
                         ? static_cast<int>(dim_x[3] * scale_h)
                         : static_cast<int>(dim_x[2] * scale_h));
        out_w_tmp = (data_layout == DataLayout::kNCHW
                         ? static_cast<int>(dim_x[4] * scale_w)
                         : static_cast<int>(dim_x[3] * scale_w));
        // protect when input shape is -1
        out_d_tmp = out_d_tmp > 0 ? out_d_tmp : -1;
        out_h_tmp = out_h_tmp > 0 ? out_h_tmp : -1;
        out_w_tmp = out_w_tmp > 0 ? out_w_tmp : -1;
      }
    } else {
      out_d_tmp = out_d;
      out_h_tmp = out_h;
      out_w_tmp = out_w;
    }
  }

  if (out_size && config.is_runtime) {
1760
    auto out_size_dim = out_size.dims();
1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789
    PADDLE_ENFORCE_EQ(
        out_size_dim.size(),
        1,
        phi::errors::InvalidArgument(
            "OutSize's dimension size must be 1, but got size is %d.",
            out_size_dim.size()));
    PADDLE_ENFORCE_EQ(out_size_dim[0],
                      3,
                      phi::errors::InvalidArgument(
                          "OutSize's dim[0] must be 3, but got size is %d.",
                          out_size_dim[0]));
    // dims will be seted in kernel
    output->set_dtype(x.dtype());
    output->share_lod(x);
    return;
  }

  phi::DDim dim_out;
  if (data_layout == DataLayout::kNCHW) {
    dim_out = {dim_x[0], dim_x[1], out_d_tmp, out_h_tmp, out_w_tmp};
  } else {
    dim_out = {dim_x[0], out_d_tmp, out_h_tmp, out_w_tmp, dim_x[4]};
  }
  output->set_dims(dim_out);
  output->set_dtype(x.dtype());
}

void InterpolateInferMeta(
    const MetaTensor& x,
1790 1791 1792
    const MetaTensor& out_size,
    const paddle::optional<std::vector<const MetaTensor*>>& size_tensor,
    const MetaTensor& scale_tensor,
1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859
    const std::string& data_layout_str,
    int out_d,
    int out_h,
    int out_w,
    const std::vector<float>& scale,
    const std::string& interp_method,
    bool align_corners,
    int align_mode,
    MetaTensor* output,
    MetaConfig config) {
  auto dim_x = x.dims();  // NCHW format
  PADDLE_ENFORCE(
      dim_x.size() == 3 || dim_x.size() == 4 || dim_x.size() == 5,
      phi::errors::Unimplemented(
          "Input(X) dimension must be 3, 4 or 5, but got dimension = %d .",
          dim_x.size()));
  if (dim_x.size() == 3) {
    // shape check for 1D interpolate for input tensor shape NCHW
    Interpolate1DInferShapeCheck(x,
                                 out_size,
                                 size_tensor,
                                 scale_tensor,
                                 data_layout_str,
                                 out_d,
                                 out_h,
                                 out_w,
                                 scale,
                                 interp_method,
                                 align_corners,
                                 align_mode,
                                 output,
                                 config);
  } else if (dim_x.size() == 4) {
    // shape check for 2D interpolate for input tensor shape NCHW
    Interpolate2DInferShapeCheck(x,
                                 out_size,
                                 size_tensor,
                                 scale_tensor,
                                 data_layout_str,
                                 out_d,
                                 out_h,
                                 out_w,
                                 scale,
                                 interp_method,
                                 align_corners,
                                 align_mode,
                                 output,
                                 config);
  } else {  // dim_x.size() == 5
    // shape check for 3D interpolate for input tensor shape NCDHW
    Interpolate3DInferShapeCheck(x,
                                 out_size,
                                 size_tensor,
                                 scale_tensor,
                                 data_layout_str,
                                 out_d,
                                 out_h,
                                 out_w,
                                 scale,
                                 interp_method,
                                 align_corners,
                                 align_mode,
                                 output,
                                 config);
  }
}

T
Thomas Young 已提交
1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958
void LambInferMeta(const MetaTensor& param,
                   const MetaTensor& grad,
                   const MetaTensor& learning_rate,
                   const MetaTensor& moment1,
                   const MetaTensor& moment2,
                   const MetaTensor& beta1_pow,
                   const MetaTensor& beta2_pow,
                   const MetaTensor& master_param,
                   const MetaTensor& skip_update,
                   float weight_decay,
                   float beta1,
                   float beta2,
                   float epsilon,
                   bool multi_precision,
                   MetaTensor* param_out,
                   MetaTensor* moment1_out,
                   MetaTensor* moment2_out,
                   MetaTensor* beta1_pow_out,
                   MetaTensor* beta2_pow_out,
                   MetaTensor* master_param_outs) {
  auto lr_dims = learning_rate.dims();
  PADDLE_ENFORCE_NE(
      phi::product(lr_dims),
      0,
      phi::errors::InvalidArgument(
          "The number of LearningRate shall not be 0, but received %d. Maybe "
          "the Input variable LearningRate has not "
          "been initialized. You may need to confirm "
          "if you put exe.run(startup_program) "
          "after optimizer.minimize function.",
          phi::product(lr_dims)));
  PADDLE_ENFORCE_EQ(
      phi::product(lr_dims),
      1,
      phi::errors::InvalidArgument(
          "Learning rate should have 1 dimension, but received %d.",
          phi::product(lr_dims)));
  auto beta1_pow_dims = beta1_pow.dims();
  PADDLE_ENFORCE_GE(phi::product(beta1_pow_dims),
                    1,
                    phi::errors::InvalidArgument(
                        "The size of Beta1 power accumulator should be "
                        "greater than 0, but received %d.",
                        phi::product(beta1_pow_dims)));
  auto beta2_pow_dims = beta2_pow.dims();
  PADDLE_ENFORCE_GE(phi::product(beta2_pow_dims),
                    1,
                    phi::errors::InvalidArgument(
                        "The size of Beta2 power accumulator should be "
                        "greater than 0, but received %d.",
                        phi::product(beta2_pow_dims)));

  auto param_dims = param.dims();
  PADDLE_ENFORCE_EQ(
      param_dims,
      moment1.dims(),
      phi::errors::InvalidArgument(
          "Param and Moment1 input of LambOp should have same dimension. But "
          "received Param dims: [%s], Moment1 dims: [%s].",
          param_dims,
          moment1.dims()));
  PADDLE_ENFORCE_EQ(
      param_dims,
      moment2.dims(),
      errors::InvalidArgument(
          "Param and Moment2 input of AdamOp should have same dimension. But "
          "received Param dims: [%s], Moment2 dims: [%s].",
          param_dims,
          moment2.dims()));

  PADDLE_ENFORCE_NOT_NULL(
      param_out, errors::NotFound("The output param_out can not be nullptr"));
  PADDLE_ENFORCE_NOT_NULL(
      moment1_out,
      errors::NotFound("The output moment1_out can not be nullptr"));
  PADDLE_ENFORCE_NOT_NULL(
      moment2_out,
      errors::NotFound("The output moment2_out can not be nullptr"));
  PADDLE_ENFORCE_NOT_NULL(
      beta1_pow_out,
      errors::NotFound("The output beta1_pow_out can not be nullptr"));
  PADDLE_ENFORCE_NOT_NULL(
      beta2_pow_out,
      errors::NotFound("The output beta2_pow_out can not be nullptr"));

  param_out->set_dims(param_dims);
  param_out->set_dtype(param.dtype());

  moment1_out->set_dims(param_dims);
  moment1_out->set_dtype(moment1.dtype());
  moment2_out->set_dims(param_dims);
  moment2_out->set_dtype(moment2.dtype());

  beta1_pow_out->set_dims(beta1_pow_dims);
  beta1_pow_out->set_dtype(beta1_pow.dtype());
  beta2_pow_out->set_dims(beta2_pow_dims);
  beta2_pow_out->set_dtype(beta2_pow.dtype());
}

1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
void LogspaceInferMeta(const MetaTensor& start,
                       const MetaTensor& stop,
                       const MetaTensor& number,
                       const MetaTensor& base,
                       MetaTensor* out) {
  auto s_dims = start.dims();
  PADDLE_ENFORCE_EQ(
      (s_dims.size() == 1) && (s_dims[0] == 1),
      true,
      phi::errors::InvalidArgument("The shape of Input(Start) must be [1],"
                                   "but received input shape is [%s].",
                                   s_dims));
  auto e_dims = stop.dims();
  PADDLE_ENFORCE_EQ(
      (e_dims.size() == 1) && (e_dims[0] == 1),
      true,
      phi::errors::InvalidArgument("The shape of Input(Stop) must be [1],"
                                   "but received input shape is [%s].",
                                   e_dims));
  auto num_dims = number.dims();
  PADDLE_ENFORCE_EQ(
      (num_dims.size() == 1) && (num_dims[0] == 1),
      true,
      phi::errors::InvalidArgument("The shape of Input(Num) must be [1],"
                                   "but received input shape is [%s].",
                                   num_dims));
  auto b_dims = base.dims();
  PADDLE_ENFORCE_EQ(
      (b_dims.size() == 1) && (b_dims[0] == 1),
      true,
      phi::errors::InvalidArgument("The shape of Input(Base) must be [1],"
                                   "but received input shape is [%s].",
                                   b_dims));
  out->set_dims(phi::make_ddim({-1}));
  out->set_dtype(start.dtype());
}

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
void MergedAdamInferMeta(
    const std::vector<const MetaTensor*>& param,
    const std::vector<const MetaTensor*>& grad,
    const std::vector<const MetaTensor*>& learning_rate,
    const std::vector<const MetaTensor*>& moment1,
    const std::vector<const MetaTensor*>& moment2,
    const std::vector<const MetaTensor*>& beta1_pow,
    const std::vector<const MetaTensor*>& beta2_pow,
    const paddle::optional<std::vector<const MetaTensor*>>& master_param,
    const Scalar& beta1,
    const Scalar& beta2,
    const Scalar& epsilon,
    bool multi_precision,
    bool use_global_beta_pow,
    std::vector<MetaTensor*> param_out,
    std::vector<MetaTensor*> moment1_out,
    std::vector<MetaTensor*> moment2_out,
    std::vector<MetaTensor*> beta1_pow_out,
    std::vector<MetaTensor*> beta2_pow_out,
    std::vector<MetaTensor*> master_param_out) {}

2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032
void MergedMomentumInferMeta(
    const std::vector<const MetaTensor*>& param,
    const std::vector<const MetaTensor*>& grad,
    const std::vector<const MetaTensor*>& velocity,
    const std::vector<const MetaTensor*>& learning_rate,
    const paddle::optional<std::vector<const MetaTensor*>>& master_param,
    float mu,
    bool use_nesterov,
    const std::vector<std::string>& regularization_method,
    const std::vector<float>& regularization_coeff,
    bool multi_precision,
    float rescale_grad,
    std::vector<MetaTensor*> param_out,
    std::vector<MetaTensor*> velocity_out,
    std::vector<MetaTensor*> master_param_out) {}

2033
void MeshgridInferMeta(const std::vector<const MetaTensor*>& inputs,
H
hong 已提交
2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048
                       std::vector<MetaTensor*> outputs) {
  const size_t inputs_num = inputs.size();

  auto out_shape = std::vector<int>(inputs_num);

  for (size_t i = 0; i < inputs.size(); i++) {
    out_shape[i] = inputs[i]->dims()[0];
  }
  auto out_dims = phi::make_ddim(std::vector<int>(out_shape));
  for (size_t i = 0; i < outputs.size(); ++i) {
    outputs[i]->set_dims(out_dims);
    outputs[i]->set_dtype(inputs[0]->dtype());
  }
}

2049 2050 2051 2052
void MomentumInferMeta(const MetaTensor& param,
                       const MetaTensor& grad,
                       const MetaTensor& velocity,
                       const MetaTensor& learning_rate,
2053
                       const MetaTensor& master_param,
2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095
                       float mu,
                       bool use_nesterov,
                       const std::string& regularization_method,
                       float regularization_coeff,
                       bool multi_precision,
                       float rescale_grad,
                       MetaTensor* param_out,
                       MetaTensor* velocity_out,
                       MetaTensor* master_param_out) {
  PADDLE_ENFORCE_NE(
      param_out,
      nullptr,
      errors::NotFound("Output(ParamOut) of Momentum should not be null."));
  PADDLE_ENFORCE_NE(
      velocity_out,
      nullptr,
      errors::NotFound("Output(VelocityOut) of Momentum should not be null."));

  auto lr_dims = learning_rate.dims();
  PADDLE_ENFORCE_NE(
      phi::product(lr_dims),
      0,
      errors::InvalidArgument("Maybe the Input variable LearningRate has not "
                              "been initialized. You may need to confirm "
                              "if you put exe.run(startup_program) "
                              "after optimizer.minimize function."));
  PADDLE_ENFORCE_EQ(
      phi::product(lr_dims),
      1,
      errors::InvalidArgument("Learning_rate should be a scalar. But Received "
                              "LearningRate's dim [%s]",
                              phi::product(lr_dims)));

  auto param_dim = param.dims();
  param_out->set_dims(param_dim);
  velocity_out->set_dims(param_dim);

  if (master_param_out) {
    master_param_out->set_dims(param_dim);
  }
}

2096 2097
void MultiDotInferMeta(const std::vector<const MetaTensor*>& x,
                       MetaTensor* out) {
2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169
  auto inputs_dims = GetMetaTensorsDim(x);

  const size_t inputs_num = inputs_dims.size();
  PADDLE_ENFORCE_GT(
      inputs_num,
      static_cast<size_t>(1),
      phi::errors::InvalidArgument(
          "The number of input tensors in multi_dot op should > 1."));

  const size_t n = inputs_dims.size();
  auto first_dim = inputs_dims[0];

  bool is_vector = false;
  phi::DDim out_dim;

  PADDLE_ENFORCE_LT(
      first_dim.size(),
      static_cast<size_t>(3),
      phi::errors::InvalidArgument(
          "multi_dot: the first input tensor must be 1D or 2D but got[%d]!",
          static_cast<int>(first_dim.size())));

  // If the first tensor is 1D of size n view it as a row vector (1, n)
  if (first_dim.size() == 1) {
    first_dim = phi::make_ddim({1, static_cast<int>(first_dim[0])});
    is_vector = true;
  }

  auto last_dim = inputs_dims[n - 1];
  PADDLE_ENFORCE_LT(
      last_dim.size(),
      static_cast<size_t>(3),
      phi::errors::InvalidArgument(
          "the last input tensor of multi_dot must be 1D or 2D but got[%d]!",
          static_cast<int>(first_dim.size())));

  // If the last tensor is 1D of size n view it as a column vector (n, 1)
  if (last_dim.size() == 1) {
    last_dim = phi::make_ddim({static_cast<int>(last_dim[0]), 1});
    out_dim = is_vector ? phi::make_ddim({1}) : phi::make_ddim({first_dim[0]});
  } else {
    out_dim = is_vector ? phi::make_ddim({last_dim[1]})
                        : phi::make_ddim({first_dim[0], last_dim[1]});
  }

  auto width = first_dim[1];
  for (size_t i = 1; i < n - 1; i++) {
    PADDLE_ENFORCE_EQ(inputs_dims[i].size(),
                      static_cast<size_t>(2),
                      phi::errors::InvalidArgument(
                          "the input tensor of multi_dot op must be 2D."));

    const auto& tmp_dim = inputs_dims[i];
    PADDLE_ENFORCE_EQ(
        tmp_dim[0],
        width,
        phi::errors::InvalidArgument(
            "the input matrix does not meet the multiplication requirements."));
    width = tmp_dim[1];
  }

  PADDLE_ENFORCE_EQ(
      last_dim[0],
      width,
      phi::errors::InvalidArgument(
          "the input matrix does not meet the multiplication requirements."));

  out->set_dims(out_dim);
  out->set_dtype(x.at(0)->dtype());
  out->share_lod(*x.at(0));
}

2170
void MultiplexInferMeta(const std::vector<const MetaTensor*>& ins,
2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213
                        const MetaTensor& ids,
                        MetaTensor* out) {
  PADDLE_ENFORCE_NE(
      ins.empty(),
      true,
      phi::errors::InvalidArgument("MultiInput(X) shouldn't be empty."));
  auto ids_dim = ids.dims();
  PADDLE_ENFORCE_EQ(ids_dim.size(),
                    2,
                    phi::errors::PreconditionNotMet(
                        "The index tensor must be a vector with 2 dimensions"));
  PADDLE_ENFORCE_EQ(
      ids_dim[1],
      1,
      phi::errors::PreconditionNotMet(
          "The index tensor must be a vector with batchSize x 1."));

  auto ins_dims = GetMetaTensorsDim(ins);
  auto num_ins = ins_dims.size();
  PADDLE_ENFORCE_GT(
      num_ins,
      1,
      phi::errors::InvalidArgument("multiplex operator should have more than "
                                   "one candidate input tensors."));

  auto in_dim = ins_dims[0];
  PADDLE_ENFORCE_GE(
      in_dim.size(),
      2,
      phi::errors::InvalidArgument(
          "The rank of candidate tensors must be not less than 2."));
  for (size_t i = 1; i < num_ins; i++) {
    auto dim = ins_dims[i];
    PADDLE_ENFORCE_EQ(
        in_dim,
        dim,
        phi::errors::PreconditionNotMet(
            "All the candidate tensors must have the same size."));
  }
  out->set_dims(in_dim);
  out->set_dtype(ins[0]->dtype());
}

F
From00 已提交
2214 2215
void PsroiPoolInferMeta(const MetaTensor& x,
                        const MetaTensor& rois,
2216
                        const MetaTensor& rois_num,
F
From00 已提交
2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238
                        int pooled_height,
                        int pooled_width,
                        int output_channels,
                        float spatial_scale,
                        MetaTensor* out) {
  auto input_dims = x.dims();
  auto rois_dims = rois.dims();

  PADDLE_ENFORCE_EQ(
      input_dims.size(),
      4,
      errors::InvalidArgument("The format of input tensor is NCHW"));
  PADDLE_ENFORCE_EQ(rois_dims.size(),
                    2,
                    errors::InvalidArgument(
                        "ROIs should be a 2-D LoDTensor of shape (num_rois, 4) "
                        "given as [(x1, y1, x2, y2), ...]"));
  PADDLE_ENFORCE_EQ(rois_dims[1],
                    4,
                    errors::InvalidArgument(
                        "ROIs should be a 2-D LoDTensor of shape (num_rois, 4) "
                        "given as [(x1, y1, x2, y2), ...]"));
2239 2240
  if (rois_num) {
    auto rois_num_dims = rois_num.dims();
F
From00 已提交
2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288
    PADDLE_ENFORCE_EQ(
        rois_num_dims.size(),
        1,
        errors::InvalidArgument("The second dimension of RoisNum should "
                                "be 1, but received dimension is %d",
                                rois_num_dims.size()));
  }

  PADDLE_ENFORCE_EQ(
      input_dims[1],
      output_channels * pooled_height * pooled_width,
      errors::InvalidArgument(
          "the channel of X(%d) "
          "should be equal to the product of "
          "output_channels(%d), pooled_height(%d) and pooled_width(%d)",
          input_dims[1],
          output_channels,
          pooled_height,
          pooled_width));

  PADDLE_ENFORCE_GT(pooled_height,
                    0,
                    errors::InvalidArgument(
                        "The pooled output height must be greater than 0"));
  PADDLE_ENFORCE_GT(pooled_width,
                    0,
                    errors::InvalidArgument(
                        "The pooled output width must be greater than 0"));
  PADDLE_ENFORCE_GT(output_channels,
                    1,
                    errors::InvalidArgument(
                        "The pooled output channels must greater than 1"));
  PADDLE_ENFORCE_GT(
      spatial_scale,
      0.0f,
      errors::InvalidArgument("The spatial scale must greater than 0."));

  auto out_dims = input_dims;
  out_dims[0] = rois_dims[0];
  out_dims[1] =
      output_channels;  // input_dims[1] / (pooled_height * pooled_width);
  out_dims[2] = pooled_height;
  out_dims[3] = pooled_width;

  out->set_dims(out_dims);
  out->set_dtype(x.dtype());
}

H
hong 已提交
2289 2290 2291 2292 2293
void RmspropInferMeta(const MetaTensor& param,
                      const MetaTensor& mean_square,
                      const MetaTensor& grad,
                      const MetaTensor& moment,
                      const MetaTensor& learning_rate,
2294
                      const MetaTensor& mean_grad,
H
hong 已提交
2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343
                      float epsilon,
                      float decay,
                      float momentum,
                      bool centered,
                      MetaTensor* param_out,
                      MetaTensor* moment_out,
                      MetaTensor* mean_square_out,
                      MetaTensor* mean_grad_out) {
  if (centered) {
    PADDLE_ENFORCE_NOT_NULL(
        mean_grad_out,
        phi::errors::InvalidArgument(
            "Output(MeanGradOut) of RmspropOp should not be null."));
  }

  auto param_dim = param.dims();
  PADDLE_ENFORCE_EQ(param_dim,
                    moment.dims(),
                    phi::errors::InvalidArgument(
                        "Param and Momentum input of RmspropOp "
                        "should have the same dimension. But received "
                        "Param's dim [%s] and Moment [%s]",
                        param_dim,
                        moment.dims()));
  PADDLE_ENFORCE_EQ(param_dim,
                    mean_square.dims(),
                    phi::errors::InvalidArgument(
                        "Param and Momentum input of RmspropOp "
                        "should have the same dimension. But received "
                        "Param's dim [%s] and MeanSquare [%s]",
                        param_dim,
                        mean_square.dims()));

  auto lr_dim = learning_rate.dims();
  PADDLE_ENFORCE_EQ(phi::product(lr_dim),
                    1,
                    phi::errors::InvalidArgument(
                        "Learning Rate of RmspropOp should be a scalar. But "
                        "received LearningRate's dim [%s]",
                        phi::product(lr_dim)));

  param_out->set_dims(param_dim);
  param_out->set_dtype(param.dtype());
  moment_out->set_dims(param_dim);
  moment_out->set_dtype(moment.dtype());
  mean_square_out->set_dims(param_dim);
  mean_square_out->set_dtype(mean_square.dtype());
  if (centered) {
    mean_grad_out->set_dims(param_dim);
2344
    mean_grad_out->set_dtype(mean_grad.dtype());
H
hong 已提交
2345 2346 2347
  }
}

2348
void RnnInferMeta(const MetaTensor& x,
2349 2350
                  const std::vector<const MetaTensor*>& pre_state,
                  const std::vector<const MetaTensor*>& weight_list,
2351
                  const MetaTensor& sequence_length,
2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373
                  float dropout_prob,
                  bool is_bidirec,
                  int input_size,
                  int hidden_size,
                  int num_layers,
                  const std::string& mode,
                  int seed,
                  bool is_test,
                  MetaTensor* out,
                  MetaTensor* dropout_state,
                  std::vector<MetaTensor*> state,
                  MetaTensor* reserve) {
  auto in_dims = x.dims();

  PADDLE_ENFORCE_EQ(
      in_dims.size(),
      3,
      phi::errors::InvalidArgument("The rank of Input in RNN  must be 3. But "
                                   "received Input's rank is %d.",
                                   in_dims.size()));

  if (sequence_length) {
2374
    auto seq_dims = sequence_length.dims();
2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432
    PADDLE_ENFORCE_EQ(
        in_dims[1],
        seq_dims[0],
        phi::errors::InvalidArgument(
            "The size of SequenceLength has to equal the batch_size. But "
            "received batch_size is %d and the size of SequenceLength is %d.",
            in_dims[1],
            seq_dims[0]));
  }

  PADDLE_ENFORCE_EQ(pre_state[0]->dims().size(),
                    3,
                    phi::errors::InvalidArgument(
                        "The rank of PreState in RNN  must be 3. But "
                        "the received rank is %d.",
                        pre_state[0]->dims().size()));
  size_t i = 0;
  for (; i < pre_state.size(); ++i) {
    PADDLE_ENFORCE_EQ(
        in_dims[1],
        pre_state[i]->dims()[1],
        phi::errors::InvalidArgument(
            "The second dimension size (representing for batch size) of "
            "Input and PreState should be equal. But received %d and %d.",
            in_dims[1],
            pre_state[i]->dims()[1]));
    PADDLE_ENFORCE_EQ(
        pre_state[0]->dims(),
        pre_state[i]->dims(),
        phi::errors::InvalidArgument(
            "The dims of all tensors in PreState should be same. But "
            "received PreState[0] is %s and PreState[%d] is %s.",
            pre_state[0]->dims(),
            i,
            pre_state[i]->dims()));
  }
  size_t num_state = mode == "LSTM" ? 2 : 1;
  PADDLE_ENFORCE_EQ(i,
                    num_state,
                    phi::errors::InvalidArgument(
                        "The number of tensors in PreState of %s should be %d, "
                        "but received %d.",
                        mode,
                        2,
                        i));

  auto out_dims = in_dims;
  out_dims[2] = is_bidirec ? hidden_size * 2 : hidden_size;
  out->set_dims(out_dims);
  out->set_dtype(x.dtype());

  int state_num = pre_state.size();
  for (int i = 0; i < state_num; ++i) {
    state[i]->set_dims(pre_state[i]->dims());
    state[i]->set_dtype(x.dtype());
  }
}

Z
zyfncg 已提交
2433
void SgdInferMeta(const MetaTensor& param,
H
hong 已提交
2434 2435
                  const MetaTensor& learning_rate,
                  const MetaTensor& grad,
2436
                  const MetaTensor& master_param,
H
hong 已提交
2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455
                  bool multi_precision,
                  MetaTensor* param_out,
                  MetaTensor* master_param_out) {
  PADDLE_ENFORCE_NOT_NULL(param_out,
                          phi::errors::InvalidArgument(
                              "Output(ParamOut) of SGDOp should not be null."));

  auto lr_dims = learning_rate.dims();
  PADDLE_ENFORCE_EQ(phi::product(lr_dims),
                    1,
                    phi::errors::InvalidArgument(
                        "Learning rate should have 1 element. But received "
                        "LearningRate dims [%s]",
                        phi::product(lr_dims)));

  param_out->set_dims(param.dims());
  param_out->set_dtype(param.dtype());
}

2456
void StackInferMeta(const std::vector<const MetaTensor*>& x,
C
csy0225 已提交
2457
                    int axis,
2458 2459
                    MetaTensor* out,
                    MetaConfig config) {
C
csy0225 已提交
2460 2461 2462 2463 2464 2465 2466
  PADDLE_ENFORCE_GT(x.size(),
                    0UL,
                    phi::errors::InvalidArgument(
                        "Number of Inputs(x) must be larger than 0, but"
                        " received value is:%d.",
                        x.size()));
  const auto& input_dims = GetMetaTensorsDim(x);
2467 2468 2469 2470
  // we reuse concat logic to compute out_dim. we set concat_axis==-1 to check
  // every axis in input_tensors.
  auto out_dim =
      phi::funcs::ComputeAndCheckShape(config.is_runtime, input_dims, -1);
C
csy0225 已提交
2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488
  int rank = input_dims[0].size();
  PADDLE_ENFORCE_GE(
      axis,
      -(rank + 1),
      phi::errors::InvalidArgument(
          "Attr(axis) must be inside [-(rank+1), rank+1), where rank = %d, "
          "but received axis is:%d.",
          rank,
          axis));
  PADDLE_ENFORCE_LT(
      axis,
      rank + 1,
      phi::errors::InvalidArgument(
          "Attr(axis) must be inside [-(rank+1), rank+1), where rank = %d, "
          "but received axis is:%d",
          rank,
          axis));
  if (axis < 0) axis += (rank + 1);
2489
  auto vec = phi::vectorize<int>(out_dim);
C
csy0225 已提交
2490 2491 2492 2493 2494 2495
  vec.insert(vec.begin() + axis, input_dims.size());
  out->set_dims(phi::make_ddim(vec));
  out->set_dtype(x.at(0)->dtype());
  out->share_lod(*x.at(0));
}

2496
void UnchangedMultiInferMeta(const std::vector<const MetaTensor*>& x,
2497 2498
                             std::vector<MetaTensor*> out) {
  for (size_t i = 0; i < x.size(); ++i) {
2499 2500 2501
    if (out[i]) {
      out[i]->share_meta(*x[i]);
    }
2502 2503 2504
  }
}

2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522
void UpdateLossScalingInferMeta(const std::vector<const MetaTensor*>& xs,
                                const MetaTensor& found_infinite,
                                const MetaTensor& prev_loss_scaling,
                                const MetaTensor& in_good_steps,
                                const MetaTensor& in_bad_steps,
                                std::vector<MetaTensor*> outs,
                                MetaTensor* loss_scaling,
                                MetaTensor* out_good_steps,
                                MetaTensor* out_bad_steps) {
  PADDLE_ENFORCE_EQ(xs.size(),
                    outs.size(),
                    phi::errors::InvalidArgument(
                        "The input(X) and output(Out) should have same size in "
                        "Operator(update_loss_scaling), size of input(X) is %d "
                        "and size of output(Out) is %d.",
                        xs.size(),
                        outs.size()));
  for (size_t i = 0; i < xs.size(); ++i) {
2523 2524 2525 2526
    if (xs[i] != nullptr && outs[i] != nullptr) {
      outs[i]->set_dims(xs[i]->dims());
      outs[i]->set_dtype(xs[i]->dtype());
    }
2527 2528 2529 2530 2531 2532 2533 2534
  }
  loss_scaling->set_dims({1});
  out_good_steps->set_dims({1});
  out_good_steps->set_dtype(DataType::INT32);
  out_bad_steps->set_dims({1});
  out_bad_steps->set_dtype(DataType::INT32);
}

0
0x45f 已提交
2535 2536
void WarpctcInferMeta(const MetaTensor& logits,
                      const MetaTensor& label,
2537 2538
                      const MetaTensor& logits_length,
                      const MetaTensor& labels_length,
0
0x45f 已提交
2539 2540
                      int blank,
                      bool norm_by_times,
Z
Zhong Hui 已提交
2541
                      MetaTensor* warpctcgrad,
0
0x45f 已提交
2542 2543 2544 2545
                      MetaTensor* loss) {
  auto logits_dims = logits.dims();
  int sequence_width = 0;

2546
  if (logits_length) {
0
0x45f 已提交
2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571
    sequence_width = logits_dims[2];
  } else {
    sequence_width =
        static_cast<int>(phi::product(logits_dims) / logits_dims[0]);
  }

  PADDLE_ENFORCE_GE(
      blank,
      0,
      errors::InvalidArgument(
          "The value of Attr(blank) should be in interval [0, %d), "
          "but received %d",
          blank));
  PADDLE_ENFORCE_LT(
      blank,
      sequence_width,
      errors::InvalidArgument(
          "The value of Attr(blank) should be in interval [0, %d), "
          "but received %d",
          blank));

  loss->set_dims({-1, 1});
  loss->set_dtype(logits.dtype());
}

2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596
void WhereInferMeta(const MetaTensor& condition,
                    const MetaTensor& x,
                    const MetaTensor& y,
                    MetaTensor* out) {
  auto cond_dims = condition.dims();
  auto x_dims = x.dims();
  auto y_dims = y.dims();
  PADDLE_ENFORCE_EQ(
      cond_dims,
      x_dims,
      phi::errors::InvalidArgument(
          "The dims of Inputs(Condition) and Inputs(X) should be same. "
          "But received Condition's shape is [%s], X's shape is [%s]",
          cond_dims,
          x_dims));
  PADDLE_ENFORCE_EQ(x_dims,
                    y_dims,
                    phi::errors::InvalidArgument(
                        "The dims of Inputs(X) and Inputs(Y) should be same. "
                        "But received X's shape is [%s], Y's shape is [%s]",
                        x_dims,
                        y_dims));
  out->share_meta(x);
}

2597 2598 2599
void Yolov3LossInferMeta(const MetaTensor& x,
                         const MetaTensor& gt_box,
                         const MetaTensor& gt_label,
2600
                         const MetaTensor& gt_score,
2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704
                         const std::vector<int>& anchors,
                         const std::vector<int>& anchor_mask,
                         int class_num,
                         float ignore_thresh,
                         int downsample_ratio,
                         bool use_label_smooth,
                         float scale_x_y,
                         MetaTensor* loss,
                         MetaTensor* objectness_mask,
                         MetaTensor* gt_match_mask) {
  auto dim_x = x.dims();
  auto dim_gtbox = gt_box.dims();
  auto dim_gtlabel = gt_label.dims();
  int anchor_num = anchors.size() / 2;
  int mask_num = anchor_mask.size();

  PADDLE_ENFORCE_EQ(dim_x.size(),
                    4,
                    phi::errors::InvalidArgument(
                        "Input(X) should be a 4-D tensor. But received "
                        "X dimension size(%s)",
                        dim_x.size()));
  PADDLE_ENFORCE_EQ(
      dim_x[2],
      dim_x[3],
      phi::errors::InvalidArgument("Input(X) dim[3] and dim[4] should be euqal."
                                   "But received dim[3](%s) != dim[4](%s)",
                                   dim_x[2],
                                   dim_x[3]));
  PADDLE_ENFORCE_EQ(
      dim_x[1],
      mask_num * (5 + class_num),
      phi::errors::InvalidArgument(
          "Input(X) dim[1] should be equal to (anchor_mask_number * (5 "
          "+ class_num))."
          "But received dim[1](%s) != (anchor_mask_number * "
          "(5+class_num)(%s).",
          dim_x[1],
          mask_num * (5 + class_num)));
  PADDLE_ENFORCE_EQ(
      dim_gtbox.size(),
      3,
      phi::errors::InvalidArgument("Input(GTBox) should be a 3-D tensor, but "
                                   "received gtbox dimension size(%s)",
                                   dim_gtbox.size()));
  PADDLE_ENFORCE_EQ(
      dim_gtbox[2],
      4,
      phi::errors::InvalidArgument("Input(GTBox) dim[2] should be 4",
                                   "But receive dim[2](%s) != 5. ",
                                   dim_gtbox[2]));
  PADDLE_ENFORCE_EQ(dim_gtlabel.size(),
                    2,
                    phi::errors::InvalidArgument(
                        "Input(GTLabel) should be a 2-D tensor,"
                        "But received Input(GTLabel) dimension size(%s) != 2.",
                        dim_gtlabel.size()));
  PADDLE_ENFORCE_EQ(
      dim_gtlabel[0],
      dim_gtbox[0],
      phi::errors::InvalidArgument(
          "Input(GTBox) dim[0] and Input(GTLabel) dim[0] should be same,"
          "But received Input(GTLabel) dim[0](%s) != "
          "Input(GTBox) dim[0](%s)",
          dim_gtlabel[0],
          dim_gtbox[0]));
  PADDLE_ENFORCE_EQ(
      dim_gtlabel[1],
      dim_gtbox[1],
      phi::errors::InvalidArgument(
          "Input(GTBox) and Input(GTLabel) dim[1] should be same,"
          "But received Input(GTBox) dim[1](%s) != Input(GTLabel) "
          "dim[1](%s)",
          dim_gtbox[1],
          dim_gtlabel[1]));
  PADDLE_ENFORCE_GT(anchors.size(),
                    0,
                    phi::errors::InvalidArgument(
                        "Attr(anchors) length should be greater then 0."
                        "But received anchors length(%s)",
                        anchors.size()));
  PADDLE_ENFORCE_EQ(anchors.size() % 2,
                    0,
                    phi::errors::InvalidArgument(
                        "Attr(anchors) length should be even integer."
                        "But received anchors length(%s)",
                        anchors.size()));
  for (size_t i = 0; i < anchor_mask.size(); i++) {
    PADDLE_ENFORCE_LT(
        anchor_mask[i],
        anchor_num,
        phi::errors::InvalidArgument(
            "Attr(anchor_mask) should not crossover Attr(anchors)."
            "But received anchor_mask[i](%s) > anchor_num(%s)",
            anchor_mask[i],
            anchor_num));
  }
  PADDLE_ENFORCE_GT(class_num,
                    0,
                    phi::errors::InvalidArgument(
                        "Attr(class_num) should be an integer greater then 0."
                        "But received class_num(%s) < 0",
                        class_num));

2705 2706
  if (gt_score) {
    auto dim_gtscore = gt_score.dims();
2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743
    PADDLE_ENFORCE_EQ(
        dim_gtscore.size(),
        2,
        phi::errors::InvalidArgument("Input(GTScore) should be a 2-D tensor"
                                     "But received GTScore dimension(%s)",
                                     dim_gtbox.size()));
    PADDLE_ENFORCE_EQ(
        dim_gtscore[0],
        dim_gtbox[0],
        phi::errors::InvalidArgument(
            "Input(GTBox) and Input(GTScore) dim[0] should be same"
            "But received GTBox dim[0](%s) != GTScore dim[0](%s)",
            dim_gtbox[0],
            dim_gtscore[0]));
    PADDLE_ENFORCE_EQ(
        dim_gtscore[1],
        dim_gtbox[1],
        phi::errors::InvalidArgument(
            "Input(GTBox) and Input(GTScore) dim[1] should be same"
            "But received GTBox dim[1](%s) != GTScore dim[1](%s)",
            dim_gtscore[1],
            dim_gtbox[1]));
  }

  std::vector<int64_t> dim_out({dim_x[0]});
  loss->set_dims(phi::make_ddim(dim_out));
  loss->set_dtype(x.dtype());

  std::vector<int64_t> dim_obj_mask({dim_x[0], mask_num, dim_x[2], dim_x[3]});
  objectness_mask->set_dims(phi::make_ddim(dim_obj_mask));
  objectness_mask->set_dtype(x.dtype());

  std::vector<int64_t> dim_gt_match_mask({dim_gtbox[0], dim_gtbox[1]});
  gt_match_mask->set_dims(phi::make_ddim(dim_gt_match_mask));
  gt_match_mask->set_dtype(x.dtype());
}

2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 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 2824 2825 2826 2827 2828 2829 2830 2831
void GraphSendUERecvInferMeta(const MetaTensor& x,
                              const MetaTensor& y,
                              const MetaTensor& src_index,
                              const MetaTensor& dst_index,
                              const std::string& message_op,
                              const std::string& reduce_op,
                              const IntArray& out_size,
                              MetaTensor* out,
                              MetaTensor* dst_count) {
  auto src_index_dims = src_index.dims();
  if (src_index_dims.size() == 2) {
    PADDLE_ENFORCE_EQ(src_index_dims[1],
                      1,
                      phi::errors::InvalidArgument(
                          "The last dim of Src_index should be 1 when it "
                          "is 2D, but we get %d",
                          src_index_dims[1]));
  } else {
    PADDLE_ENFORCE_EQ(
        src_index_dims.size(),
        1,
        phi::errors::InvalidArgument(
            "The Src_index should be 1D, when it is not 2D, but we get %d",
            src_index_dims.size()));
  }

  auto dst_index_dims = dst_index.dims();
  if (dst_index_dims.size() == 2) {
    PADDLE_ENFORCE_EQ(dst_index_dims[1],
                      1,
                      phi::errors::InvalidArgument(
                          "The last dim of Dst_index should be 1 when it "
                          "is 2D, but we get %d",
                          dst_index_dims[1]));
  } else {
    PADDLE_ENFORCE_EQ(
        dst_index_dims.size(),
        1,
        phi::errors::InvalidArgument("The Dst_index should be 1D, "
                                     "when it is not 2D, but we get %d",
                                     dst_index_dims.size()));
  }

  PADDLE_ENFORCE_EQ(src_index_dims[0],
                    dst_index_dims[0],
                    phi::errors::InvalidArgument(
                        "Src_index and Dst_index should have the same shape."));

  auto y_dims = y.dims();
  PADDLE_ENFORCE_EQ(
      y_dims[0],
      src_index_dims[0],
      phi::errors::InvalidArgument(
          "Expect Input Y to have size %d as Src_index on the first dimension, "
          "but we get %d",
          src_index_dims[0],
          y_dims[0]));

  auto x_dims = x.dims();
  if (reduce_op == "MEAN") {
    dst_count->set_dims({-1});
    dst_count->set_dtype(DataType::INT32);
  }

  // Infer out's shape according to x and e(need broadcasting condition)
  out->set_dtype(x.dtype());
  auto x_dims1 = phi::vectorize<int>(x_dims);
  auto y_dims1 = phi::vectorize<int>(y_dims);
  std::vector<int> x_dims2(x_dims1.begin() + 1, x_dims1.end());
  std::vector<int> y_dims2(y_dims1.begin() + 1, y_dims1.end());

  int max_dim = std::max(x_dims2.size(), y_dims2.size());
  int axis = std::abs(static_cast<int>(x_dims2.size() - y_dims2.size()));
  std::vector<int> x_dims_array(max_dim);
  std::vector<int> y_dims_array(max_dim);
  std::vector<int> out_dims_array(max_dim);
  // Only need to broadcast dimensions other than the 0th dimension.
  phi::funcs::GetBroadcastDimsArrays(phi::make_ddim(x_dims2),
                                     phi::make_ddim(y_dims2),
                                     x_dims_array.data(),
                                     y_dims_array.data(),
                                     out_dims_array.data(),
                                     max_dim,
                                     axis);
  out_dims_array.insert(out_dims_array.begin(), -1);
  out->set_dims(phi::make_ddim(out_dims_array));
}

2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901
void GraphSendUVInferMeta(const MetaTensor& x,
                          const MetaTensor& y,
                          const MetaTensor& src_index,
                          const MetaTensor& dst_index,
                          const std::string& message_op,
                          MetaTensor* out) {
  auto src_index_dims = src_index.dims();
  if (src_index_dims.size() == 2) {
    PADDLE_ENFORCE_EQ(src_index_dims[1],
                      1,
                      phi::errors::InvalidArgument(
                          "The last dim of Src_index should be 1 when it "
                          "is 2D, but we get %d",
                          src_index_dims[1]));
  } else {
    PADDLE_ENFORCE_EQ(
        src_index_dims.size(),
        1,
        phi::errors::InvalidArgument(
            "The Src_index should be 1D, when it is not 2D, but we get %d",
            src_index_dims.size()));
  }

  auto dst_index_dims = dst_index.dims();
  if (dst_index_dims.size() == 2) {
    PADDLE_ENFORCE_EQ(dst_index_dims[1],
                      1,
                      phi::errors::InvalidArgument(
                          "The last dim of Dst_index should be 1 when it "
                          "is 2D, but we get %d",
                          dst_index_dims[1]));
  } else {
    PADDLE_ENFORCE_EQ(
        dst_index_dims.size(),
        1,
        phi::errors::InvalidArgument("The Dst_index should be 1D, "
                                     "when it is not 2D, but we get %d",
                                     dst_index_dims.size()));
  }

  PADDLE_ENFORCE_EQ(src_index_dims[0],
                    dst_index_dims[0],
                    phi::errors::InvalidArgument(
                        "Src_index and Dst_index should have the same shape."));

  // Infer out's shape according to x and y(need broadcasting condition)
  out->set_dtype(x.dtype());
  auto x_dims = x.dims();
  auto y_dims = y.dims();
  auto x_dims1 = phi::vectorize<int>(x_dims);
  auto y_dims1 = phi::vectorize<int>(y_dims);
  std::vector<int> x_dims2(x_dims1.begin() + 1, x_dims1.end());
  std::vector<int> y_dims2(y_dims1.begin() + 1, y_dims1.end());
  int max_dim = std::max(x_dims2.size(), y_dims2.size());
  int axis = std::abs(static_cast<int>(x_dims2.size() - y_dims2.size()));
  std::vector<int> x_dims_array(max_dim);
  std::vector<int> y_dims_array(max_dim);
  std::vector<int> out_dims_array(max_dim);
  // Only need to broadcast dimensions other than the 0th dimension.
  phi::funcs::GetBroadcastDimsArrays(phi::make_ddim(x_dims2),
                                     phi::make_ddim(y_dims2),
                                     x_dims_array.data(),
                                     y_dims_array.data(),
                                     out_dims_array.data(),
                                     max_dim,
                                     axis);
  out_dims_array.insert(out_dims_array.begin(), src_index_dims[0]);
  out->set_dims(phi::make_ddim(out_dims_array));
}

2902
}  // namespace phi
H
hong 已提交
2903

2904
PD_REGISTER_INFER_META_FN(batch_norm_infer, phi::BatchNormInferInferMeta);