interpolate_v2_op.cc 29.4 KB
Newer Older
X
xiaoting 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
   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. */

#include "paddle/fluid/operators/interpolate_v2_op.h"
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
17 18 19
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
X
xiaoting 已提交
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

namespace paddle {
namespace operators {

using framework::Tensor;
using DataLayout = framework::DataLayout;

static void Interpolate1DInferShapeCheck(framework::InferShapeContext* ctx) {
  auto dim_x = ctx->GetInputDim("X");
  auto interp_method = ctx->Attrs().Get<std::string>("interp_method");

  PADDLE_ENFORCE_EQ("linear", interp_method,
                    platform::errors::InvalidArgument(
                        "Interpolation method can only be \"linear\" when"
                        "Input(X) dimension is 3, but got method = %s .",
                        interp_method));
  const DataLayout data_layout = framework::StringToDataLayout(
      ctx->Attrs().Get<std::string>("data_layout"));

  if (ctx->HasInputs("SizeTensor")) {
    // top prority size
    auto inputs_name = ctx->Inputs("SizeTensor");
    PADDLE_ENFORCE_EQ(
        inputs_name.size(), 1,
        platform::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()));
    int out_w = ctx->Attrs().Get<int>("out_w");
    framework::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]};
    }
    ctx->SetOutputDim("Out", dim_out);

    return;
  }

  int out_w;
  if (ctx->HasInput("Scale")) {
    auto scale_tensor = ctx->GetInputDim("Scale");
    PADDLE_ENFORCE_EQ(
        scale_tensor.size(), 1,
        platform::errors::InvalidArgument(
            "Scale's dimension size must be 1, but got dimension = %d .",
            scale_tensor.size()));
    PADDLE_ENFORCE_EQ(
        scale_tensor[0], 1,
        platform::errors::InvalidArgument(
            "Scale's shape must be 1, but got shape = %d .", scale_tensor[0]));
73
    out_w = -1;
X
xiaoting 已提交
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123
  } else {
    auto scale = ctx->Attrs().Get<std::vector<float>>("scale");
    if (scale.size() > 0) {
      float scale_w = -1;
      scale_w = scale[0];
      PADDLE_ENFORCE_EQ(scale_w > 0, true, platform::errors::InvalidArgument(
                                               "scale  of Op(interpolate) "
                                               "should be greater than 0."));
      if (scale_w > 0.) {
        // round down
        out_w = (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 = out_w > 0 ? out_w : -1;
      }
    } else {
      out_w = ctx->Attrs().Get<int>("out_w");
    }
  }

  if (ctx->HasInput("OutSize") && ctx->IsRuntime()) {
    auto out_size_dim = ctx->GetInputDim("OutSize");
    PADDLE_ENFORCE_EQ(
        out_size_dim.size(), 1,
        platform::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, platform::errors::InvalidArgument(
                                              "OutSize's dim[0] must be 1"));
    ctx->ShareLoD("X", "Out");
    return;
  }

  framework::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]};
  }
  ctx->SetOutputDim("Out", dim_out);
}

static void Interpolate2DInferShapeCheck(framework::InferShapeContext* ctx) {
  auto dim_x = ctx->GetInputDim("X");
  auto interp_method = ctx->Attrs().Get<std::string>("interp_method");

  PADDLE_ENFORCE(
      "bilinear" == interp_method || "nearest" == interp_method ||
          "bicubic" == interp_method,
124 125 126 127
      platform::errors::InvalidArgument(
          "Interpolation method can only be \"bilinear\" or \"nearest\" when "
          "Input(X) dimension is 4, but got method = %s.",
          interp_method));
X
xiaoting 已提交
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
  const DataLayout data_layout = framework::StringToDataLayout(
      ctx->Attrs().Get<std::string>("data_layout"));

  if (ctx->HasInputs("SizeTensor")) {
    // top prority size
    auto inputs_name = ctx->Inputs("SizeTensor");
    PADDLE_ENFORCE_EQ(
        inputs_name.size(), 2,
        platform::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()));
    int out_h = ctx->Attrs().Get<int>("out_h");
    int out_w = ctx->Attrs().Get<int>("out_w");
    framework::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]};
    }
    ctx->SetOutputDim("Out", dim_out);

    return;
  }

  int out_h, out_w;
  if (ctx->HasInput("Scale")) {
    auto scale_tensor = ctx->GetInputDim("Scale");
    PADDLE_ENFORCE_EQ(
        scale_tensor.size(), 1,
        platform::errors::InvalidArgument(
            "Scale's dimension size must be 1, but got dimension = %d .",
            scale_tensor.size()));
    PADDLE_ENFORCE_EQ(scale_tensor[0] == 2 || scale_tensor[0] == 1, true,
                      platform::errors::InvalidArgument(
                          "Scale's shape must be 2 or 1, but got shape = %d .",
                          scale_tensor[0]));
166 167
    out_h = -1;
    out_w = -1;
X
xiaoting 已提交
168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270
  } else {
    auto scale = ctx->Attrs().Get<std::vector<float>>("scale");
    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 && scale_h > 0, true,
          platform::errors::InvalidArgument("scale  of Op(interpolate) "
                                            "should be greater than 0."));
      if (scale_h > 0. && scale_w > 0.) {
        // round down
        out_h = (data_layout == DataLayout::kNCHW
                     ? static_cast<int>(dim_x[2] * scale_h)
                     : static_cast<int>(dim_x[1] * scale_h));
        out_w = (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 = out_h > 0 ? out_h : -1;
        out_w = out_w > 0 ? out_w : -1;
      }
    } else {
      out_h = ctx->Attrs().Get<int>("out_h");
      out_w = ctx->Attrs().Get<int>("out_w");
    }
  }

  if (ctx->HasInput("OutSize") && ctx->IsRuntime()) {
    auto out_size_dim = ctx->GetInputDim("OutSize");
    PADDLE_ENFORCE_EQ(
        out_size_dim.size(), 1,
        platform::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,
        platform::errors::InvalidArgument(
            "OutSize's dim[0] must be 2, but got dimention = %d .",
            out_size_dim[0]));
    ctx->ShareLoD("X", "Out");
    return;
  }

  framework::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]};
  }
  ctx->SetOutputDim("Out", dim_out);
}

static void Interpolate3DInferShapeCheck(framework::InferShapeContext* ctx) {
  auto dim_x = ctx->GetInputDim("X");
  auto interp_method = ctx->Attrs().Get<std::string>("interp_method");

  PADDLE_ENFORCE_EQ(
      "trilinear", interp_method,
      platform::errors::InvalidArgument(
          "Interpolation method can only be \"trilinear\" when Input(X) "
          "dimension is 5, but got method = %s .",
          interp_method));
  const DataLayout data_layout = framework::StringToDataLayout(
      ctx->Attrs().Get<std::string>("data_layout"));

  if (ctx->HasInputs("SizeTensor")) {
    // top prority size
    auto inputs_name = ctx->Inputs("SizeTensor");
    PADDLE_ENFORCE_EQ(
        inputs_name.size(), 3,
        platform::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()));
    int out_d = ctx->Attrs().Get<int>("out_d");
    int out_h = ctx->Attrs().Get<int>("out_h");
    int out_w = ctx->Attrs().Get<int>("out_w");
    framework::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]};
    }
    ctx->SetOutputDim("Out", dim_out);

    return;
  }

  int out_d, out_h, out_w;
  if (ctx->HasInput("Scale")) {
    auto scale_tensor = ctx->GetInputDim("Scale");
    PADDLE_ENFORCE_EQ(
        scale_tensor.size(), 1,
        platform::errors::InvalidArgument(
            "Scale's dimension size must be 1, but got size = %d .",
            scale_tensor.size()));
    PADDLE_ENFORCE_EQ(scale_tensor[0] == 3 || scale_tensor[0] == 1, true,
                      platform::errors::InvalidArgument(
                          "Scale's shape must be 3 or 1, but got shape = %d .",
                          scale_tensor[0]));
271 272 273
    out_d = -1;
    out_h = -1;
    out_w = -1;
X
xiaoting 已提交
274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311
  } else {
    auto scale = ctx->Attrs().Get<std::vector<float>>("scale");
    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 && scale_h > 0 && scale_d > 0, true,
          platform::errors::InvalidArgument("scale  of Op(interpolate) "
                                            "should be greater than 0."));
      if (scale_d > 0. && scale_h > 0. && scale_w > 0.) {
        // round down
        out_d = (data_layout == DataLayout::kNCHW
                     ? static_cast<int>(dim_x[2] * scale_d)
                     : static_cast<int>(dim_x[1] * scale_d));
        out_h = (data_layout == DataLayout::kNCHW
                     ? static_cast<int>(dim_x[3] * scale_h)
                     : static_cast<int>(dim_x[2] * scale_h));
        out_w = (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 = out_d > 0 ? out_d : -1;
        out_h = out_h > 0 ? out_h : -1;
        out_w = out_w > 0 ? out_w : -1;
      }
    } else {
      out_d = ctx->Attrs().Get<int>("out_d");
      out_h = ctx->Attrs().Get<int>("out_h");
      out_w = ctx->Attrs().Get<int>("out_w");
    }
  }

  if (ctx->HasInput("OutSize") && ctx->IsRuntime()) {
    auto out_size_dim = ctx->GetInputDim("OutSize");
312 313 314 315 316
    PADDLE_ENFORCE_EQ(
        out_size_dim.size(), 1,
        platform::errors::InvalidArgument(
            "OutSize's dimension size must be 1, but got size is %d.",
            out_size_dim.size()));
X
xiaoting 已提交
317
    PADDLE_ENFORCE_EQ(out_size_dim[0], 3,
318 319 320
                      platform::errors::InvalidArgument(
                          "OutSize's dim[0] must be 3, but got size is %d.",
                          out_size_dim[0]));
X
xiaoting 已提交
321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339
    ctx->ShareLoD("X", "Out");
    return;
  }

  framework::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]};
  }
  ctx->SetOutputDim("Out", dim_out);
}

class InterpolateV2Op : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(framework::InferShapeContext* ctx) const override {
340 341
    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "Interpolate");
    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "Interpolate");
X
xiaoting 已提交
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364

    auto dim_x = ctx->GetInputDim("X");  // NCHW format
    PADDLE_ENFORCE(
        dim_x.size() == 3 || dim_x.size() == 4 || dim_x.size() == 5,
        platform::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(ctx);
    } else if (dim_x.size() == 4) {
      // shape check for 2D interpolate for input tensor shape NCHW
      Interpolate2DInferShapeCheck(ctx);
    } else {  // dim_x.size() == 5
      // shape check for 3D interpolate for input tensor shape NCDHW
      Interpolate3DInferShapeCheck(ctx);
    }
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
365 366 367 368 369 370 371 372 373 374 375 376 377 378 379
    framework::DataLayout layout = framework::DataLayout::kAnyLayout;
    framework::LibraryType library = framework::LibraryType::kPlain;
    auto data_type = OperatorWithKernel::IndicateVarDataType(ctx, "X");

#ifdef PADDLE_WITH_MKLDNN
    auto interp_method = ctx.Attr<std::string>("interp_method");
    // TODO(danqing): support other interp_method
    if (this->CanMKLDNNBeUsed(ctx, data_type) &&
        (interp_method == "nearest" || interp_method == "bilinear")) {
      layout = framework::DataLayout::kMKLDNN;
      library = framework::LibraryType::kMKLDNN;
    }
#endif

    return framework::OpKernelType(data_type, ctx.GetPlace(), layout, library);
X
xiaoting 已提交
380 381 382 383 384
  }

  framework::OpKernelType GetKernelTypeForVar(
      const std::string& var_name, const Tensor& tensor,
      const framework::OpKernelType& expected_kernel_type) const override {
385 386 387 388 389 390 391 392 393 394 395 396 397 398 399
#ifdef PADDLE_WITH_MKLDNN
    if ((expected_kernel_type.data_layout_ == framework::DataLayout::kMKLDNN) &&
        (tensor.layout() != framework::DataLayout::kMKLDNN)) {
      auto attrs = Attrs();
      auto ar = paddle::framework::AttrReader(attrs);
      const std::string data_format = ar.Get<std::string>("data_layout");
      auto dl = framework::StringToDataLayout(data_format);
      // Some models may have intentionally set "AnyLayout" for pool
      // op. Treat this as NCHW (default data_format value)
      if (dl != framework::DataLayout::kAnyLayout) {
        return framework::OpKernelType(expected_kernel_type.data_type_,
                                       tensor.place(), dl);
      }
    }
#endif
X
xiaoting 已提交
400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469
    if (var_name == "SizeTensor" || var_name == "Scale") {
      return expected_kernel_type;
    }
    return framework::OpKernelType(expected_kernel_type.data_type_,
                                   tensor.place(), tensor.layout());
  }
};

class InterpolateV2OpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X",
             "The input tensor of interpolate operator, "
             "This is a 4-D tensor with shape of [N, C, H, W] or a "
             "5-D tensor with shape of [N, C, D, H, W].");
    AddInput("OutSize",
             "This is a 1-D tensor with two numbers to specify output size. "
             "It should be [output_height, output_width] when input is a 4-D "
             "tensor and should be [output_depth, output_height, output_width] "
             "when input is a 5-D tensor. It has a higher priority than "
             "the attr(out_d), attr(out_h), attr(out_w) and attr(scale).")
        .AsDispensable();
    AddInput("SizeTensor",
             "(vector<Tensor<int32>>, optional). If provided, interpolate will "
             "use this. The shape of the tensor in vector MUST BE [1]. "
             "It has the highest priority compare with Input(OutSize) and "
             "attr(out_d), attr(out_h), attr(out_w) and attr(scale).")
        .AsDuplicable()
        .AsDispensable();
    AddInput("Scale",
             "This is a 1-D tensor with one number to specify output scale. "
             "It has the higher priority compare with attr(scale).")
        .AsDispensable();
    AddOutput("Out",
              "The output tensor of interpolate operator, "
              "This is a tensor in same rank with Input(X).");

    AddAttr<std::string>(
        "data_layout",
        "(string, default NCHW) Only used in "
        "an optional string from: \"NHWC\", \"NCHW\". "
        "Specify that the data format of the input and output data is "
        "channel_first or channel_last.")
        .SetDefault("NCHW");
    AddAttr<int>("out_d", "output depth of interpolate op.").SetDefault(0);
    AddAttr<int>("out_h", "output height of interpolate op.").SetDefault(0);
    AddAttr<int>("out_w", "output width of interpolate op.").SetDefault(0);
    AddAttr<std::vector<float>>("scale", "scale_d factor of interpolate op.")
        .SetDefault(std::vector<float>{});
    AddAttr<std::string>("interp_method",
                         "(string, default \"bilinear\"), interpolation "
                         "method, can be \"linear\" for linear interpolation"
                         ",\"bilinear\" for "
                         "bilinear interpolation, \"trilinear\" for trilinear "
                         "interpolation and \"nearest\" for nearest "
                         "neighbor interpolation, and \"bicubic\" for bicubic"
                         "interpolation.")
        .SetDefault("bilinear");
    AddAttr<bool>(
        "align_corners",
        "an optional bool. Defaults to True. "
        "If True, the centers of 4 corner pixels of the input and output "
        "tensors are aligned, preserving the values at the corner pixels, "
        "If False, are not aligned")
        .SetDefault(true);
    AddAttr<int>("align_mode",
                 "(int, default \'1\'), optional for bilinear interpolation, "
                 "can be \'0\' for src_idx = scale*(dst_indx+0.5)-0.5 , "
                 "can be \'1\' for src_idx = scale*dst_index .")
        .SetDefault(1);
470 471 472
    AddAttr<bool>("use_mkldnn",
                  "(bool, default false) Only used in mkldnn kernel")
        .SetDefault(false);
X
xiaoting 已提交
473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614
    AddComment(R"DOC(
          This operator samples input X to given output shape by using specified
          interpolation method, the interpolation methods can be \"nearest\"
          for nearest neighbor interpolation and \"bilinear\" for bilinear 
          interpolation and \"linear\" for linear interpolation..

          Nearest neighbor interpolation is to perform nearest neighbor interpolation
          in both the 3rd dimension(in height direction) and the 4th dimension(in width 
          direction) on input tensor.
           
          Linear interpolation is the method of using a line connecting two known quantities 
          to determine the value of an unknown quantity between the two known quantities. 
          
          Bilinear interpolation is an extension of linear interpolation for 
          interpolating functions of two variables (e.g. H-direction and 
          W-direction in this op) on a rectilinear 2D grid. The key idea is 
          to perform linear interpolation first in one direction, and then 
          again in the other direction.

          Trilinear interpolation is an extension of linear interpolation for 
          interpolating functions of three variables (e.g. D-direction, 
          H-direction and W-direction in this op) on a rectilinear 3D grid. 
          The linear interpolation is performed on three directions.

          Bicubic interpolation is an extension of cubic interpolation for interpolating
          data points on a two-dimensional regular grid. The interpolated surface is
          smoother than corresponding surfaces obtained by bilinear interpolation or
          nearest-neighbor interpolation.

          Align_corners and align_mode are optional parameters,the calculation method 
          of interpolation can be selected by them.
          
          Example:

          For scale:
          
            if align_corners = True and out_{size}>1 :

              scale_{factor} = (in_{size}-1.0)/(out_{size}-1.0)
            
            else:
              
              scale_{factor} = float(in_{size}/out_{size})
            
          
          Nearest neighbor interpolation:
          
          if:
              align_corners = False

              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:

              H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor
              W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor

          else:
              align_corners = True

              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:

              H_out = round(H_{in} * scale_{factor})
              W_out = round(W_{in} * scale_{factor})

          Bilinear interpolation:

          if:
              align_corners = False , align_mode = 0
              
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5


          else:
           
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:

              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}

          Trilinear interpolation:

          if:
              align_corners = False , align_mode = 0
              
              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:
              
              D_out = (D_{in}+0.5) * scale_{factor} - 0.5
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5


          else:
           
              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:

              D_out = D_{in} * scale_{factor}
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}

          Bicubic interpolation:

          if:
              align_corners = False
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5
          else:
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}

          For details of nearest neighbor interpolation, please refer to Wikipedia: 
          https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation

          For details of bilinear interpolation, please refer to Wikipedia: 
          https://en.wikipedia.org/wiki/Bilinear_interp_v2olation

          For details of trilinear interpolation, please refer to Wikipedia: 
          https://en.wikipedia.org/wiki/Trilinear_interp_v2olation

          For details of bicubic interpolation, please refer to Wikipedia:
          https://en.wikipedia.org/wiki/Bicubic_interpolation
         )DOC");
  }
};

class InterpolateV2OpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(framework::InferShapeContext* ctx) const override {
615 616 617 618
    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "InterpolateGrad");
    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
                   "Out@GRAD", "InterpolateGrad");

X
xiaoting 已提交
619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
    auto dim_x = ctx->GetInputDim("X");
    if (ctx->HasOutput(framework::GradVarName("X"))) {
      ctx->SetOutputDim(framework::GradVarName("X"), dim_x);
    }
  }

  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
                                       ctx, framework::GradVarName("Out")),
                                   ctx.GetPlace());
  }

  framework::OpKernelType GetKernelTypeForVar(
      const std::string& var_name, const Tensor& tensor,
      const framework::OpKernelType& expected_kernel_type) const override {
    if (var_name == "SizeTensor" || var_name == "Scale") {
      return expected_kernel_type;
    }
    return framework::OpKernelType(expected_kernel_type.data_type_,
                                   tensor.place(), tensor.layout());
  }
};

template <typename T>
class InterpolateV2GradMaker : public framework::SingleGradOpMaker<T> {
 public:
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;

 protected:
  void Apply(GradOpPtr<T> op) const override {
    op->SetType(this->ForwardOpType() + "_grad");
    op->SetInput("X", this->Input("X"));
    if (this->HasInput("SizeTensor") > 0) {
      op->SetInput("SizeTensor", this->Input("SizeTensor"));
    }
    if (this->HasInput("OutSize") > 0) {
      op->SetInput("OutSize", this->Input("OutSize"));
    }
    if (this->HasInput("Scale") > 0) {
      op->SetInput("Scale", this->Input("Scale"));
    }
    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op->SetAttrMap(this->Attrs());
  }
};

DECLARE_NO_NEED_BUFFER_VARS_INFERER(InterpolateV2GradNoNeedBufferVarsInferer,
                                    "X");

}  // namespace operators
}  // namespace paddle

673 674 675
// interp_v2 support scale_factor whose input type is list, this operation is
// not
// compatible with interp_op, so a new one is added in paddle2.0
X
xiaoting 已提交
676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708
namespace ops = paddle::operators;
REGISTER_OPERATOR(bilinear_interp_v2, ops::InterpolateV2Op,
                  ops::InterpolateV2OpMaker,
                  ops::InterpolateV2GradMaker<paddle::framework::OpDesc>,
                  ops::InterpolateV2GradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(bilinear_interp_v2_grad, ops::InterpolateV2OpGrad,
                  ops::InterpolateV2GradNoNeedBufferVarsInferer);
REGISTER_OPERATOR(nearest_interp_v2, ops::InterpolateV2Op,
                  ops::InterpolateV2OpMaker,
                  ops::InterpolateV2GradMaker<paddle::framework::OpDesc>,
                  ops::InterpolateV2GradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(nearest_interp_v2_grad, ops::InterpolateV2OpGrad,
                  ops::InterpolateV2GradNoNeedBufferVarsInferer);
REGISTER_OPERATOR(trilinear_interp_v2, ops::InterpolateV2Op,
                  ops::InterpolateV2OpMaker,
                  ops::InterpolateV2GradMaker<paddle::framework::OpDesc>,
                  ops::InterpolateV2GradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(trilinear_interp_v2_grad, ops::InterpolateV2OpGrad,
                  ops::InterpolateV2GradNoNeedBufferVarsInferer);
REGISTER_OPERATOR(bicubic_interp_v2, ops::InterpolateV2Op,
                  ops::InterpolateV2OpMaker,
                  ops::InterpolateV2GradMaker<paddle::framework::OpDesc>,
                  ops::InterpolateV2GradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(bicubic_interp_v2_grad, ops::InterpolateV2OpGrad,
                  ops::InterpolateV2GradNoNeedBufferVarsInferer);
REGISTER_OP_CPU_KERNEL(bilinear_interp_v2, ops::InterpolateV2Kernel<float>,
                       ops::InterpolateV2Kernel<double>,
                       ops::InterpolateV2Kernel<uint8_t>);
REGISTER_OP_CPU_KERNEL(bilinear_interp_v2_grad,
                       ops::InterpolateV2GradKernel<float>,
                       ops::InterpolateV2GradKernel<double>);
REGISTER_OP_CPU_KERNEL(nearest_interp_v2, ops::InterpolateV2Kernel<float>,
                       ops::InterpolateV2Kernel<double>,
709 710
                       ops::InterpolateV2Kernel<int>,
                       ops::InterpolateV2Kernel<int64_t>,
X
xiaoting 已提交
711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737
                       ops::InterpolateV2Kernel<uint8_t>);
REGISTER_OP_CPU_KERNEL(nearest_interp_v2_grad,
                       ops::InterpolateV2GradKernel<float>,
                       ops::InterpolateV2GradKernel<double>);
REGISTER_OP_CPU_KERNEL(trilinear_interp_v2, ops::InterpolateV2Kernel<float>,
                       ops::InterpolateV2Kernel<double>,
                       ops::InterpolateV2Kernel<uint8_t>);
REGISTER_OP_CPU_KERNEL(trilinear_interp_v2_grad,
                       ops::InterpolateV2GradKernel<float>,
                       ops::InterpolateV2GradKernel<double>);
REGISTER_OPERATOR(linear_interp_v2, ops::InterpolateV2Op,
                  ops::InterpolateV2OpMaker,
                  ops::InterpolateV2GradMaker<paddle::framework::OpDesc>,
                  ops::InterpolateV2GradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(linear_interp_v2_grad, ops::InterpolateV2OpGrad,
                  ops::InterpolateV2GradNoNeedBufferVarsInferer);
REGISTER_OP_CPU_KERNEL(linear_interp_v2, ops::InterpolateV2Kernel<float>,
                       ops::InterpolateV2Kernel<double>,
                       ops::InterpolateV2Kernel<uint8_t>);
REGISTER_OP_CPU_KERNEL(linear_interp_v2_grad,
                       ops::InterpolateV2GradKernel<float>,
                       ops::InterpolateV2GradKernel<double>);
REGISTER_OP_CPU_KERNEL(bicubic_interp_v2, ops::InterpolateV2Kernel<float>,
                       ops::InterpolateV2Kernel<double>);
REGISTER_OP_CPU_KERNEL(bicubic_interp_v2_grad,
                       ops::InterpolateV2GradKernel<float>,
                       ops::InterpolateV2GradKernel<double>);