activation_op.h 41.9 KB
Newer Older
1
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
2

3 4 5 6 7 8 9 10 11
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. */
Q
qijun 已提交
12 13

#pragma once
D
dzhwinter 已提交
14
#include <glog/logging.h>
15
#include <algorithm>
16
#include <memory>
D
dzhwinter 已提交
17 18
#include <string>
#include <unordered_set>
19 20
#include <utility>
#include <vector>
21

22 23 24 25 26
#include <cmath>
#ifndef _USE_MATH_DEFINES
#define _USE_MATH_DEFINES
#endif

27
#include <type_traits>
Y
Yi Wang 已提交
28 29
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
30
#include "paddle/fluid/framework/tensor_util.h"
31
#include "paddle/fluid/platform/enforce.h"
32
#include "paddle/fluid/platform/float16.h"
33
#include "paddle/phi/kernels/funcs/blas/blas.h"
34 35 36 37
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif

38 39
#include "paddle/phi/kernels/funcs/activation_functor.h"

Q
qijun 已提交
40 41 42
namespace paddle {
namespace operators {

43 44
using framework::To32BitIndex;

45
using ActBwdOpFwdDeps = phi::funcs::ActBwdOpFwdDeps;
46

C
chengduo 已提交
47 48 49 50 51 52
/* The following operator can be used to process SelectedRows, because the
 * output of those operator for zero is zero too.
 */
static std::unordered_set<std::string> CanBeUsedBySelectedRows = {
    "abs", "abs_grad", "square", "square_grad", "sqrt", "sqrt_grad"};

53 54 55 56 57
inline void ExtractActivationTensor(const framework::ExecutionContext& context,
                                    const framework::Tensor** X,
                                    framework::Tensor** Out) {
  auto x_var = context.InputVar("X");
  auto out_var = context.OutputVar("Out");
58 59 60 61 62 63 64 65
  PADDLE_ENFORCE_NOT_NULL(x_var,
                          platform::errors::NotFound(
                              "Cannot get input Variable X, variable name = %s",
                              context.InputName("X")));
  PADDLE_ENFORCE_NOT_NULL(
      out_var, platform::errors::NotFound(
                   "Cannot get output Variable Out, variable name = %s",
                   context.OutputName("Out")));
66
  if (CanBeUsedBySelectedRows.count(context.Type())) {
67 68 69 70 71 72 73 74
    *X = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var);
    *Out = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
        out_var);
  } else {
    *X = context.Input<framework::Tensor>("X");
    *Out = context.Output<framework::Tensor>("Out");
  }

75 76 77 78
  PADDLE_ENFORCE_NOT_NULL(*Out, platform::errors::NotFound(
                                    "Cannot get the tensor from the Variable "
                                    "Output(Out), variable name = %s",
                                    context.OutputName("Out")));
79 80
}

81
template <ActBwdOpFwdDeps kDepValue>
82 83 84 85 86 87
inline void ExtractActivationGradTensor(
    const framework::ExecutionContext& context, const framework::Tensor** X,
    const framework::Tensor** Out, const framework::Tensor** dOut,
    framework::Tensor** dX) {
  auto out_grad_var = context.InputVar(framework::GradVarName("Out"));
  auto x_grad_var = context.OutputVar(framework::GradVarName("X"));
88 89
  const framework::Variable* out_var = nullptr;

90 91
  if (static_cast<int>(kDepValue) &
      static_cast<int>(ActBwdOpFwdDeps::kDepOut)) {
92
    out_var = context.InputVar("Out");
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
    PADDLE_ENFORCE_NOT_NULL(
        out_var, platform::errors::NotFound(
                     "Cannot get input Variable Out, variable name = %s",
                     context.InputName("Out")));
  }

  PADDLE_ENFORCE_NOT_NULL(
      out_grad_var, platform::errors::NotFound(
                        "Cannot get input Variable %s, variable name = %s",
                        framework::GradVarName("Out"),
                        context.InputName(framework::GradVarName("Out"))));
  PADDLE_ENFORCE_NOT_NULL(
      x_grad_var, platform::errors::NotFound(
                      "Cannot get output Variable %s, variable name = %s",
                      framework::GradVarName("X"),
                      context.OutputName(framework::GradVarName("X"))));
109

110
  if (CanBeUsedBySelectedRows.count(context.Type())) {
111 112 113 114
    *dOut = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(
        *out_grad_var);
    *dX = paddle::framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(
        x_grad_var);
115 116 117 118 119 120 121 122

    if (out_var) {
      *Out =
          paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*out_var);
    } else {
      *Out = *dOut;  // fake out
    }

123 124 125 126
  } else {
    *Out = context.Input<framework::Tensor>("Out");
    *dOut = context.Input<framework::Tensor>(framework::GradVarName("Out"));
    *dX = context.Output<framework::Tensor>(framework::GradVarName("X"));
127 128 129 130 131 132

    if (out_var) {
      *Out = &(out_var->Get<framework::LoDTensor>());
    } else {
      *Out = *dOut;  // fake out
    }
133
  }
134

135 136 137 138 139
  PADDLE_ENFORCE_NOT_NULL(*dX,
                          platform::errors::NotFound(
                              "Cannot get the tensor from the Variable "
                              "Output(Out), variable name = %s",
                              context.OutputName(framework::GradVarName("X"))));
140

141
  if (static_cast<int>(kDepValue) & static_cast<int>(ActBwdOpFwdDeps::kDepX)) {
C
chengduo 已提交
142
    auto x_var = context.InputVar("X");
143 144 145 146
    PADDLE_ENFORCE_NOT_NULL(x_var, platform::errors::NotFound(
                                       "Cannot get the tensor from the "
                                       "Variable Input(X), variable name = %s",
                                       context.InputName("X")));
147
    if (CanBeUsedBySelectedRows.count(context.Type())) {
148
      *X = paddle::framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_var);
C
chengduo 已提交
149
    } else {
150
      *X = context.Input<framework::Tensor>("X");
C
chengduo 已提交
151
    }
152
  } else {
153
    VLOG(10) << " Inplace activation of Op : " << context.Type();
154 155 156
    *X = *dX;
  }
}
C
chengduo 已提交
157

158 159 160 161 162
template <typename DeviceContext, typename Functor>
class ActivationKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
C
chengduo 已提交
163

164 165 166 167
  void Compute(const framework::ExecutionContext& context) const override {
    const framework::Tensor* X = nullptr;
    framework::Tensor* Out = nullptr;
    ExtractActivationTensor(context, &X, &Out);
C
chengduo 已提交
168
    Out->mutable_data<T>(context.GetPlace());
169

170 171 172 173
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "Activation"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Output", "Out", "Activation"));
Q
QI JUN 已提交
174 175
    auto* place =
        context.template device_context<DeviceContext>().eigen_device();
Q
qijun 已提交
176
    Functor functor;
177 178 179 180 181

    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = context.Attr<float>(attr.first);
    }
182 183 184 185 186 187 188 189
    // use 32bit index to speed up computation
    bool use_32bit_index = out.size() < Eigen::NumTraits<int>::highest();
    bool is_gpu_place = platform::is_gpu_place(context.GetPlace());
    if (use_32bit_index && is_gpu_place) {
      functor(*place, To32BitIndex(x), To32BitIndex(out));
    } else {
      functor(*place, x, out);
    }
Q
qijun 已提交
190 191 192
  }
};

Q
QI JUN 已提交
193
template <typename DeviceContext, typename Functor>
194 195
class ActivationGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
Q
qijun 已提交
196
 public:
197
  using T = typename Functor::ELEMENT_TYPE;
Q
qijun 已提交
198
  void Compute(const framework::ExecutionContext& context) const override {
199 200 201
    const framework::Tensor *X, *Out, *dOut;
    framework::Tensor* dX = nullptr;
    X = Out = dOut = nullptr;
202 203
    ExtractActivationGradTensor<Functor::FwdDeps()>(context, &X, &Out, &dOut,
                                                    &dX);
Q
qijun 已提交
204
    dX->mutable_data<T>(context.GetPlace());
205 206 207 208 209 210 211 212
    auto dout = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dOut, "Input", "Out@GRAD", "ActivationGrad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Input", "Out", "ActivationGrad"));
    auto dx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dX, "Input", "X@GRAD", "ActivationGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "ActivationGrad"));
Q
QI JUN 已提交
213 214
    auto* place =
        context.template device_context<DeviceContext>().eigen_device();
Q
qijun 已提交
215
    Functor functor;
216 217 218 219
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = context.Attr<float>(attr.first);
    }
220 221 222 223 224 225 226 227 228
    // use 32bit index to speed up computation
    bool use_32bit_index = out.size() < Eigen::NumTraits<int>::highest();
    bool is_gpu_place = platform::is_gpu_place(context.GetPlace());
    if (use_32bit_index && is_gpu_place) {
      functor(*place, To32BitIndex(x), To32BitIndex(out), To32BitIndex(dout),
              To32BitIndex(dx));
    } else {
      functor(*place, x, out, dout, dx);
    }
Q
qijun 已提交
229 230 231
  }
};

232 233 234 235 236 237 238 239 240
template <typename T>
struct BaseActivationFunctor {
  using ELEMENT_TYPE = T;

  using AttrPair = std::vector<std::pair<const char*, float*>>;

  AttrPair GetAttrs() { return AttrPair(); }
};

241 242 243 244 245 246
#define USE_PHI_FUNCTOR(name)                         \
  template <typename T>                               \
  using name##Functor = phi::funcs::name##Functor<T>; \
  template <typename T>                               \
  using name##GradFunctor = phi::funcs::name##GradFunctor<T>;

247 248 249 250 251 252 253 254
#define USE_PHI_DOUBLE_GRAD_FUNCTOR(name) \
  template <typename T>                   \
  using name##GradGradFunctor = phi::funcs::name##GradGradFunctor<T>;

#define USE_PHI_TRIPLE_GRAD_FUNCTOR(name) \
  template <typename T>                   \
  using name##TripleGradFunctor = phi::funcs::name##TripleGradFunctor<T>;

255 256 257 258 259 260 261 262 263 264 265
USE_PHI_FUNCTOR(Cos)
USE_PHI_FUNCTOR(Tan)
USE_PHI_FUNCTOR(Acos)
USE_PHI_FUNCTOR(Sin)
USE_PHI_FUNCTOR(Asin)
USE_PHI_FUNCTOR(Atan)
USE_PHI_FUNCTOR(Sinh)
USE_PHI_FUNCTOR(Cosh)
USE_PHI_FUNCTOR(Asinh)
USE_PHI_FUNCTOR(Acosh)
USE_PHI_FUNCTOR(Atanh)
266 267 268 269 270 271 272
USE_PHI_FUNCTOR(Tanh)
USE_PHI_DOUBLE_GRAD_FUNCTOR(Tanh)
USE_PHI_TRIPLE_GRAD_FUNCTOR(Tanh)
USE_PHI_FUNCTOR(BRelu)
USE_PHI_FUNCTOR(ThresholdedRelu)
USE_PHI_FUNCTOR(LeakyRelu)
USE_PHI_DOUBLE_GRAD_FUNCTOR(LeakyRelu)
273 274 275 276 277 278
USE_PHI_FUNCTOR(HardShrink)
USE_PHI_FUNCTOR(SoftShrink)
USE_PHI_FUNCTOR(TanhShrink)
USE_PHI_FUNCTOR(Silu)
USE_PHI_FUNCTOR(ELU)
USE_PHI_DOUBLE_GRAD_FUNCTOR(ELU)
279 280 281 282 283
USE_PHI_FUNCTOR(Sigmoid)
USE_PHI_DOUBLE_GRAD_FUNCTOR(Sigmoid)
USE_PHI_TRIPLE_GRAD_FUNCTOR(Sigmoid)
USE_PHI_FUNCTOR(LogSigmoid)
USE_PHI_FUNCTOR(HardSigmoid)
284 285 286 287 288
USE_PHI_FUNCTOR(Log)
USE_PHI_DOUBLE_GRAD_FUNCTOR(Log)
USE_PHI_FUNCTOR(Log2)
USE_PHI_FUNCTOR(Log10)
USE_PHI_FUNCTOR(Log1p)
Y
YuanRisheng 已提交
289 290 291
USE_PHI_FUNCTOR(Swish)
USE_PHI_FUNCTOR(HardSwish)
USE_PHI_FUNCTOR(Pow)
292 293 294

template <typename T>
using ELUGradNegativeAlphaFunctor = phi::funcs::ELUGradNegativeAlphaFunctor<T>;
295

Y
YuanRisheng 已提交
296 297 298 299 300 301 302 303 304 305 306 307
template <typename T>
using RoundFunctor = phi::funcs::RoundFunctor<T>;

template <typename T>
using FloorFunctor = phi::funcs::FloorFunctor<T>;

template <typename T>
using CeilFunctor = phi::funcs::CeilFunctor<T>;

template <typename T>
using ZeroGradFunctor = phi::funcs::ZeroGradFunctor<T>;

308
// exp(x) = e^x
309
template <typename T>
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
struct ExpFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.exp();
  }
};

template <typename T>
struct ExpGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * out;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

// expm1(x) = e^x - 1
template <typename T>
struct Expm1Functor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.expm1();
  }
};

template <typename T>
struct Expm1GradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * out + dout;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};
R
ronnywang 已提交
351

Q
qijun 已提交
352
// relu(x) = max(x, 0)
353 354

template <typename T>
355 356 357
using ReluCPUFunctor = phi::funcs::ReluCPUFunctor<T>;
template <typename T>
using ReluGradFunctor = phi::funcs::ReluGradFunctor<T>;
Q
qijun 已提交
358

359
template <typename T>
360
using ReluGradGradFunctor = phi::funcs::ReluGradGradFunctor<T>;
361

362 363
template <typename T>
using ReluCUDAFunctor = phi::funcs::ReluCUDAFunctor<T>;
Q
qijun 已提交
364

365
// sqrt(x) = x^(1/2)
Z
zhoukunsheng 已提交
366
template <typename T>
367 368 369 370
struct SqrtFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.sqrt();
Z
zhoukunsheng 已提交
371
  }
372 373 374 375 376 377 378 379 380 381
};

template <typename T>
struct SqrtGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = static_cast<T>(0.5) * dout / out;
  }

382 383 384
  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
Z
zhoukunsheng 已提交
385 386
};

387
// rsqrt(x) = x^(-1/2)
Q
qijun 已提交
388
template <typename T>
389 390 391 392 393 394
struct RsqrtFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.rsqrt();
  }
};
Q
qijun 已提交
395

396 397 398 399 400 401
template <typename T>
struct RsqrtGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = static_cast<T>(-0.5) * dout * out * out * out;
Q
qijun 已提交
402
  }
403

404 405 406
  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
Q
qijun 已提交
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
// reciprocal(x) = 1 / x
template <typename T>
struct ReciprocalFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = static_cast<T>(1) / x;
  }
};

template <typename T>
struct ReciprocalGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * static_cast<T>(-1) * out * out;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

// square(x) = x^2
template <typename T>
struct SquareFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) = x.square();
  }
};

template <typename T>
struct SquareGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    dx.device(d) = dout * static_cast<T>(2) * x;
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

451 452 453 454 455 456 457 458 459
// relu6(x) = min(max(0, x), 6)
template <typename T>
struct Relu6Functor : public BaseActivationFunctor<T> {
  float threshold;

  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }

F
fengjiayi 已提交
460 461 462
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) =
463
        x.cwiseMax(static_cast<T>(0)).cwiseMin(static_cast<T>(threshold));
464 465 466 467 468 469 470 471 472
  }
};

template <typename T>
struct Relu6GradFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
F
fengjiayi 已提交
473 474 475
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
D
dzhwinter 已提交
476 477 478 479
    dx.device(d) =
        dout *
        ((out > static_cast<T>(0)) * (out < static_cast<T>(threshold)))
            .template cast<T>();
480
  }
481

482 483 484
  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
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
// For numerical stability, using the following formula instead of softplus(x) =
// log(1 + exp(x))
// softplus(x) = log(1 + exp(beta * x)) / beta when beta * x <= threshold(beta =
// 1, threshold = 20 by default), otherwise x
template <typename T>
struct SoftplusFunctor : public BaseActivationFunctor<T> {
  float beta;
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"beta", &beta}, {"threshold", &threshold}};
  }

  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) {
    auto x_beta = static_cast<T>(beta) * x;
    out.device(d) = (x_beta > static_cast<T>(threshold))
                        .select(x, (static_cast<T>(1) + x_beta.exp()).log() /
                                       static_cast<T>(beta));
  }
};

// For numerical stability, using the following formula instead of
// d(softplus(x))/dx = 1 / (1 + exp(-x))
// d(softplus(x))/dx = 1 / (1 + exp(-beta * x)) when beta * x <= threshold(beta
// = 1, threshold = 20 by default), otherwise x
template <typename T>
struct SoftplusGradFunctor : public BaseActivationFunctor<T> {
  float beta;
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"beta", &beta}, {"threshold", &threshold}};
  }

  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) {
    auto x_beta = static_cast<T>(beta) * x;
    dx.device(d) =
        (x_beta > static_cast<T>(threshold))
            .select(dout, dout / (static_cast<T>(1) + (-x_beta).exp()));
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

// mish(x) = x * tanh(softplus(x))
// softplus(x) = x, if x > threshold
//             = ln(1 + exp(x)), otherwise
template <typename T>
struct MishFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }

  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) {
    auto sp = (x > static_cast<T>(threshold))
                  .select(x, (static_cast<T>(1) + x.exp()).log());
    out.device(d) = x * sp.tanh();
  }
};

// dx = dout * (tanh(sp) + x * (1 - tanh(sp) ** 2) * (1 - exp(-sp)))
// sp = softplus(x)
template <typename T>
struct MishGradFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }

  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) {
    auto sp = (x > static_cast<T>(threshold))
                  .select(x, (static_cast<T>(1) + x.exp()).log());
    auto gsp = static_cast<T>(1) - (-sp).exp();
    auto tsp = sp.tanh();
    dx.device(d) = dout * (tsp + x * (static_cast<T>(1) - tsp * tsp) * gsp);
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

// softsign(x) = x / (1 + |x|)
template <typename T>
struct SoftsignFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) {
    out.device(d) = x / (static_cast<T>(1) + x.abs());
  }
};

// d(softsign(x))/dx = 1 / (1 + |x|)^2
// Taken from https://en.wikipedia.org/wiki/Activation_function
template <typename T>
struct SoftsignGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) {
    dx.device(d) =
        dout * (static_cast<T>(1) / (static_cast<T>(1) + x.abs()).square());
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

595 596 597 598 599 600
template <typename T>
struct SoftReluFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
601

F
fengjiayi 已提交
602 603
  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
604 605
    auto tmp = static_cast<T>(threshold);
    auto temp = x.cwiseMax(-tmp).cwiseMin(tmp);
F
fengjiayi 已提交
606
    out.device(d) = (static_cast<T>(1) + temp.exp()).log();
607 608 609
  }
};

610 611 612 613 614 615
template <typename T>
struct SoftReluGradFunctor : public BaseActivationFunctor<T> {
  float threshold;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"threshold", &threshold}};
  }
F
fengjiayi 已提交
616 617 618
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
619
    auto tmp = static_cast<T>(threshold);
620
    auto temp = ((out > -tmp) * (out < tmp)).template cast<T>();
F
fengjiayi 已提交
621
    dx.device(d) = dout * (static_cast<T>(1) - (-out).exp()) * temp;
622
  }
623

624 625 626
  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
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
template <typename DeviceContext, typename T>
class ELUGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* X = context.Input<framework::Tensor>("X");
    auto* Out = context.Input<framework::Tensor>("Out");
    auto* dOut =
        context.Input<framework::Tensor>(framework::GradVarName("Out"));
    auto* dX = context.Output<framework::Tensor>(framework::GradVarName("X"));
    const float alpha = context.Attr<float>("alpha");
    dX->mutable_data<T>(context.GetPlace());

    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "elu_grad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Input", "Out", "elu_grad"));
    auto dout = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dOut, "Input", "dOut", "elu_grad"));
    auto dx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(dX, "Output", "dX", "elu_grad"));
    auto* place =
        context.template device_context<DeviceContext>().eigen_device();

    if (alpha > 0) {
      ELUGradFunctor<T> functor;
      functor.alpha = alpha;
      functor(*place, x, out, dout, dx);
    } else {
      ELUGradNegativeAlphaFunctor<T> functor;
      functor.alpha = alpha;
      functor(*place, x, out, dout, dx);
    }
  }
};

664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705
template <typename T>
struct CELUFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }

  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) =
        (x < static_cast<T>(0))
            .select(static_cast<T>(alpha) *
                        ((x / static_cast<T>(alpha)).exp() - static_cast<T>(1)),
                    x);
  }
};

template <typename T>
struct CELUGradFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    auto temp_a_pos = static_cast<T>(alpha > 0);
    auto temp_a_neg = static_cast<T>(alpha <= 0);
    auto temp_x_pos = (x > static_cast<T>(0)).template cast<T>();
    auto temp_x_neg = (x <= static_cast<T>(0)).template cast<T>();

    // dx = dout, if alpha > 0 and x > 0
    // dx = dout * (x/alpha).exp(), if alpha > 0 and x <= 0
    // dx = dout , if alpha < 0 and x > 0
    // dx = dout * (x/alpha).exp(), if alpha < 0 and x <=0
    dx.device(d) =
        dout * temp_a_pos * temp_x_pos +
        dout * (x / static_cast<T>(alpha)).exp() * temp_a_pos * temp_x_neg +
        dout * temp_a_neg * temp_x_pos +
        dout * (x / static_cast<T>(alpha)).exp() * temp_a_neg * temp_x_neg;
  }

706
  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
707 708
};

709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
template <typename T>
struct LogitFunctor {
  template <typename Device, typename X, typename Out, typename P>
  void operator()(Device d, X x, Out out, P p, float eps) const {
    // logit(x) = ln(x/(1-x))
    auto tmp_x =
        (x.cwiseMin(static_cast<T>(1.0 - eps))).cwiseMax(static_cast<T>(eps));

    if (!eps) {
      out.device(d) = (x < static_cast<T>(0.0) || x > static_cast<T>(1.0))
                          .select(p.constant(static_cast<T>(NAN)),
                                  (tmp_x / (static_cast<T>(1) - tmp_x)).log());
    } else {
      out.device(d) = (tmp_x / (static_cast<T>(1) - tmp_x)).log();
    }
  }
};

template <typename T>
struct LogitGradFunctor {
  template <typename Device, typename X, typename dOut, typename dX, typename P>
  void operator()(Device d, X x, dOut dout, dX dx, P p, float eps) const {
    // logit(x)' = 1/(x*(1-x))
    dx.device(d) =
        (x < static_cast<T>(eps) || x > static_cast<T>(1.0 - eps))
            .select(p.constant(static_cast<T>(0)),
                    dout * (static_cast<T>(1) / ((static_cast<T>(1) - x) * x)));
  }
};

template <typename T>
struct STanhFunctor : public BaseActivationFunctor<T> {
  float scale_a;
  float scale_b;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"scale_a", &scale_a}, {"scale_b", &scale_b}};
  }

  template <typename Device, typename X, typename Out>
  void operator()(Device d, X x, Out out) const {
    out.device(d) =
        static_cast<T>(scale_b) * (static_cast<T>(scale_a) * x).tanh();
  }
};

template <typename T>
struct STanhGradFunctor : public BaseActivationFunctor<T> {
  float scale_a;
  float scale_b;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"scale_a", &scale_a}, {"scale_b", &scale_b}};
  }

  template <typename Device, typename X, typename Out, typename dOut,
            typename dX>
  void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
    auto a = static_cast<T>(scale_a);
    auto b = static_cast<T>(scale_b);
    auto temp = (a * x).tanh() * (a * x).tanh();
    dx.device(d) = dout * a * b * (static_cast<T>(1) - temp);
  }

  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
};

Z
Zhong Hui 已提交
774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791
template <typename T>
struct AbsGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* Out, const framework::Tensor* ddX,
                  framework::Tensor* ddOut, framework::Tensor* dOut,
                  framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "AbsGradGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "AbsGradGrad"));
    if (ddOut) {
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "AbsGradGrad"));
      ddout.device(*d) = ddx * x.sign();
    }
  }
792
  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
793 794
};

795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830
template <typename T>
struct CELUGradGradFunctor : public BaseActivationFunctor<T> {
  float alpha;
  typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
    return {{"alpha", &alpha}};
  }
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* ddX, framework::Tensor* ddOut,
                  const framework::Tensor* dOut, framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "CELUGradGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "CELUGradGrad"));

    if (dX) {
      auto dx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dX, "Output", "DX", "CELUGradGrad"));
      auto dout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dOut, "Output", "DOut", "CELUGradGrad"));
      dx.device(*d) = ddx * dout / static_cast<T>(alpha) *
                      (x / static_cast<T>(alpha)).exp() *
                      (x <= static_cast<T>(0)).template cast<T>();
    }

    if (ddOut) {
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "CELUGradGrad"));
      ddout.device(*d) = ddx *
                         ((x > static_cast<T>(0)).template cast<T>() +
                          (x / static_cast<T>(alpha)).exp() *
                              (x <= static_cast<T>(0)).template cast<T>())
                             .template cast<T>();
    }
  }
831
  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
832 833
};

834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895
template <typename T>
struct SqrtGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* Out,
                  const framework::Tensor* ddX, framework::Tensor* ddOut,
                  framework::Tensor* dOut, const framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "SqrtGradGrad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Output", "Out", "SqrtGradGrad"));
    // sqrt GradGrad: ddy = 0.5 * ddx / y, dy = -1 * dx * ddx
    // calculate dy first, so ddy can inplace ddx
    if (dOut) {
      auto dx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dX, "Output", "DX", "SqrtGradGrad"));
      auto dout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dOut, "Output", "DOut", "SqrtGradGrad"));
      dout.device(*d) = dx * ddx * static_cast<T>(-1) / out;
    }
    if (ddOut) {
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "SqrtGradGrad"));
      ddout.device(*d) = ddx * static_cast<T>(0.5) / out;
    }
  }
  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

template <typename T>
struct RsqrtGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* Out,
                  const framework::Tensor* ddX, framework::Tensor* ddOut,
                  framework::Tensor* dOut, const framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "RsqrtGradGrad"));
    auto out = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(Out, "Output", "Out", "RsqrtGradGrad"));

    // rsqrt GradGrad: ddy = -0.5 * ddx * y * y * y, dy = (3/y) * dx * ddx
    if (dOut) {
      auto dx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dX, "Output", "DX", "RsqrtGradGrad"));
      auto dout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dOut, "Output", "DOut", "RsqrtGradGrad"));
      dout.device(*d) = (static_cast<T>(3.0) / out) * dx * ddx;
    }
    if (ddOut) {
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "RsqrtGradGrad"));
      ddout.device(*d) = ddx * static_cast<T>(-0.5) * out * out * out;
    }
  }
  static constexpr ActBwdOpFwdDeps FwdDeps() {
    return ActBwdOpFwdDeps::kDepOut;
  }
};

896 897 898 899 900 901 902
template <typename T>
struct SquareGradGradFunctor : public BaseActivationFunctor<T> {
  template <typename Device>
  void operator()(const Device& dev, const framework::Tensor* X,
                  const framework::Tensor* ddX, framework::Tensor* ddOut,
                  const framework::Tensor* dOut, framework::Tensor* dX) const {
    auto* d = dev.eigen_device();
903 904 905 906
    auto ddx = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(ddX, "Input", "DDX", "SquareGradGrad"));
    auto x = framework::EigenVector<T>::Flatten(
        GET_DATA_SAFELY(X, "Input", "X", "SquareGradGrad"));
907 908
    // square GradGrad: ddy=2x*ddx, dx=2dy*ddx
    // calculate dx first, so ddy can inplace ddx
909
    if (dX) {
910 911 912 913
      auto dx = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dX, "Output", "DX", "SquareGradGrad"));
      auto dout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(dOut, "Output", "DOut", "SquareGradGrad"));
914 915
      dx.device(*d) = ddx * static_cast<T>(2) * dout;
    }
916
    if (ddOut) {
917 918
      auto ddout = framework::EigenVector<T>::Flatten(
          GET_DATA_SAFELY(ddOut, "Output", "DDOut", "SquareGradGrad"));
919 920
      ddout.device(*d) = ddx * static_cast<T>(2) * x;
    }
921
  }
922
  static constexpr ActBwdOpFwdDeps FwdDeps() { return ActBwdOpFwdDeps::kDepX; }
923 924 925 926 927 928 929 930 931 932 933 934
};

// TODO(dengkaipeng): double gradient calculation for Square/Sqrt need
// DOut(dy) as input(not output), tensor extraction is different from
// others. Impliment extraction kernel seperately here.
inline void ExtractDoubleGradTensorWithInputDOut(
    const framework::ExecutionContext& ctx, const framework::Tensor** X,
    const framework::Tensor** ddX, framework::Tensor** dX,
    const framework::Tensor** dOut, framework::Tensor** ddOut) {
  // extract ddX(output), ddOut(input)
  auto ddx_var = ctx.InputVar("DDX");
  auto ddo_var = ctx.OutputVar("DDOut");
935 936 937 938
  PADDLE_ENFORCE_NOT_NULL(
      ddx_var, platform::errors::NotFound(
                   "Cannot get input Variable Out, variable name = %s",
                   ctx.InputName("DDX")));
939 940 941 942
  *ddX = ctx.Input<framework::Tensor>("DDX");
  if (ddo_var) {
    *ddOut = ctx.Output<framework::Tensor>("DDOut");
  }
943 944 945 946 947
  PADDLE_ENFORCE_NOT_NULL(
      ddX,
      platform::errors::NotFound(
          "Cannot get the tensor from the Variable DDX, variable name = %s",
          ctx.OutputName("DDX")));
948 949 950

  // extract x(input), dx(output)
  auto x_var = ctx.InputVar("X");
951 952
  PADDLE_ENFORCE_NOT_NULL(
      x_var, platform::errors::NotFound(
953
                 "Cannot get input Variable Out, variable name = %s",
954
                 ctx.InputName("X")));
955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980
  auto dx_var = ctx.OutputVar("DX");
  *X = ctx.Input<framework::Tensor>("X");
  if (dx_var) {
    *dX = ctx.Output<framework::Tensor>("DX");
  }

  // extract dOut(input)
  auto dout_var = ctx.InputVar("DOut");
  if (dout_var) {
    *dOut = ctx.Input<framework::Tensor>("DOut");
  }
}

template <typename DeviceContext, typename Functor>
class SquareDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *X, *ddX, *dOut;
    X = ddX = dOut = nullptr;
    framework::Tensor *dX, *ddOut;
    dX = ddOut = nullptr;

    ExtractDoubleGradTensorWithInputDOut(ctx, &X, &ddX, &dX, &dOut, &ddOut);

981 982
    if (dX) dX->mutable_data<T>(X->dims(), ctx.GetPlace());
    if (ddOut) ddOut->mutable_data<T>(ctx.GetPlace());
983 984 985 986 987 988 989 990

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    functor(place, X, ddX, ddOut, dOut, dX);
  }
};

991
template <typename DeviceContext, typename Functor>
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
class CELUDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *X, *ddX, *dOut;
    X = ddX = dOut = nullptr;
    framework::Tensor *dX, *ddOut;
    dX = ddOut = nullptr;

    ExtractDoubleGradTensorWithInputDOut(ctx, &X, &ddX, &dX, &dOut, &ddOut);

    if (dX) dX->mutable_data<T>(X->dims(), ctx.GetPlace());
    if (ddOut) ddOut->mutable_data<T>(ctx.GetPlace());

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    auto attrs = functor.GetAttrs();
    for (auto& attr : attrs) {
      *attr.second = ctx.Attr<float>(attr.first);
    }
    functor(place, X, ddX, ddOut, dOut, dX);
  }
};

template <typename DeviceContext, typename Functor>
1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031
class SqrtDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *Out, *dX, *ddX;
    Out = dX = ddX = nullptr;
    framework::Tensor *ddOut, *dOut;
    ddOut = dOut = nullptr;

    // extract ddx(input), ddout(output)
    auto ddx_var = ctx.InputVar("DDX");
    auto ddo_var = ctx.OutputVar("DDOut");
1032 1033 1034 1035
    PADDLE_ENFORCE_NOT_NULL(
        ddx_var, platform::errors::NotFound(
                     "Cannot get input Variable DDX, variable name = %s",
                     ctx.InputName("DDX")));
1036 1037 1038 1039
    ddX = ctx.Input<framework::Tensor>("DDX");
    if (ddo_var) {
      ddOut = ctx.Output<framework::Tensor>("DDOut");
    }
1040 1041 1042 1043
    PADDLE_ENFORCE_NOT_NULL(
        ddX, platform::errors::NotFound(
                 "Cannot get input Variable DDX, variable name = %s",
                 ctx.InputName("DDX")));
1044 1045 1046

    // extract out(input), dout(output)
    auto out_var = ctx.InputVar("Out");
1047 1048 1049 1050
    PADDLE_ENFORCE_NOT_NULL(
        out_var, platform::errors::NotFound(
                     "Cannot get input Variable Out, variable name = %s",
                     ctx.InputName("Out")));
1051 1052 1053 1054 1055 1056 1057 1058
    auto dout_var = ctx.OutputVar("DOut");
    Out = ctx.Input<framework::Tensor>("Out");
    if (dout_var) {
      dOut = ctx.Output<framework::Tensor>("DOut");
    }

    // extract dx(input)
    auto dx_var = ctx.InputVar("DX");
1059 1060 1061 1062
    PADDLE_ENFORCE_NOT_NULL(
        dx_var, platform::errors::NotFound(
                    "Cannot get input Variable DX, variable name = %s",
                    ctx.InputName("DX")));
1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076
    if (dx_var) {
      dX = ctx.Input<framework::Tensor>("DX");
    }

    if (dOut) dOut->mutable_data<T>(Out->dims(), ctx.GetPlace());
    if (ddOut) ddOut->mutable_data<T>(Out->dims(), ctx.GetPlace());

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    functor(place, Out, ddX, ddOut, dOut, dX);
  }
};

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 1106 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
// rsqrt Grad: dx = -0.5 * dy * y * y * y
// rsqrt GradGrad: ddy = -0.5 * ddx * y * y * y, dy = (3 / y) * dx * ddx
template <typename DeviceContext, typename Functor>
class RsqrtDoubleGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  using T = typename Functor::ELEMENT_TYPE;
  void Compute(const framework::ExecutionContext& ctx) const override {
    const framework::Tensor *Out, *dX, *ddX;
    Out = dX = ddX = nullptr;
    framework::Tensor *ddOut, *dOut;
    ddOut = dOut = nullptr;

    // extract ddx(input), ddout(output)
    auto ddx_var = ctx.InputVar("DDX");
    auto ddo_var = ctx.OutputVar("DDOut");
    PADDLE_ENFORCE_NOT_NULL(
        ddx_var, platform::errors::NotFound(
                     "Cannot get input Variable DDX, variable name = %s",
                     ctx.InputName("DDX")));
    ddX = ctx.Input<framework::Tensor>("DDX");
    if (ddo_var) {
      ddOut = ctx.Output<framework::Tensor>("DDOut");
    }
    PADDLE_ENFORCE_NOT_NULL(
        ddX, platform::errors::NotFound(
                 "Cannot get input Variable DDX, variable name = %s",
                 ctx.InputName("DDX")));

    // extract out(input), dout(output)
    auto out_var = ctx.InputVar("Out");
    PADDLE_ENFORCE_NOT_NULL(
        out_var, platform::errors::NotFound(
                     "Cannot get input Variable Out, variable name = %s",
                     ctx.InputName("Out")));
    auto dout_var = ctx.OutputVar("DOut");
    Out = ctx.Input<framework::Tensor>("Out");
    if (dout_var) {
      dOut = ctx.Output<framework::Tensor>("DOut");
    }

    // extract dx(input)
    auto dx_var = ctx.InputVar("DX");
    PADDLE_ENFORCE_NOT_NULL(
        dx_var, platform::errors::NotFound(
                    "Cannot get input Variable DX, variable name = %s",
                    ctx.InputName("DX")));
    if (dx_var) {
      dX = ctx.Input<framework::Tensor>("DX");
    }

    if (dOut) dOut->mutable_data<T>(Out->dims(), ctx.GetPlace());
    if (ddOut) ddOut->mutable_data<T>(Out->dims(), ctx.GetPlace());

    auto& place = ctx.template device_context<DeviceContext>();

    Functor functor;
    functor(place, Out, ddX, ddOut, dOut, dX);
  }
};

1138 1139 1140 1141 1142 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
template <typename DeviceContext, typename T>
class LogitKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* out = context.Output<framework::Tensor>("Out");
    auto* in = context.Input<framework::Tensor>("X");
    auto eps = context.Attr<float>("eps");
    out->mutable_data<T>(in->place());

    auto eigen_out = framework::EigenVector<T>::Flatten(*out);
    auto eigen_in = framework::EigenVector<T>::Flatten(*in);
    auto& place =
        *context.template device_context<DeviceContext>().eigen_device();
    auto eigen_p = framework::EigenVector<T>::Flatten(*out);

    LogitFunctor<T> functor;
    functor(place, eigen_in, eigen_out, eigen_p, eps);
  }
};

template <typename DeviceContext, typename T>
class LogitGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* x = context.Input<framework::Tensor>("X");
    auto* dout =
        context.Input<framework::Tensor>(framework::GradVarName("Out"));
    auto* dx = context.Output<framework::Tensor>(framework::GradVarName("X"));
    auto eps = context.Attr<float>("eps");
    dx->mutable_data<T>(dout->place());

    auto eigen_x = framework::EigenVector<T>::Flatten(*x);
    auto eigen_dout = framework::EigenVector<T>::Flatten(*dout);
    auto eigen_dx = framework::EigenVector<T>::Flatten(*dx);
    auto& place =
        *context.template device_context<DeviceContext>().eigen_device();
    auto eigen_p = framework::EigenVector<T>::Flatten(*x);

    LogitGradFunctor<T> functor;
    functor(place, eigen_x, eigen_dout, eigen_dx, eigen_p, eps);
  }
};

Q
qijun 已提交
1181 1182
}  // namespace operators
}  // namespace paddle
1183

1184 1185 1186 1187 1188 1189 1190 1191
#define FOR_EACH_ACTIVATION_OP(__macro)                                      \
  __macro(reciprocal, Reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \
  __macro(soft_relu, SoftRelu, SoftReluFunctor, SoftReluGradFunctor);        \
  __macro(stanh, STanh, STanhFunctor, STanhGradFunctor);                     \
  __macro(softplus, Softplus, SoftplusFunctor, SoftplusGradFunctor);         \
  __macro(softsign, Softsign, SoftsignFunctor, SoftsignGradFunctor);         \
  __macro(relu6, Relu6, Relu6Functor, Relu6GradFunctor);                     \
  __macro(mish, Mish, MishFunctor, MishGradFunctor);
反馈
建议
客服 返回
顶部