386 lines
16 KiB
C++
386 lines
16 KiB
C++
/*******************************************************************************
|
|
* Copyright (c) 2015-2018 Skymind, Inc.
|
|
*
|
|
* This program and the accompanying materials are made available under the
|
|
* terms of the Apache License, Version 2.0 which is available at
|
|
* https://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.
|
|
*
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
******************************************************************************/
|
|
|
|
//
|
|
// @author GS <sgazeos@gmail.com>
|
|
//
|
|
|
|
#include <ops/declarable/helpers/legacy_helpers.h>
|
|
#include <NDArrayFactory.h>
|
|
#include <ops/ops.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
namespace helpers {
|
|
template <typename T>
|
|
static void reluDerivative__(NDArray* theFirst, NDArray* theSecond) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return x > (T) 0.f ? y : T(0.f);
|
|
};
|
|
|
|
theFirst->applyPairwiseLambda<T>(theSecond, functor, nullptr);
|
|
}
|
|
|
|
void reluDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), reluDerivative__, (theFirst, theSecond), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void reluDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
|
|
T zero = (T) 0.f;
|
|
auto functor = LAMBDA_TT(x, y, zero){
|
|
return x > zero ? y : zero;
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
|
|
/*
|
|
auto x = input->bufferAsT<T>();
|
|
auto y = epsilon->bufferAsT<T>();
|
|
auto z = output->bufferAsT<T>();
|
|
|
|
int length = input->lengthOf();
|
|
|
|
T zero = (T) 0.f;
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR
|
|
for (int e = 0; e < length; e++) {
|
|
z[e] = x[e] > zero ? y[e] : zero;
|
|
}
|
|
*/
|
|
}
|
|
|
|
void reluDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), reluDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void relu6Derivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return x > (T)0.f && x < (T)6.f? y : T(0.f);
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void relu6Derivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), relu6Derivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void leakyReluDerivative_(NDArray* input, NDArray* epsilon, NDArray* output, const float alpha) {
|
|
|
|
const T alphaT = static_cast<T>(alpha);
|
|
|
|
auto functor = LAMBDA_TT(x, y, alphaT) {
|
|
return x < 0 ? alphaT * y : y;
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void leakyReluDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput, const float alpha) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), leakyReluDerivative_, (theFirst, theSecond, theOutput, alpha), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void eluDerivative_(NDArray* input, NDArray* epsilon, NDArray* output, const float alpha) {
|
|
|
|
const T alphaT = static_cast<T>(alpha);
|
|
|
|
auto functor = LAMBDA_TT(x, y, alphaT){
|
|
return y * nd4j::math::nd4j_eluderivative<T,T>(x, alphaT);
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void eluDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput, const float alpha) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), eluDerivative_, (theFirst, theSecond, theOutput, alpha), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void seluDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return y * simdOps::SELUDerivative<T>::op(x, nullptr);
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void seluDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), seluDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void cubeDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return y * (3 * x * x);
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void cubeDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), cubeDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
//return (x >= X(0.f) ? y: -y);
|
|
template <typename T>
|
|
static void reduceNorm1_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return x > T(0.f)? y : -y;
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void reduceNorm1(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), reduceNorm1_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename T>
|
|
static void sigmCrossEntropy_(NDArray* logits, NDArray* labels, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return nd4j::math::nd4j_max<T>(x, (T)0.f) - x * y + nd4j::math::nd4j_log<T,T>((T)1.f + nd4j::math::nd4j_exp<T,T>(-nd4j::math::nd4j_abs(x)));
|
|
};
|
|
|
|
logits->applyPairwiseLambda<T>(labels, functor, output);
|
|
}
|
|
|
|
void sigmCrossEntropy(nd4j::LaunchContext * context, NDArray* logits, NDArray* labels, NDArray* output) {
|
|
BUILD_SINGLE_SELECTOR(logits->dataType(), sigmCrossEntropy_, (logits, labels, output), FLOAT_TYPES);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename T>
|
|
static void sigmCrossEntropyGrad_(NDArray* logits, NDArray* labels, NDArray* output) {
|
|
// 1 - labels - 1 / (1 + exp(logits))
|
|
auto functor = LAMBDA_TT(x, y) {
|
|
if(x <= 0)
|
|
return static_cast<T>(1.) - y - static_cast<T>(1.) / (static_cast<T>(1.) + nd4j::math::nd4j_exp<T,T>(x));
|
|
auto e = nd4j::math::nd4j_exp<T,T>(-x);
|
|
return static_cast<T>(1.) - y - e / (static_cast<T>(1.) + e);
|
|
};
|
|
|
|
logits->applyPairwiseLambda<T>(labels, functor, output);
|
|
}
|
|
|
|
void sigmCrossEntropyGrad(nd4j::LaunchContext * context, NDArray* logits, NDArray* labels, NDArray* output) {
|
|
BUILD_SINGLE_SELECTOR(logits->dataType(), sigmCrossEntropyGrad_, (logits, labels, output), FLOAT_TYPES);
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template <typename T>
|
|
static void tanhDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
T th = nd4j::math::nd4j_tanh<T,T>(x);
|
|
return y * ((T)1.0f - (th * th));
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void tanhDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), tanhDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
// return static_cast<X>(d2) * simdOps::HardTanhDerivative<X>::op(d1, nullptr);
|
|
template <typename T>
|
|
static void hardTanhDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
T th = nd4j::math::nd4j_tanh<T,T>(x);
|
|
return y * simdOps::HardTanhDerivative<T>::op(x, nullptr);
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void hardTanhDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), hardTanhDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void rationalTanhDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return y * simdOps::RationalTanhDerivative<T>::op(x, nullptr);
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void rationalTanhDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), rationalTanhDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void rectifiedTanhDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return x > (T) 0.0f ? y * (nd4j::math::nd4j_tanhderivative<T,T>(x)) : (T) 0.0f;
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void rectifiedTanhDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), rectifiedTanhDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
// X f = (X) 1.0f + nd4j::math::nd4j_abs<X>(d1);
|
|
// return (X) d2 * ((X) 1.0f / (f * f));
|
|
|
|
template <typename T>
|
|
static void softSignDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
T ss = (T)1.f + nd4j::math::nd4j_abs<T>(x);
|
|
return y * ((T) 1.0f / (ss * ss));
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void softSignDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), softSignDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void softPlusDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
T p = nd4j::math::nd4j_pow<T, T, T>(static_cast<T>(M_E), x);
|
|
return y * (p / (p + 1.));
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void softPlusDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), softPlusDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
///
|
|
/// \param theFirst
|
|
/// \param theSecond
|
|
/// \param theOutput
|
|
template <typename T>
|
|
static void sigmoidDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
T s = nd4j::math::nd4j_sigmoid<T,T>(x);
|
|
return y * (s * ((T) 1.0f - s));
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void sigmoidDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), sigmoidDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void hardSigmoidDerivative_(NDArray* input, NDArray* epsilon, NDArray* output) {
|
|
auto functor = LAMBDA_TT(x, y){
|
|
return y * simdOps::HardSigmoidDerivative<T>::op(x, nullptr);
|
|
};
|
|
|
|
input->applyPairwiseLambda<T>(epsilon, functor, output);
|
|
}
|
|
|
|
void hardSigmoidDerivative(nd4j::LaunchContext * context, NDArray* theFirst, NDArray* theSecond, NDArray* theOutput) {
|
|
BUILD_SINGLE_SELECTOR(theFirst->dataType(), hardSigmoidDerivative_, (theFirst, theSecond, theOutput), FLOAT_TYPES);
|
|
}
|
|
|
|
template <typename T>
|
|
static void logSumExp_(NDArray* input, NDArray* axis, NDArray* output) {
|
|
// reduce along axis with
|
|
std::unique_ptr<NDArray> tempInput(input->dup());
|
|
input->applyTransform(transform::Exp, tempInput.get());
|
|
std::vector<int> axisVector;
|
|
if (axis != nullptr) {
|
|
axisVector.resize(axis->lengthOf());
|
|
for (size_t i = 0; i < axisVector.size(); ++i)
|
|
axisVector[i] = axis->e<int>(i);
|
|
}
|
|
tempInput->reduceAlongDimension(reduce::Sum, output, axisVector);
|
|
output->applyTransform(transform::Log, nullptr, nullptr);
|
|
}
|
|
|
|
template <typename T>
|
|
static void logSumExp_(NDArray* input, NDArray* subtrah, NDArray* axis, NDArray* output) {
|
|
// reduce along axis with
|
|
std::unique_ptr<NDArray> tempInput(input->dup());
|
|
input->applyPairwiseTransform(pairwise::Subtract, subtrah, tempInput.get(), nullptr);
|
|
tempInput->applyTransform(transform::Exp, nullptr, nullptr);
|
|
|
|
std::vector<int> axisVector;
|
|
if (axis != nullptr) {
|
|
axisVector.resize(axis->lengthOf());
|
|
for (size_t i = 0; i < axisVector.size(); ++i)
|
|
axisVector[i] = axis->e<int>(i);
|
|
}
|
|
tempInput->reduceAlongDimension(reduce::Sum, output, axisVector);
|
|
output->applyTransform(transform::Log, nullptr, nullptr);
|
|
}
|
|
|
|
void logSumExp(nd4j::LaunchContext * context, NDArray* input, NDArray* axis, NDArray* output) {
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), logSumExp_, (input, axis, output), FLOAT_TYPES);
|
|
}
|
|
|
|
void logSumExp(nd4j::LaunchContext * context, NDArray* input, NDArray* subtrah, NDArray* axis, NDArray* output) {
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), logSumExp_, (input, subtrah, axis, output), FLOAT_TYPES);
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
template <typename T>
|
|
static void weightedCrossEntropyWithLogitsFunctor_(NDArray const* targets, NDArray const* input, NDArray const* weights, NDArray* output) {
|
|
|
|
T posWeight = weights->e<T>(0);
|
|
|
|
auto mainRoutineT1 = LAMBDA_TT(_x, _z, posWeight) {
|
|
T targetWeight = (1. + (posWeight - (T)1.f) * _z);
|
|
return (1. - _z) * _x +
|
|
targetWeight * (nd4j::math::nd4j_log<T,T>((T)1.f + nd4j::math::nd4j_exp<T,T>(-nd4j::math::nd4j_abs(_x))) +
|
|
nd4j::math::nd4j_max(-_x, T(0.f))
|
|
);
|
|
};
|
|
|
|
auto mainRoutineT2 = LAMBDA_TTT(_x, _z, _w) {
|
|
return (((T)1.0 - _z) * _x) +
|
|
_w * (nd4j::math::nd4j_log<T,T>(T(1.) + nd4j::math::nd4j_exp<T,T>(-nd4j::math::nd4j_abs(_x))) +
|
|
nd4j::math::nd4j_max(-_x, T(0.f)));
|
|
};
|
|
|
|
|
|
if (weights->isScalar()) {
|
|
const_cast<NDArray*>(input)->applyPairwiseLambda<T>(const_cast<NDArray*>(targets), mainRoutineT1, output);
|
|
}
|
|
else
|
|
{
|
|
std::unique_ptr<NDArray> targetVector(new NDArray(*weights));
|
|
targetVector->applyScalar(scalar::Add, -1.f);
|
|
|
|
std::unique_ptr<NDArray> targetTensor(new NDArray(*targets));
|
|
*targetTensor = (*targetVector * *targetTensor) + T(1.f);
|
|
const_cast<NDArray*>(input)->applyTriplewiseLambda<T>(const_cast<NDArray*>(targets), targetTensor.get(), mainRoutineT2, output);
|
|
}
|
|
}
|
|
|
|
void weightedCrossEntropyWithLogitsFunctor(nd4j::LaunchContext * context, NDArray const* targets, NDArray const* input, NDArray const* weights, NDArray* output) {
|
|
BUILD_SINGLE_SELECTOR(targets->dataType(), weightedCrossEntropyWithLogitsFunctor_, (targets, input, weights, output), FLOAT_TYPES);
|
|
}
|
|
|
|
}
|
|
}
|
|
} |