cavis/libnd4j/include/ops/declarable/helpers/cpu/updaterAdaGrad.cpp

92 lines
3.7 KiB
C++

/*******************************************************************************
* Copyright (c) 2019-2020 Konduit K.K.
*
* 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 Oleh Semeniv (oleg.semeniv@gmail.com)
//
#include <ops/declarable/helpers/updatersHelpers.h>
#include <execution/Threads.h>
#include <math/platformmath.h>
#include <math/templatemath.h>
namespace sd {
namespace ops {
namespace helpers {
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename T>
static void adaGradUpdater_(const NDArray& gradient, const NDArray& initState, NDArray& update, NDArray& stateH, const double dLr, const double dEpsilon) {
const T* grad = gradient.bufferAsT<T>();
const T* init = initState.bufferAsT<T>();
T* up = update.bufferAsT<T>();
T* st = stateH.bufferAsT<T>();
const T lr = static_cast<T>(dLr);
const T epsilon = static_cast<T>(dEpsilon);
bool bEws1 = 1 == gradient.ews() && 1 == update.ews() && 1 == stateH.ews() && 1 == initState.ews();
bool bSameOrdering = gradient.ordering() == update.ordering() && update.ordering() == stateH.ordering() && stateH.ordering() == initState.ordering();
if (bEws1 && bSameOrdering) {
auto func = PRAGMA_THREADS_FOR{
for (auto i = start; i < stop; i++) {
st[i] = init[i] + grad[i] * grad[i];
up[i] = (lr * grad[i]) / (math::nd4j_sqrt<T, T>(st[i]) + epsilon);
}
};
samediff::Threads::parallel_for(func, 0, gradient.lengthOf(), 1);
return;
}
bool bXZsame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), update.getShapeInfo());
bool bXInSame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), initState.getShapeInfo());
bool bXStSame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), stateH.getShapeInfo());
auto func = PRAGMA_THREADS_FOR{
int coords[MAX_RANK];
for (auto i = start; i < stop; i++) {
shape::index2coordsCPU(start, i, gradient.getShapeInfo(), coords);
const auto xOffset = shape::getOffset(gradient.getShapeInfo(), coords);
const auto zOffset = bXZsame ? xOffset : shape::getOffset(update.getShapeInfo(), coords);
const auto initOffset = bXInSame ? xOffset : shape::getOffset(initState.getShapeInfo(), coords);
const auto stOffset = bXStSame ? xOffset : shape::getOffset(stateH.getShapeInfo(), coords);
st[stOffset] = init[initOffset] + grad[xOffset] * grad[xOffset];
up[zOffset] = (lr * grad[xOffset]) / (math::nd4j_sqrt<T, T>(st[stOffset]) + epsilon);
}
};
samediff::Threads::parallel_for(func, 0, gradient.lengthOf(), 1);
return;
}
void updaterAdaGrad(sd::LaunchContext* context, const NDArray& gradient, const NDArray& initState, NDArray& update, NDArray& stateH,
const double dLr, const double dEpsilon) {
BUILD_SINGLE_SELECTOR(gradient.dataType(), adaGradUpdater_, (gradient, initState, update, stateH, dLr, dEpsilon), FLOAT_TYPES);
}
}
}
}