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

/*******************************************************************************
 * 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 adamUpdater_(const NDArray& gradient, const NDArray& initStateU, const NDArray& initStateM, NDArray& update, 
                        NDArray& stateU, NDArray& stateM,  const double dLr, const double dBeta1, const double dBeta2, 
                        const double dEpsilon, const int nIteration) {

    const T* grad = gradient.bufferAsT<T>();
    const T* initU = initStateU.bufferAsT<T>();
    const T* initM = initStateM.bufferAsT<T>();

    T* up = update.bufferAsT<T>();
    T* stU = stateU.bufferAsT<T>();
    T* stM = stateM.bufferAsT<T>();

    const T lr = static_cast<T>(dLr);
    const T beta1 = static_cast<T>(dBeta1);
    const T beta2 = static_cast<T>(dBeta2);
    const T epsilon = static_cast<T>(dEpsilon);
    const T iteration = static_cast<T>(nIteration);

    const T beta1T = sd::math::nd4j_pow<T, T, T>(beta1, (iteration + 1));
    const T beta2T = sd::math::nd4j_pow<T, T, T>(beta2, (iteration + 1));

    T epsilonT = lr * sd::math::nd4j_sqrt<T, T>(1. - beta2T) / (1.0 - beta1T);
    if (sd::math::nd4j_isnan(epsilonT) || 0 == epsilonT || sd::math::nd4j_isinf(epsilonT))
        epsilonT = epsilon;

    bool bEws1 = 1 == gradient.ews() && 1 == update.ews() && 1 == stateM.ews() && 1 == initStateM.ews() && 1 == stateU.ews() && 1 == initStateU.ews();
    bool bSameOrdering = gradient.ordering() == update.ordering() &&
        update.ordering() == stateU.ordering() &&
        stateU.ordering() == initStateU.ordering() &&
        stateU.ordering() == initStateM.ordering() && stateM.ordering() == initStateM.ordering();

    if (bEws1 && bSameOrdering) {
            
            auto func = PRAGMA_THREADS_FOR{
                 for (auto i = start; i < stop; i++) {
                     stM[i] = beta1 * initM[i] + grad[i] * (1 - beta1);
                     stU[i] = beta2 * initU[i] + grad[i] * grad[i] * (1 - beta2);

                     up[i] = (stM[i] * epsilonT) / (sd::math::nd4j_sqrt<T, T>(stU[i]) + epsilon);
                 }
            };

           samediff::Threads::parallel_for(func, 0, gradient.lengthOf(), 1);
           return;
    }

    bool bXZsame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), update.getShapeInfo());
    bool bXInVSame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), initStateU.getShapeInfo());
    bool bXStVSame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), stateU.getShapeInfo());
    bool bXInMSame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), initStateM.getShapeInfo());
    bool bXStMSame = shape::haveSameShapeAndStrides(gradient.getShapeInfo(), stateM.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 initUOffset = bXInVSame ? xOffset : shape::getOffset(initStateU.getShapeInfo(), coords);
            const auto stUOffset = bXStVSame ? xOffset : shape::getOffset(stateU.getShapeInfo(), coords);
            const auto initMOffset = bXInVSame ? xOffset : shape::getOffset(initStateM.getShapeInfo(), coords);
            const auto stMOffset = bXStMSame ? xOffset : shape::getOffset(stateM.getShapeInfo(), coords);
            
            stM[stMOffset] = beta1 * initM[initMOffset] + grad[xOffset] * (1 - beta1);
            stU[stUOffset] = beta2 * initU[initUOffset] + grad[xOffset] * grad[xOffset] * (1 - beta2);

            up[zOffset] = (stM[stMOffset] * epsilonT) / (sd::math::nd4j_sqrt<T, T>(stU[stUOffset]) + epsilon);
        }
    };

    samediff::Threads::parallel_for(func, 0, gradient.lengthOf(), 1);
    return;
}

void updaterAdam(sd::LaunchContext* context, const NDArray& gradient, const NDArray& initStateU, const NDArray& initStateM, NDArray& update, NDArray& stateU, NDArray& stateM,  const double dLr, const double dBeta1, const double dBeta2, const double dEpsilon, const int nIteration) {
    BUILD_SINGLE_SELECTOR(gradient.dataType(), adamUpdater_, (gradient, initStateU, initStateM, update, stateU, stateM, dLr, dBeta1, dBeta2, dEpsilon, nIteration), FLOAT_TYPES);
}

}
}
}