cavis/libnd4j/include/ops/declarable/generic/updaters/adaMaxUpdater.cpp

94 lines
4.1 KiB
C++
Raw Normal View History

Learning updaters for gradient (#335) * libnd4j raw implementation of sgd upader Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j some corrections and simple test added Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j some corrections after discussion Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j integrate applyScalar Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j raw implementation of rmsPropUpdater on cpu Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j fix operations declaration Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j rmsPropUpdater added, test cases for sgd, etc Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j fixed several typos Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j some fixes and improvements for rmsPropUpdater based on Java tests Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j fixed cuda implementation, update tests and corrected behavior according java tests Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j adaGrad updater added Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j one minor fix for ada grad Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j several more fixes for ada_grad Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j nesterovs updater added Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j fixed nesterovs updater behavior, several typos and rename file Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j one minor typo Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j ada max updater added Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j fixed several typos in adaMax updater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j fixed several typos in adaMaxUpdater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j several fixes for adaMax, added Adam Updater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j adaDeltaUpdater added, minor fixes for adamUpdater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j several fixes for adaDeltaUpdater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j nadamUpdater added Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j one more correction for nadam updater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j several fixes for nadam updater and added amsGradUpdater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j several typos fixed in amsGradUpdater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j some corrections and added f order support rmsProp updater Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j added support of f order for all updaters and modify tests for testing in place Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j fixed issues for updates when not in place mode used, added tests for f order Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j added input shape checks Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j some corrections for different cases handling Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j some code clean up and optimize per request Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j updaters refactoring after review Signed-off-by: Oleg <oleg.semeniv@gmail.com> * SgdUpdater wrapper Signed-off-by: raver119 <raver119@gmail.com> * first test Signed-off-by: raver119 <raver119@gmail.com> * RmsPropUpdater added Signed-off-by: raver119 <raver119@gmail.com> * NadamUpdater + NesterovsUpdater Signed-off-by: raver119 <raver119@gmail.com> * AmsGradUpdater Signed-off-by: raver119 <raver119@gmail.com> * AdamUpdater added Signed-off-by: raver119 <raver119@gmail.com> * AdaGradUpdater + AdaDeltaUpdater + AdaMaxUpdater Signed-off-by: raver119 <raver119@gmail.com> * AdaGradUpdater test added Signed-off-by: raver119 <raver119@gmail.com> * libnd4j remove input parameters parsing through NDArray, split implementation of helpers to separate files, added some rename, etc Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j next step to split operations implementation into separate files Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j merge master and minor corrections Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j revert some changes of split implementation Signed-off-by: Oleg <oleg.semeniv@gmail.com> * libnd4j forgot to add header file Signed-off-by: Oleg <oleg.semeniv@gmail.com> * public default constructors Signed-off-by: raver119 <raver119@gmail.com> * ImportClassMapping updated Signed-off-by: raver119 <raver119@gmail.com> Co-authored-by: raver119 <raver119@gmail.com>
2020-03-23 05:28:31 +01:00
/*******************************************************************************
* 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/headers/updaters.h>
#include <ops/declarable/CustomOperations.h>
#include <helpers/ConstantTadHelper.h>
#include <execution/Threads.h>
#include <array/NDArray.h>
namespace sd {
namespace ops {
CONFIGURABLE_OP_IMPL(ada_max_updater, 3, 3, true, 0, 0) {
const auto gradient = INPUT_VARIABLE(0);
const auto initStateU = INPUT_VARIABLE(1);
const auto initStateM = INPUT_VARIABLE(2);
auto update = OUTPUT_VARIABLE(0);
auto stateU = OUTPUT_VARIABLE(1);
auto stateM = OUTPUT_VARIABLE(2);
// todo maybe we need an error like on Java side
if (gradient->isEmpty() || initStateU->isEmpty() || initStateM->isEmpty())
return Status::OK();
REQUIRE_TRUE(gradient->isSameShape(initStateU), 0, "ADA_MAX UPDATER OP: input state V must have the same shape as gradient,"
" expected shape %s, but got %s!", ShapeUtils::shapeAsString(gradient->getShapeInfo()).c_str(),
ShapeUtils::shapeAsString(initStateU->getShapeInfo()).c_str());
REQUIRE_TRUE(gradient->isSameShape(initStateM), 0, "ADA_MAX UPDATER OP: input state M must have the same shape as gradient,"
" expected shape %s, but got %s!", ShapeUtils::shapeAsString(gradient->getShapeInfo()).c_str(),
ShapeUtils::shapeAsString(initStateM->getShapeInfo()).c_str());
bool bParamsSupply = 7 == block.width() || 4 == block.getTArguments()->size();
int iteration = block.getIArguments()->size() > 0 ? INT_ARG(0) : 0;
REQUIRE_TRUE(bParamsSupply, 0, "ADA_MAX UPDATER OP: learning rate, beta 1, beta 2 and epsilon were not provided!");
double dLr, dBeta1, dBeta2, dEpsilon;
if (block.width() > 3) {
const auto lr = INPUT_VARIABLE(3);
const auto beta1 = INPUT_VARIABLE(4);
const auto beta2 = INPUT_VARIABLE(5);
const auto epsilon = INPUT_VARIABLE(6);
REQUIRE_TRUE(lr->isScalar(), 0, "ADA_MAX UPDATER OP: Learning rate has to be a scalar, but instead got rank %i!", lr->rankOf());
REQUIRE_TRUE(beta1->isScalar(), 0, "ADA_MAX UPDATER OP: beta 1 has to be a scalar, but instead got rank %i!", beta1->rankOf());
REQUIRE_TRUE(beta2->isScalar(), 0, "ADA_MAX UPDATER OP: beta 2 has to be a scalar, but instead got rank %i!", beta2->rankOf());
REQUIRE_TRUE(epsilon->isScalar(), 0, "ADA_MAX UPDATER OP: Epsilon has to be a scalar, but instead got rank %i!", epsilon->rankOf());
dLr = lr->e<double>(0);
dBeta1 = beta1->e<double>(0);
dBeta2 = beta2->e<double>(0);
dEpsilon = epsilon->e<double>(0);
}
else {
dLr = T_ARG(0);
dBeta1 = T_ARG(1);
dBeta2 = T_ARG(2);
dEpsilon = T_ARG(3);
}
helpers::updaterAdaMax(block.launchContext(), *gradient, *initStateU, *initStateM, *update, *stateU, *stateM, dLr, dBeta1, dBeta2, dEpsilon, iteration);
return Status::OK();
}
DECLARE_TYPES(ada_max_updater) {
getOpDescriptor()->setAllowedInputTypes({ ALL_FLOATS })
->setSameMode(true);
}
}
}