cavis/libnd4j/include/helpers/impl/GradCheck.cpp

/*******************************************************************************
 * 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 Yurii Shyrma (iuriish@yahoo.com), created on 16.07.2018
//

#include <helpers/GradCheck.h>
#include <array/NDArrayFactory.h>


namespace sd {

//////////////////////////////////////////////////////////////////////////
void GradCheck::fillGradArrays(const LossFunc loss, const std::vector<NDArray*>& gradArrs) {

	const int numInGradArrs = gradArrs.size();

	// fill input gradient arrays in accordance to type of loss function
	switch(loss) {

		case MEAN:
			for(int i = 0; i < numInGradArrs; ++i)
				*gradArrs[i] = 1. / gradArrs[i]->lengthOf();
			break;

		case SUM:
			for(int i = 0; i < numInGradArrs; ++i)
				*gradArrs[i] = 1.;
			break;

		default:
			throw std::invalid_argument("GradCheck::fillGradArrays: invalid type of loss function !");
	}
}

//////////////////////////////////////////////////////////////////////////
bool GradCheck::checkGrad(ops::DeclarableOp& opFF, ops::DeclarableOp& opBP, const OpArgsHolder& argsHolderFF, const OpArgsHolder& argsHolderBP,
	                      const std::vector<bool>& whatArrsToCheck, const std::vector<double>& idxRange, const LossFunc loss) {

	const int numInArrsFF     = argsHolderFF.getNumInArrs();						// at the same time numInArrsFF = number of output arrays in opBP
	const int numInGradArrsBP = argsHolderBP.getNumInArrs() - numInArrsFF;  		// because argsHolderBP.getNumInArrs() = numInArrsFF + numInGradArrsBP
	const std::vector<NDArray*>& inArrsFF = argsHolderFF.getInArrs();
	const std::vector<NDArray*>& inArrsBP = argsHolderBP.getInArrs();

	// fill input gradient arrays in accordance to kind of loss function
	fillGradArrays(loss, std::vector<NDArray*>(&inArrsBP[numInArrsFF], &inArrsBP[numInArrsFF + numInGradArrsBP]));

	// back prop pass
	ResultSet outArrsBP = opBP.execute(argsHolderBP);		// number of output arrays in back prop = numInArrsFF;

	NDArray tmpScalar(sd::DataType::DOUBLE, inArrsFF[0]->getContext()); // scalar = 0

	for(int i = 0; i < numInArrsFF; ++i) {							// loop through input array

		if(!whatArrsToCheck.empty() && static_cast<bool>(whatArrsToCheck[i]) == false)
			continue;

		const Nd4jLong idxStart = static_cast<Nd4jLong>(idxRange[0] * inArrsFF[i]->lengthOf());
		const Nd4jLong idxEnd   = static_cast<Nd4jLong>(idxRange[1] * inArrsFF[i]->lengthOf());

		for(Nd4jLong j = idxStart; j < idxEnd; ++j) {			// loop through all elements for current array

			const double orig = inArrsFF[i]->e<double>(j);

			// add epsilon, feed forward
			inArrsFF[i]->p<double>(j, orig + EPSILON);
			ResultSet outArrsFF = opFF.execute(argsHolderFF);
			int numOutArrs = outArrsFF.size();
			double scorePlus = 0.;

			for(int k = 0; k < numOutArrs; ++k) {                // loop through output arrays
				if(loss == SUM)
					outArrsFF.at(k)->reduceNumber(reduce::Sum, tmpScalar);
				else
					outArrsFF.at(k)->reduceNumber(reduce::Mean, tmpScalar);
				scorePlus += tmpScalar.e<double>(0);
			}

			// subtract epsilon, feed forward
			inArrsFF[i]->p<double>(j, orig - EPSILON);
			outArrsFF = opFF.execute(argsHolderFF);
			double scoreMinus = 0.;

			for(int k = 0; k < numOutArrs; ++k) {            // loop through output arrays
				if(loss == SUM)
					outArrsFF.at(k)->reduceNumber(reduce::Sum, tmpScalar);
				else
					outArrsFF.at(k)->reduceNumber(reduce::Mean, tmpScalar);
				scoreMinus += tmpScalar.e<double>(0);
			}

			// restore initial element value
			inArrsFF[i]->p<double>(j, orig);

			// calculate numerical gradient
			const double numericalGrad = (scorePlus - scoreMinus) / (2 * EPSILON);
			if(std::isnan(numericalGrad) || std::isinf(numericalGrad)) {
				printf("GradCheck::checkGrad: got wrong value for numerical gradient for input array # %i and its element at position %lld ! \n", i, j);
				throw std::runtime_error("");
			}

			// get analytical gradient
			const double analyticGrad = outArrsBP.at(i)->e<double>(j);
			if(std::isnan(analyticGrad) || std::isinf(analyticGrad)) {
				printf("GradCheck::checkGrad: got wrong value for analytical gradient for input array # %i and its element at position %lld ! \n", i, j);
				throw std::runtime_error("");
			}

			// printf("%lld: num = %.15f, ana = %.15f\n", j, numericalGrad, analyticGrad);

			// calculate relative error
			double relError;
			if(numericalGrad == 0. && analyticGrad == 0.)
				relError = 0.;
			else
                relError = math::nd4j_abs<double>(analyticGrad - numericalGrad) / (math::nd4j_abs<double>(analyticGrad) + math::nd4j_abs<double>(numericalGrad));

            // verify result
            if(relError > MAXRELERR || std::isnan(relError)) {

            	if(math::nd4j_abs<double>(analyticGrad - numericalGrad) < MINABSERR)
            		continue;
            	printf("numericalGrad = %.15f,  analyticGrad = %.15f \n", numericalGrad, analyticGrad);
            	printf("GradCheck::checkGrad: got RELERROR = %f > MAXRELERROR(%f) for input array # %i and its element at position %lld ! \n", relError, MAXRELERR, i, j);
            	return false;
            }
		}
	}

	return true;
}



}