320 lines
15 KiB
C++
320 lines
15 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 Yurii Shyrma (iuriish@yahoo.com), created on 23.11.2017
|
|
//
|
|
|
|
#include <system/op_boilerplate.h>
|
|
#if NOT_EXCLUDED(OP_huber_loss)
|
|
|
|
#include <ops/declarable/CustomOperations.h>
|
|
|
|
namespace sd {
|
|
namespace ops {
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
CUSTOM_OP_IMPL(huber_loss, 3, 1, false, 1, 1) {
|
|
auto predictions = INPUT_VARIABLE(0);
|
|
auto weights = INPUT_VARIABLE(1);
|
|
auto labels = INPUT_VARIABLE(2);
|
|
auto output = OUTPUT_VARIABLE(0);
|
|
|
|
int reductionMode = INT_ARG(0); // 0 - "none"; 1 - "weighted_sum"; 2 - "weighted_mean"; 3 - "weighted_sum_by_nonzero_weights"
|
|
// FIXME: double?
|
|
double delta = T_ARG(0);
|
|
|
|
// input validation
|
|
REQUIRE_TRUE(labels->isSameShape(predictions), 0, "HUBER_LOSS OP: labels and predictions arrays must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(labels).c_str(), ShapeUtils::shapeAsString(predictions).c_str());
|
|
// weights array can be single scalar or has the same rank as labels, and must be broadcastable to labels
|
|
REQUIRE_TRUE(weights->isScalar() || weights->rankOf() == labels->rankOf(), 0, "HUBER_LOSS OP: weights array should be scalar or have the same rank as labels array, but got %i and %i correspondingly!", weights->rankOf(), labels->rankOf());
|
|
// check whether broadcast operation is possible for weights array
|
|
REQUIRE_TRUE(weights->isScalar() || ShapeUtils::areShapesBroadcastable(*weights, *labels), 0, "HUBER_LOSS OP: shapes of weights and labels arrays should be broadcastable, but got weights = %s and labels = %s instead!", ShapeUtils::shapeAsString(weights).c_str(), ShapeUtils::shapeAsString(labels).c_str());
|
|
// only 4 possible reduction modes exist
|
|
REQUIRE_TRUE(reductionMode==0 || reductionMode==1 || reductionMode==2 || reductionMode==3, 0, "HUBER_LOSS OP: reduction mode value is not acceptable, possible values are 0, 1, 2, 3, but got %i instead!", reductionMode);
|
|
|
|
// perform weights broadcasting/tile to predictions if needed
|
|
auto weightsBroad = weights;
|
|
if(!weights->isScalar() && !weights->isSameShape(predictions))
|
|
weightsBroad = new NDArray(weights->tileToShape(predictions->shapeInfo()));
|
|
|
|
auto error = *predictions - *labels;
|
|
error.applyTransform(transform::Abs, error);
|
|
NDArray quadratic(error.shapeInfo(), block.getWorkspace());
|
|
error.applyScalar(scalar::MinPairwise, delta, quadratic);
|
|
|
|
NDArray E = quadratic * quadratic * 0.5f + (error - quadratic)*delta;
|
|
|
|
// multiply E on weights
|
|
E *= *weightsBroad;
|
|
|
|
switch (reductionMode) {
|
|
case 0: { // 0 - "none", un-reduced weighted losses with the same shape as labels.
|
|
output->assign(E);
|
|
break;
|
|
}
|
|
case 1: { // 1 - "weighted_sum", output is scalar and equal to sum of all elements of E array
|
|
E.reduceNumber(reduce::Sum, *output);
|
|
break;
|
|
}
|
|
case 2: { // 2 - "weighted_mean", output is scalar and equal to sum of all elements of E array divided by sum of all elements of weightsBroad array
|
|
NDArray sum;
|
|
sum.setContext(block.launchContext());
|
|
if (weights->isScalar())
|
|
sum = *weights * E.lengthOf();
|
|
else
|
|
sum = weightsBroad->reduceNumber(reduce::Sum);
|
|
|
|
if (sum.e<double>(0) == 0.)
|
|
*output = 0.;
|
|
else
|
|
output->assign(E.reduceNumber(reduce::Sum) / sum);
|
|
break;
|
|
}
|
|
case 3: { // 3 - "weighted_sum_by_nonzero_weights", output is scalar and equal to scalar sum of all elements of E array divided by number of non-zero weights
|
|
Nd4jLong numOfNonZeroWeights = 0;
|
|
if(weights->isScalar()) {
|
|
if(weights->e<double>(0) != 0.)
|
|
numOfNonZeroWeights = E.lengthOf();
|
|
}
|
|
else {
|
|
numOfNonZeroWeights = weightsBroad->reduceNumber(reduce::CountNonZero).e<Nd4jLong>(0);
|
|
}
|
|
|
|
if (numOfNonZeroWeights == 0)
|
|
(*output) = 0.;
|
|
else
|
|
output->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights));
|
|
break;
|
|
}
|
|
}
|
|
|
|
if(weightsBroad != weights)
|
|
delete weightsBroad;
|
|
|
|
return Status::OK();
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
DECLARE_TYPES(huber_loss) {
|
|
|
|
getOpDescriptor()->setAllowedInputTypes(sd::DataType::ANY)->setAllowedOutputTypes({ALL_FLOATS});
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
DECLARE_SHAPE_FN(huber_loss) {
|
|
|
|
auto predictionsShapeInfo = inputShape->at(0);
|
|
auto weightsShapeInfo = inputShape->at(1);
|
|
auto labelsShapeInfo = inputShape->at(2);
|
|
|
|
// labels and predictions must have the same shapes
|
|
REQUIRE_TRUE(shape::shapeEquals(labelsShapeInfo, predictionsShapeInfo), 0, "HUBER_LOSS OP: labels and predictions arrays must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(labelsShapeInfo).c_str(), ShapeUtils::shapeAsString(predictionsShapeInfo).c_str());
|
|
// weights array can be single scalar or has the same rank as labels, and must be broadcastable to labels
|
|
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || shape::rank(weightsShapeInfo) == shape::rank(labelsShapeInfo), 0, "HUBER_LOSS OP: weights array should be scalar or have the same rank as labels array, but got %i and %i correspondingly!", shape::rank(weightsShapeInfo), shape::rank(labelsShapeInfo));
|
|
// check whether broadcast operation is possible for weights array
|
|
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || ShapeUtils::areShapesBroadcastable(weightsShapeInfo, labelsShapeInfo), 0, "HUBER_LOSS OP: shapes of weights and labels arrays should be broadcastable, but got weights = %s and labels = %s instead!", ShapeUtils::shapeAsString(weightsShapeInfo).c_str(), ShapeUtils::shapeAsString(labelsShapeInfo).c_str());
|
|
|
|
DataType outType = DataTypeUtils::pickFloatingType(ArrayOptions::dataType(predictionsShapeInfo));
|
|
Nd4jLong const* outShapeInfo = nullptr;
|
|
|
|
if(INT_ARG(0) != 0) // in this case output is scalar
|
|
outShapeInfo = ConstantShapeHelper::getInstance().scalarShapeInfo(outType);
|
|
else // in this case output has the same shape as labels and predictions
|
|
outShapeInfo = ConstantShapeHelper::getInstance().createShapeInfo(ShapeDescriptor(outType, shape::order(labelsShapeInfo), shape::shapeOf(labelsShapeInfo), shape::rank(labelsShapeInfo)));
|
|
|
|
return SHAPELIST(outShapeInfo);
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
CUSTOM_OP_IMPL(huber_loss_grad, 3, 3, false, 1, 1) {
|
|
|
|
auto predictions = INPUT_VARIABLE(0);
|
|
auto weights = INPUT_VARIABLE(1);
|
|
auto labels = INPUT_VARIABLE(2);
|
|
|
|
auto dLdp = OUTPUT_VARIABLE(0); // dL/dpredictions
|
|
auto dLdw = OUTPUT_VARIABLE(1); // dL/dweights
|
|
auto dLdl = OUTPUT_VARIABLE(2); // dL/dlabels
|
|
|
|
auto delta = T_ARG(0);
|
|
|
|
int reductionMode = INT_ARG(0); // 0 - "none"; 1 - "weighted_sum"; 2 - "weighted_mean"; 3 - "weighted_sum_by_nonzero_weights"
|
|
// take into account Alex's proposition to treat "none" the same as "weighted_sum" mode when calculating gradients
|
|
if(reductionMode == 0)
|
|
reductionMode = 1;
|
|
|
|
// inputs validation
|
|
REQUIRE_TRUE(labels->isSameShape(predictions), 0, "HUBER_LOSS_GRAD OP: labels and predictions arrays must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(labels).c_str(), ShapeUtils::shapeAsString(predictions).c_str());
|
|
// weights array can be single scalar or has the same rank as labels, and must be broadcastable to labels
|
|
REQUIRE_TRUE(weights->isScalar() || weights->rankOf() == labels->rankOf(), 0, "HUBER_LOSS_GRAD OP: weights array should be scalar or have the same rank as labels array, but got %i and %i correspondingly!", weights->rankOf(), labels->rankOf());
|
|
// check whether broadcast operation is possible for weights array
|
|
REQUIRE_TRUE(weights->isScalar() || ShapeUtils::areShapesBroadcastable(*weights, *labels), 0, "HUBER_LOSS_GRAD OP: shapes of weights and labels arrays should be broadcastable, but got weights = %s and labels = %s instead!", ShapeUtils::shapeAsString(weights).c_str(), ShapeUtils::shapeAsString(labels).c_str());
|
|
// only 4 possible reduction modes exist
|
|
REQUIRE_TRUE(reductionMode==0 || reductionMode==1 || reductionMode==2 || reductionMode==3, 0, "HUBER_LOSS_GRAD OP: reduction mode value is not acceptable, possible values are 0, 1, 2, 3, but got %i instead!", reductionMode);
|
|
|
|
// perform weights broadcasting/tile to labels if needed
|
|
auto weightsBroad = weights;
|
|
if(!weights->isScalar() && !weights->isSameShape(predictions))
|
|
weightsBroad = new NDArray(weights->tileToShape(predictions->shapeInfo()));
|
|
|
|
NDArray diff = *predictions - *labels;
|
|
NDArray absDiff(diff);
|
|
absDiff.applyTransform(transform::Abs, absDiff);
|
|
NDArray quadratic(absDiff);
|
|
absDiff.applyScalar(scalar::MinPairwise, delta, quadratic);
|
|
|
|
NDArray E = quadratic * quadratic * 0.5f + (absDiff - quadratic)*delta;
|
|
|
|
NDArray lteMask(diff.shapeInfo(), BOOL, true, block.launchContext());
|
|
absDiff.applyScalar(scalar::LessThanOrEqual, delta, lteMask);
|
|
|
|
NDArray gtMask(diff.shapeInfo(), BOOL, true, block.launchContext());
|
|
absDiff.applyScalar(scalar::GreaterThan, delta, gtMask);
|
|
|
|
NDArray signDiff(diff);
|
|
diff.applyTransform(transform::Sign, signDiff);
|
|
|
|
|
|
auto gtMaskFloat = gtMask.cast(diff.dataType());
|
|
auto lteMaskFloat = lteMask.cast(diff.dataType());
|
|
|
|
|
|
dLdp->assign( lteMaskFloat * diff + gtMaskFloat * delta * signDiff);
|
|
dLdl->assign(-lteMaskFloat * diff - gtMaskFloat * delta * signDiff);
|
|
|
|
switch (reductionMode) {
|
|
|
|
case 1: { // 1 - "none" and "weighted_sum", output is scalar and equal to sum of all elements of E array
|
|
|
|
*dLdp *= *weightsBroad;
|
|
*dLdl *= *weightsBroad;
|
|
|
|
if(weights->isScalar())
|
|
dLdw->assign(E.reduceNumber(reduce::Sum));
|
|
else if(weights != weightsBroad) {
|
|
std::vector<int> axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo());
|
|
E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true);
|
|
}
|
|
else
|
|
dLdw->assign(E);
|
|
break;
|
|
}
|
|
case 2: { // 2 - "weighted_mean", output is scalar and equal to sum of all elements of E array divided by sum of all elements of weightsBroad array
|
|
|
|
NDArray sum;
|
|
sum.setContext(block.launchContext());
|
|
if (weights->isScalar())
|
|
sum = (*weights) * E.lengthOf();
|
|
else
|
|
sum = weightsBroad->reduceNumber(reduce::Sum);
|
|
|
|
if (sum.e<double>(0) == 0.) {
|
|
*dLdp = 0.;
|
|
*dLdl = 0.;
|
|
*dLdw = 0.;
|
|
}
|
|
else {
|
|
|
|
*dLdp *= *weightsBroad / sum;
|
|
*dLdl *= *weightsBroad / sum;
|
|
|
|
if(weights->isScalar())
|
|
*dLdw = 0.;
|
|
else if(weights != weightsBroad) {
|
|
std::vector<int> axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo());
|
|
((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum)).reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true);
|
|
}
|
|
else
|
|
dLdw->assign((E * sum - (E * *weightsBroad).reduceNumber(reduce::Sum)) / (sum*sum));
|
|
}
|
|
break;
|
|
}
|
|
case 3: { // 3 - "weighted_sum_by_nonzero_weights", output is scalar and equal to scalar sum of all elements of E array divided by number of non-zero weights
|
|
|
|
Nd4jLong numOfNonZeroWeights = 0;
|
|
if(weights->isScalar()) {
|
|
if(weights->e<double>(0) != 0.)
|
|
numOfNonZeroWeights = E.lengthOf();
|
|
}
|
|
else
|
|
numOfNonZeroWeights = weightsBroad->reduceNumber(reduce::CountNonZero).e<Nd4jLong>(0);
|
|
|
|
if (numOfNonZeroWeights == 0) {
|
|
*dLdp = 0.;
|
|
*dLdl = 0.;
|
|
*dLdw = 0.;
|
|
}
|
|
else {
|
|
auto numOfNonZeroWeightsScalar = NDArrayFactory::create(dLdw->dataType(), numOfNonZeroWeights, block.launchContext());
|
|
|
|
if(weights->isScalar())
|
|
dLdw->assign(E.reduceNumber(reduce::Sum) / double(numOfNonZeroWeights));
|
|
else if(weights != weightsBroad) {
|
|
std::vector<int> axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo());
|
|
E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true);
|
|
*dLdw /= numOfNonZeroWeightsScalar;
|
|
}
|
|
else
|
|
dLdw->assign(E / numOfNonZeroWeightsScalar);
|
|
|
|
NDArray temp = *weightsBroad / numOfNonZeroWeightsScalar;
|
|
*dLdp *= temp;
|
|
*dLdl *= temp;
|
|
}
|
|
break;
|
|
}
|
|
}
|
|
|
|
if(weightsBroad != weights)
|
|
delete weightsBroad;
|
|
|
|
return Status::OK();
|
|
}
|
|
|
|
DECLARE_TYPES(huber_loss_grad) {
|
|
|
|
getOpDescriptor()->setAllowedInputTypes(sd::DataType::ANY)->setAllowedOutputTypes({ALL_FLOATS});
|
|
}
|
|
|
|
DECLARE_SHAPE_FN(huber_loss_grad) {
|
|
|
|
auto predictionsShapeInfo = inputShape->at(0);
|
|
auto weightsShapeInfo = inputShape->at(1);
|
|
auto labelsShapeInfo = inputShape->at(2);
|
|
|
|
// labels and predictions must have the same shapes
|
|
REQUIRE_TRUE(shape::shapeEquals(labelsShapeInfo, predictionsShapeInfo), 0, "HUBER_LOSS_GRAD OP: labels and predictions arrays must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(labelsShapeInfo).c_str(), ShapeUtils::shapeAsString(predictionsShapeInfo).c_str());
|
|
// weights array can be single scalar or has the same rank as labels, and must be broadcastable to labels
|
|
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || shape::rank(weightsShapeInfo) == shape::rank(labelsShapeInfo), 0, "HUBER_LOSS_GRAD OP: weights array should be scalar or have the same rank as labels array, but got %i and %i correspondingly!", shape::rank(weightsShapeInfo), shape::rank(labelsShapeInfo));
|
|
// check whether broadcast operation is possible for weights array
|
|
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || ShapeUtils::areShapesBroadcastable(weightsShapeInfo, labelsShapeInfo), 0, "HUBER_LOSS_GRAD OP: shapes of weights and labels arrays should be broadcastable, but got weights = %s and labels = %s instead!", ShapeUtils::shapeAsString(weightsShapeInfo).c_str(), ShapeUtils::shapeAsString(labelsShapeInfo).c_str());
|
|
|
|
DataType outType = DataTypeUtils::pickFloatingType(ArrayOptions::dataType(predictionsShapeInfo));
|
|
|
|
auto dLdpShapeInfo = ShapeBuilders::copyShapeInfoAndType(predictionsShapeInfo, outType, false, block.getWorkspace());
|
|
auto dLdwShapeInfo = ShapeBuilders::copyShapeInfoAndType(weightsShapeInfo, outType, false, block.getWorkspace());
|
|
auto dLdlShapeInfo = ShapeBuilders::copyShapeInfoAndType(labelsShapeInfo, outType, false, block.getWorkspace());
|
|
|
|
return SHAPELIST(dLdpShapeInfo, dLdwShapeInfo, dLdlShapeInfo);
|
|
}
|
|
|
|
|
|
}
|
|
}
|
|
|
|
#endif |