383 lines
21 KiB
C++
383 lines
21 KiB
C++
#pragma clang diagnostic push
|
|
#pragma ide diagnostic ignored "cert-err58-cpp"
|
|
|
|
/*******************************************************************************
|
|
* Copyright (c) 2015-2019 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 24.11.2017
|
|
// @author Paul Dubs
|
|
//
|
|
|
|
#include <system/op_boilerplate.h>
|
|
#if NOT_EXCLUDED(OP_mean_pairwssqerr_loss)
|
|
|
|
#include <ops/declarable/CustomOperations.h>
|
|
#include <numeric>
|
|
#include <iostream>
|
|
|
|
namespace sd {
|
|
namespace ops {
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
CUSTOM_OP_IMPL(mean_pairwssqerr_loss, 3, 1, false, 0, 1) {
|
|
/*
|
|
* Implementation of mean pairwise squared error loss
|
|
*
|
|
* For context on where this loss function may be useful see:
|
|
*
|
|
* Wei, Z., Zhang, J., Shen, X., Lin, Z., Mech, R., Hoai, M. and Samaras, D., 2018.
|
|
* Good view hunting: learning photo composition from dense view pairs. In Proceedings of the IEEE Conference on
|
|
* Computer Vision and Pattern Recognition (pp. 5437-5446).
|
|
*
|
|
* The paper defines the loss function as:
|
|
*
|
|
* L(y,q) = 1/((n*(n-1))/2) * (sum_(i,j=1..n,i!=j)((y_i - y_j) - (q_i - q_j))^2)
|
|
*
|
|
* with y: predictions, q: labels, n: length of y and q
|
|
*
|
|
* As creating those pairs is computationally expensive, we implement a mathematically equivalent function:
|
|
*
|
|
* L(y,q) = 4/(n*(n-1)) * (n * sum (y_i - q_i)^2 - (sum y_i - q_i)^2)
|
|
*
|
|
* This equivalency can be derived as:
|
|
*
|
|
* sum_(i,j=1..n,i!=j)((y_i - y_j) - (q_i - q_j))^2 = sum_(i,j=1..n,i!=j)((y_i - q_i) - (y_j - q_j))^2
|
|
*
|
|
* To simplify the following equations we use
|
|
*
|
|
* sum_(i,j=1..n,i!=j)(d_i - d_j)^2 = sum_(i,j=1..n,i!=j)(d_i^2 + d_j^2 - 2*d_i*d_j)
|
|
*
|
|
* Due to the pairings each element will appear as both d_i and d_j exactly n-1 times. This allows us to split the sum:
|
|
*
|
|
* sum_(i,j=1..n,i!=j)(d_i^2 + d_j^2 - 2*d_i*d_j) = 2*(n-1)*sum d_i^2 - 2 * sum_(i,j=1..n,i!=j) d_i * d_j
|
|
* = 2*((n-1) * sum d_i^2 - sum_(i,j=1..n,i!=j) d_i * d_j)
|
|
*
|
|
* Now we use the following equivalency:
|
|
*
|
|
* (sum d_i)^2 = sum d_i^2 + sum_(i,j=1..n,i!=j) d_i * d_j
|
|
*
|
|
* This allows us to now use sum d_i^2 and (sum d_i)^2 as a quick way to calculate the sum:
|
|
*
|
|
* (n-1) * sum d_i^2 - sum_(i,j=1..n,i!=j) d_i * d_j = n * sum d_i^2 - (sum d_i)^2
|
|
*
|
|
* And by substituting it into the original definition we get:
|
|
*
|
|
* 1/((n*(n-1))/2) * 2*(n * sum d_i^2 - (sum d_i)^2)
|
|
*
|
|
* Which can be again simplified to
|
|
*
|
|
* 4/(n*(n-1)) * (n * sum d_i^2 - (sum d_i)^2)
|
|
*
|
|
* After substituting d_i back to (y_i - q_i) this results in the function that we actually implement.
|
|
*
|
|
*/
|
|
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"
|
|
|
|
|
|
// input validation
|
|
REQUIRE_TRUE(labels->isSameShape(predictions), 0,
|
|
"MEAN_PAIRWSSQERR_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());
|
|
// only 4 possible reduction modes exist
|
|
REQUIRE_TRUE(reductionMode==0 || reductionMode==1 || reductionMode==2 || reductionMode==3, 0, "MEAN_PAIRWSSQERR_LOSS OP: reduction mode value is not acceptable, possible values are 0, 1, 2, 3, but got %i instead!", reductionMode);
|
|
|
|
if (labels->rankOf() == 1) { // If labels and predictions are of rank 1, it means that all data entries are 0-tensor (scalar) so that the result of becomes always zero.
|
|
*output = 0.;
|
|
return Status::OK();
|
|
}
|
|
|
|
std::vector<int> reductionIdx = ShapeUtils::evalDimsToExclude(labels->rankOf(), {0});
|
|
|
|
auto n = double(labels->sizeAt(1));
|
|
auto diffs = *predictions - *labels;
|
|
|
|
auto sumOfSquares = (diffs * diffs).reduceAlongDimension(reduce::Sum, reductionIdx, true);
|
|
|
|
auto squareOfSum = diffs.reduceAlongDimension(reduce::Sum, reductionIdx, true);
|
|
squareOfSum.applyScalar(scalar::Pow, 2, squareOfSum);
|
|
|
|
|
|
auto E = ((sumOfSquares * n) - squareOfSum) * (4/(n*(n-1)));
|
|
|
|
// 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() == E.rankOf(), 0, "MEAN_PAIRWSSQERR_LOSS_GRAD OP: weights array should be scalar or have the same rank as results array, but got %i and %i correspondingly!", weights->rankOf(), E.rankOf());
|
|
// check whether broadcast operation is possible for weights array
|
|
REQUIRE_TRUE(weights->isScalar() || ShapeUtils::areShapesBroadcastable(*weights, E), 0, "MEAN_PAIRWSSQERR_LOSS_GRAD OP: shapes of weights and labels arrays should be broadcastable, but got weights = %s and results = %s instead!", ShapeUtils::shapeAsString(weights).c_str(), ShapeUtils::shapeAsString(&E).c_str());
|
|
|
|
// perform weights broadcasting/tile to labels if needed
|
|
auto weightsBroad = weights;
|
|
if(!weights->isScalar() && !weights->isSameShape(E))
|
|
weightsBroad = new NDArray(weights->tileToShape(E.shapeInfo()));
|
|
|
|
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(mean_pairwssqerr_loss) {
|
|
|
|
getOpDescriptor()->setAllowedInputTypes(sd::DataType::ANY)->setAllowedOutputTypes({ALL_FLOATS});
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
DECLARE_SHAPE_FN(mean_pairwssqerr_loss) {
|
|
|
|
auto predictionsShapeInfo = inputShape->at(0);
|
|
auto weightsShapeInfo = inputShape->at(1);
|
|
auto labelsShapeInfo = inputShape->at(2);
|
|
|
|
REQUIRE_TRUE(shape::shapeEquals(labelsShapeInfo, predictionsShapeInfo), 0,
|
|
"MEAN_PAIRWSSQERR_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());
|
|
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 shape as labels and logits minus last dimension
|
|
std::vector<int> dimensions = {-1};
|
|
outShapeInfo = ShapeUtils::evalReduceShapeInfo(shape::order(predictionsShapeInfo), dimensions, predictionsShapeInfo, false, true, block.getWorkspace());
|
|
|
|
// weights array can be single scalar or has the same rank as output, and must be broadcastable to output
|
|
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || shape::rank(weightsShapeInfo) == shape::rank(outShapeInfo), 0, "MEAN_PAIRWSSQERR_LOSS OP: weights array should be scalar or have the same rank as output array, but got %i and %i correspondingly!", shape::rank(weightsShapeInfo), shape::rank(outShapeInfo));
|
|
// check whether broadcast operation is possible for weights array
|
|
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || ShapeUtils::areShapesBroadcastable(weightsShapeInfo, outShapeInfo), 0, "MEAN_PAIRWSSQERR_LOSS OP: shapes of weights and output arrays should be broadcastable, but got weights = %s and output = %s instead!", ShapeUtils::shapeAsString(weightsShapeInfo).c_str(), ShapeUtils::shapeAsString(outShapeInfo).c_str());
|
|
}
|
|
|
|
return SHAPELIST(outShapeInfo);
|
|
}
|
|
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
CUSTOM_OP_IMPL(mean_pairwssqerr_loss_grad, 3, 3, false, 0, 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
|
|
|
|
|
|
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, "MEAN_PAIRWSSQERR_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());
|
|
// only 4 possible reduction modes exist
|
|
REQUIRE_TRUE(reductionMode==0 || reductionMode==1 || reductionMode==2 || reductionMode==3, 0, "MEAN_PAIRWSSQERR_LOSS_GRAD OP: reduction mode value is not acceptable, possible values are 0, 1, 2, 3, but got %i instead!", reductionMode);
|
|
|
|
auto n = double(labels->sizeAt(1));
|
|
auto diffs = *predictions - *labels;
|
|
|
|
std::vector<int> reductionIdx = ShapeUtils::evalDimsToExclude(labels->rankOf(), {0});
|
|
auto sumOfSquares = (diffs * diffs).reduceAlongDimension(reduce::Sum, reductionIdx, true);
|
|
|
|
auto squareOfSum = diffs.reduceAlongDimension(reduce::Sum, reductionIdx, true);
|
|
squareOfSum.applyScalar(scalar::Pow, 2, squareOfSum);
|
|
|
|
auto E = ((sumOfSquares * n) - squareOfSum) * (4/(n*(n-1)));
|
|
|
|
auto sumPred = predictions->reduceAlongDimension(reduce::Sum, reductionIdx, true);
|
|
auto sumLabel = labels->reduceAlongDimension(reduce::Sum, reductionIdx, true);
|
|
|
|
dLdp->assign(((diffs * n) - sumPred + sumLabel)*(8/(n*(n-1))));
|
|
|
|
|
|
// 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() == E.rankOf(), 0, "MEAN_PAIRWSSQERR_LOSS_GRAD OP: weights array should be scalar or have the same rank as results array, but got %i and %i correspondingly!", weights->rankOf(), E.rankOf());
|
|
// check whether broadcast operation is possible for weights array
|
|
REQUIRE_TRUE(weights->isScalar() || ShapeUtils::areShapesBroadcastable(*weights, E), 0, "MEAN_PAIRWSSQERR_LOSS_GRAD OP: shapes of weights and labels arrays should be broadcastable, but got weights = %s and results = %s instead!", ShapeUtils::shapeAsString(weights).c_str(), ShapeUtils::shapeAsString(&E).c_str());
|
|
|
|
// perform weights broadcasting/tile to labels if needed
|
|
auto weightsBroad = weights;
|
|
if(!weights->isScalar() && !weights->isSameShape(E))
|
|
weightsBroad = new NDArray(weights->tileToShape(E.shapeInfo()));
|
|
|
|
switch (reductionMode) {
|
|
|
|
case 1: { // 1 - "none" and "weighted_sum", output is scalar and equal to sum of all elements of E array
|
|
|
|
*dLdp *= *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, false, false);
|
|
}
|
|
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.;
|
|
*dLdw = 0.;
|
|
}
|
|
else {
|
|
|
|
*dLdp *= *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, false, false);
|
|
}
|
|
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.;
|
|
*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, false, false);
|
|
*dLdw /= numOfNonZeroWeightsScalar;
|
|
}
|
|
else
|
|
dLdw->assign(E / numOfNonZeroWeightsScalar);
|
|
|
|
NDArray temp = *weightsBroad / numOfNonZeroWeightsScalar;
|
|
*dLdp *= temp;
|
|
}
|
|
break;
|
|
}
|
|
}
|
|
|
|
dLdl->assign(-*dLdp);
|
|
|
|
if(weightsBroad != weights)
|
|
delete weightsBroad;
|
|
|
|
return Status::OK();
|
|
}
|
|
|
|
DECLARE_TYPES(mean_pairwssqerr_loss_grad) {
|
|
getOpDescriptor()->setAllowedInputTypes(sd::DataType::ANY)->setAllowedOutputTypes({ALL_FLOATS});
|
|
}
|
|
|
|
DECLARE_SHAPE_FN(mean_pairwssqerr_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, "MEAN_PAIRWSSQERR_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, "MEAN_PAIRWSSQERR_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, "MEAN_PAIRWSSQERR_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));
|
|
|
|
Nd4jLong *dLdpShapeInfo = ShapeBuilders::copyShapeInfoAndType(predictionsShapeInfo, outType, false, block.getWorkspace());
|
|
Nd4jLong *dLdwShapeInfo = ShapeBuilders::copyShapeInfoAndType(weightsShapeInfo, outType, false, block.getWorkspace());
|
|
Nd4jLong *dLdlShapeInfo = ShapeBuilders::copyShapeInfoAndType(labelsShapeInfo, outType, false, block.getWorkspace());
|
|
|
|
return SHAPELIST(dLdpShapeInfo, dLdwShapeInfo, dLdlShapeInfo);
|
|
}
|
|
}
|
|
}
|
|
|
|
#endif
|
|
#pragma clang diagnostic pop
|