cavis/libnd4j/include/ops/declarable/generic/loss/softmaxCrossEntropy.cpp
raver119 320924278d
Legacy API changes (#441)
* initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* another initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* another initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* one more initial commit

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next step

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Refactored buffer() and shapeInfo() methods usage with NDArray class.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt Graph class methods to use const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt choose op to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt where op shape method to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt lstsq op to use constant empty shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt matrix_diag_part op shape routine to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt determinant ops to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt mean_pairwssqerr_loss ops to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt shape methods for loss ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt log_loss op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt shape methods for ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt dilation2d ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted deconv2d ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted dynamicRNN op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods for ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods for lstm layer ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* few updates

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* first cuda tweak

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Adopt constant shapes for sconv2d ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt constant shapes for gru ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt constant shapes with shape methods for segment ops and so on.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted constant shapes with unsorted_segment_* ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted constant shapes with gamma op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods of reduce_stddev ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape methods for reduce_* ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt shape method for squeeze op.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt strided_slice shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored concat op shape method to adopt constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted shape method for mirror_pad op.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted split op shape method.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted tile ops shape methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Added const cast for mkldnn routines handles.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored logSoftMaxForVector_ routine to conform with proper data and shape pointer casts.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Cosmetic changes to proper usage of constant pointers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored a couple shape comparators for strides and addBias helpers to proper use data pointers with inplace option.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored depthToSpace helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored histogram helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored im2col helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored gather and gatherND helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage on percentile helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed gather shape with helpers and range buffer usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with space to depth helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage and constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with LUP decomposition>

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored onehot_ helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pad and prefix to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactoed softmax helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed space to batch helpers to use buffers properly.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed stack and split helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with sparse to dense helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with mindistance_ helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with tile helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed constant shape usage with legacy pairwise bool ops.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored a couple of methods to adopt constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed broadcasting with constant shape."

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const usage with inplace reverse and constant shapes with legacy reduction.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored legacy ops with const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored sort to adopt constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected sort for constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed constant shape usage with special methods.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored Context to conform with constant shape usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* CUDA broadcasting headers

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* pairwise/indexreduce/random headers

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Refactored native ops to adopt constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* legacy reduce3/scalar headers

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Corrected pullRow signature and tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected routines to proper use of constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored tests to use constant shapes properly.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored legacy ops tests to use constant shapes properly.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored buffer usage with NDArray tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed native ops tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed special concat routine.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with test.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed buffer usage with a test.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored TAD.h and tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored calcStrides* routines to use constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed miscelaneous errors with constant shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* NativeOps const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Corrected definitions for declared functions.

Signed-off-by: shugeo <sgazeos@gmail.com>

* NativeOps const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* few more const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed const shapes with shape routines.

Signed-off-by: shugeo <sgazeos@gmail.com>

* few more const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed shape method for broadcastable case.

Signed-off-by: shugeo <sgazeos@gmail.com>

* few more const changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* xw_plus_b BP shape fn restored

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed signatures with broadcasting.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Repaired backprops shape methods for a set of operations.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored broadcast bool for cuda.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored methods for 3 args with const qualifier.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed a couple of kernel signatures for broadcasting.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernels signatures for const buffers and shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pairwise methods to persistent buffers and shapes usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt const to buffers and shapes with kernels.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopt const to buffers and shapes with scalar kernels.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored indexreduce kernels signatures to use const buffers and shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pairwise kernels to adopt cons shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored pairwise bool kernels to adopt cons shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored random special ops to conform with const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored native ops to conform with const shapes and buffers under cuda platform.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Cosmetical changes only.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shapes and buffers error.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected start pos routine.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored methods to conform with const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored helpers to use proper methods instead.

Signed-off-by: shugeo <sgazeos@gmail.com>

* bunch of changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next bunch of changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* next bunch of changes

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Fixed execScalar declaration.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed execScalar declaration.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected const shape cases with sort and so on.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shapes for sort.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored kernel declarations to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernels declarations to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected kernel declarations to adopt const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernels declarations to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed segment helpers kernels declarations and so on to adopt const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shape usage with segment and solve helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed kernel declaration with adjustWeight helper.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed cuda implementations for constant shape helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted const shape usage with kernels.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Adopted top_k kernels to use const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Corrected kernels declarations to adopt const shapes with helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored NDArray definitions to adopt const shapes and buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shapes with image suppression helpers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Slight improvement with buffers.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored buffer usage.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored buffer usage with tests.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Fixed const shape usage with definitions.

Signed-off-by: shugeo <sgazeos@gmail.com>

* minor updates on cpu side

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* Refactored const shape usage with ConstantDescritor and native ops with cuda platform.

Signed-off-by: shugeo <sgazeos@gmail.com>

* Refactored tear and tile kernels to adopt with const shapes.

Signed-off-by: shugeo <sgazeos@gmail.com>

* softmax_loop fix

Signed-off-by: raver119 <raver119@gmail.com>

* update missing signature

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* softmax again

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

* few more missing consts

Signed-off-by: raver119 <raver119@gmail.com>

* new methods updated

Signed-off-by: raver119@gmail.com <raver119@gmail.com>

Co-authored-by: shugeo <sgazeos@gmail.com>
2020-05-09 08:06:14 +03:00

398 lines
19 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 25.11.2017.
//
#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_softmax_cross_entropy_loss)
#include <ops/declarable/CustomOperations.h>
namespace sd {
namespace ops {
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(softmax_cross_entropy_loss, 3, 1, false, 1, 1) {
auto logits = 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"
double labelsSmoothing = T_ARG(0);
// input validation
REQUIRE_TRUE(labels->isSameShape(logits), 0, "SOFTMAX_CROSS_ENTROPY_LOSS OP: labels and logits arrays must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(labels).c_str(), ShapeUtils::shapeAsString(logits).c_str());
// only 4 possible reduction modes exist
REQUIRE_TRUE(reductionMode==0 || reductionMode==1 || reductionMode==2 || reductionMode==3, 0, "SOFTMAX_CROSS_ENTROPY_LOSS OP: reduction mode value is not acceptable, possible values are 0, 1, 2, 3, but got %i instead!", reductionMode);
// smoothing is possible for rank of logits/labels > 1
REQUIRE_TRUE(labels->rankOf() > 1 || (labels->rankOf() == 1 && labelsSmoothing == 0.), 0, "SOFTMAX_CROSS_ENTROPY_LOSS OP: smoothing is not possible when rank of labels/ logits = 1 !");
if(!output->isScalar()) {
// weights array can be single scalar or has the same shape as output, and must be broadcastable to output shape
REQUIRE_TRUE(weights->isScalar() || weights->rankOf() == output->rankOf(), 0, "SOFTMAX_CROSS_ENTROPY_LOSS OP: weights array should be scalar or have the same rank as output array, but got %i and %i correspondingly!", weights->rankOf(), output->rankOf());
// check whether broadcast operation is possible for weights array
REQUIRE_TRUE(weights->isScalar() || ShapeUtils::areShapesBroadcastable(*weights, *output), 0, "SOFTMAX_CROSS_ENTROPY_LOSS OP: shapes of weights and output arrays should be broadcastable, but got weights = %s and output = %s instead!", ShapeUtils::shapeAsString(weights).c_str(), ShapeUtils::shapeAsString(labels).c_str());
}
// If label_smoothing is nonzero, smooth the labels towards 1/num_classes: new_onehot_labels = onehot_labels * (1 - label_smoothing) + label_smoothing / num_classes
// num_classes = labels->sizeAt(1)
NDArray* cLabels = new NDArray(labels->cast(weights->dataType()));
NDArray* newLabels = cLabels;
if(labelsSmoothing != 0.) {
newLabels = new NDArray(cLabels);
newLabels->assign((1.f - labelsSmoothing) * *cLabels + labelsSmoothing / cLabels->sizeAt(1));
}
// main formula: result = - sum_i(lables_i * log(softmax_i)) - sum over last dimension
// softmax_i = exp(logits_i) / sum_j(exp(logits_j))
// so result = sum_i( lables_i * (log(sum_j(exp(logits_j))) - logits_i) )
// for numerical stability we use shifted logits (one can approve this using simple math):
// softmax_i = exp(logits_i - maxLogit) / sum_j(exp(logits_j - maxLogit))
// maxLogit is max among logits_i
std::vector<int> dimensions = {-1};
NDArray shiftedLogits = *logits - logits->reduceAlongDimension(reduce::Max, dimensions, true);
NDArray logSumExp = shiftedLogits.transform(transform::Exp).reduceAlongDimension(reduce::Sum, dimensions, true).transform(transform::Log);
NDArray E = (*newLabels * (logSumExp - shiftedLogits)).reduceAlongDimension(reduce::Sum, dimensions);
// perform weights broadcasting/tile to E if it is necessary
auto weightsBroad = weights;
if(!weights->isScalar() && !weights->isSameShape(&E)) {
if(E.rankOf() == 1 && weights->isVector() && weights->rankOf() > 1)
weightsBroad = new NDArray(weights->reshape(weights->ordering(), {weights->lengthOf()}));
else
weightsBroad = new NDArray(weights->tileToShape(E.shapeInfo()));
}
// 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
double sum;
if (weights->isScalar())
sum = weights->e<double>(0) * E.lengthOf();
else
sum = weightsBroad->reduceNumber(reduce::Sum).e<double>(0);
if (sum == 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;
if(newLabels != cLabels)
delete newLabels;
delete cLabels;
return Status::OK();
}
//////////////////////////////////////////////////////////////////////////
DECLARE_TYPES(softmax_cross_entropy_loss) {
getOpDescriptor()->setAllowedInputTypes(0, {ALL_FLOATS})
->setAllowedInputTypes(1, {ALL_FLOATS})
->setAllowedInputTypes(2, {ALL_FLOATS, ALL_INTS})
->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
DECLARE_SHAPE_FN(softmax_cross_entropy_loss) {
auto logitsShapeInfo = inputShape->at(0);
auto weightsShapeInfo = inputShape->at(1);
auto labelsShapeInfo = inputShape->at(2);
// labels and logits must have the same shapes
REQUIRE_TRUE(shape::shapeEquals(logitsShapeInfo, labelsShapeInfo), 0, "SOFTMAX_CROSS_ENTROPY_LOSS OP: labels and logits arrays must have the same shapes, but got %s and %s correspondingly!", ShapeUtils::shapeAsString(labelsShapeInfo).c_str(), ShapeUtils::shapeAsString(logitsShapeInfo).c_str());
DataType outType = DataTypeUtils::pickFloatingType(ArrayOptions::dataType(logitsShapeInfo));
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(logitsShapeInfo), dimensions, logitsShapeInfo, 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, "SOFTMAX_CROSS_ENTROPY_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, "SOFTMAX_CROSS_ENTROPY_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(softmax_cross_entropy_loss_grad, 3, 3, false, 1, 1) {
auto logits = INPUT_VARIABLE(0);
auto weights = INPUT_VARIABLE(1);
auto labels = INPUT_VARIABLE(2);
auto dLdp = OUTPUT_VARIABLE(0); // dL/dlogits
auto dLdw = OUTPUT_VARIABLE(1); // dL/dweights
auto dLdl = OUTPUT_VARIABLE(2); // dL/dlabels
auto labelsSmoothing = 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;
std::vector<int> dimensions = {-1};
// input validation
REQUIRE_TRUE(labels->isSameShape(logits), 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: labels and logits arrays must have the same shapes, but got %s and %s correspondingly !", ShapeUtils::shapeAsString(labels).c_str(), ShapeUtils::shapeAsString(logits).c_str());
// only 4 possible reduction modes exist
REQUIRE_TRUE(reductionMode==0 || reductionMode==1 || reductionMode==2 || reductionMode==3, 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: reduction mode value is not acceptable, possible values are 0, 1, 2, 3, but got %i instead!", reductionMode);
auto lossShapeInfo = ShapeUtils::evalReduceShapeInfo(logits->ordering(), dimensions, logits->shapeInfo(), false, false, block.getWorkspace());
// weights array can be single scalar or has the same shape as loss, and must be broadcastable to loss shape
REQUIRE_TRUE(weights->isScalar() || weights->rankOf() == shape::rank(lossShapeInfo), 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: weights array should be scalar or have the same rank as loss array, but got %i and %i correspondingly!", weights->rankOf(), shape::rank(lossShapeInfo));
// check whether broadcast operation is possible for weights array
REQUIRE_TRUE(weights->isScalar() || ShapeUtils::areShapesBroadcastable(weights->shapeInfo(), lossShapeInfo), 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: shapes of weights and loss arrays should be broadcastable, but got weights = %s and loss = %s instead!", ShapeUtils::shapeAsString(weights).c_str(), ShapeUtils::shapeAsString(lossShapeInfo).c_str());
// smoothing is possible for rank of logits/labels > 1
REQUIRE_TRUE(labels->rankOf() > 1 || (labels->rankOf() == 1 && labelsSmoothing == 0.), 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: smoothing is not possible when rank of labels/ logits = 1 !");
// If label_smoothing is nonzero, smooth the labels towards 1/num_classes: new_onehot_labels = onehot_labels * (1 - label_smoothing) + label_smoothing / num_classes
// num_classes = labels->sizeAt(1)
NDArray* cLabels = new NDArray(labels->cast(weights->dataType()));
NDArray* newLabels = cLabels;
if(labelsSmoothing != 0.) {
newLabels = new NDArray(labels->shapeInfo(), dLdl->dataType(), false, block.launchContext());
newLabels->assign((1.f - labelsSmoothing) * *cLabels + labelsSmoothing / cLabels->sizeAt(1));
}
NDArray softmax = (*logits - logits->reduceAlongDimension(reduce::Max, dimensions, true)).transform(transform::Exp);
softmax /= softmax.reduceAlongDimension(reduce::Sum, dimensions, true);
// dEdp = softmax * sum_i(lables_i) - labels
dLdp->assign(softmax * newLabels->reduceAlongDimension(reduce::Sum, dimensions, true) - *newLabels);
// dEdl = -log(softmax)
dLdl->assign(-softmax.transform(transform::Log)* (1.f - labelsSmoothing));
NDArray shiftedLogits = *logits - logits->reduceAlongDimension(reduce::Max, dimensions, true);
NDArray logSumExp = shiftedLogits.transform(transform::Exp).reduceAlongDimension(reduce::Sum, dimensions, true).transform(transform::Log);
NDArray E = (*newLabels * (logSumExp - shiftedLogits)).reduceAlongDimension(reduce::Sum, dimensions);
// perform weights broadcasting/tile to E if it is necessary
auto weightsBroad = weights;
if(!weights->isScalar() && !weights->isSameShape(&E))
weightsBroad = new NDArray(weights->tileToShape(E.shapeInfo()));
dimensions = ShapeUtils::evalDimsToExclude(dLdp->rankOf(), dimensions);
switch (reductionMode) {
case 1: { // 1 - "none" and "weighted_sum", output is scalar and equal to sum of all elements of E array
if(weights->isScalar() || weights->lengthOf() == 1) {
dLdw->assign(E.reduceNumber(reduce::Sum));
*dLdp *= *weights;
*dLdl *= *weights;
}
else {
dLdp->applyBroadcast(sd::broadcast::Multiply, dimensions, *weightsBroad, *dLdp);
dLdl->applyBroadcast(sd::broadcast::Multiply, dimensions, *weightsBroad, *dLdl);
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;
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 {
if(weights->isScalar() || weights->lengthOf() == 1) {
NDArray temp = *weights / sum;
*dLdp *= temp;
*dLdl *= temp;
*dLdw = 0.;
}
else {
NDArray temp = *weightsBroad / sum;
dLdp->applyBroadcast(sd::broadcast::Multiply, dimensions, temp, *dLdp);
dLdl->applyBroadcast(sd::broadcast::Multiply, dimensions, temp, *dLdl);
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.;
*dLdl = 0.;
*dLdw = 0.;
}
else {
if(weights->isScalar() || weights->lengthOf() == 1) {
NDArray temp = *weights / numOfNonZeroWeights;
*dLdp *= temp;
*dLdl *= temp;
dLdw->assign(E.reduceNumber(reduce::Sum) / numOfNonZeroWeights);
}
else {
NDArray temp = *weightsBroad / numOfNonZeroWeights;
dLdp->applyBroadcast(sd::broadcast::Multiply, dimensions, temp, *dLdp);
dLdl->applyBroadcast(sd::broadcast::Multiply, dimensions, temp, *dLdl);
if(weights != weightsBroad) {
std::vector<int> axesToReduceAlong = ShapeUtils::evalBroadcastBackwardAxis(weights->shapeInfo(), weightsBroad->shapeInfo());
E.reduceAlongDimension(reduce::Sum, *dLdw, axesToReduceAlong, true, false, false);
*dLdw /= numOfNonZeroWeights;
}
else
dLdw->assign(E / numOfNonZeroWeights);
}
}
break;
}
}
if(weightsBroad != weights)
delete weightsBroad;
if(newLabels != cLabels)
delete newLabels;
delete cLabels;
return Status::OK();
}
//////////////////////////////////////////////////////////////////////////
DECLARE_TYPES(softmax_cross_entropy_loss_grad) {
getOpDescriptor()->setAllowedInputTypes(0, {ALL_FLOATS})
->setAllowedInputTypes(1, {ALL_FLOATS})
->setAllowedInputTypes(2, {ALL_FLOATS, ALL_INTS})
->setAllowedInputTypes(3, {ALL_FLOATS})
->setAllowedInputTypes(4, {ALL_FLOATS})
->setAllowedInputTypes(5, {ALL_FLOATS})
->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
DECLARE_SHAPE_FN(softmax_cross_entropy_loss_grad) {
auto logitsShapeInfo = inputShape->at(0);
auto weightsShapeInfo = inputShape->at(1);
auto labelsShapeInfo = inputShape->at(2);
std::vector<int> dimensions = {-1};
// labels and logits must have the same shapes
REQUIRE_TRUE(shape::shapeEquals(logitsShapeInfo, labelsShapeInfo), 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: labels and logits arrays must have the same shapes, but got %s and %s correspondingly!", ShapeUtils::shapeAsString(labelsShapeInfo).c_str(), ShapeUtils::shapeAsString(logitsShapeInfo).c_str());
auto lossShapeInfo = ShapeUtils::evalReduceShapeInfo(shape::order(logitsShapeInfo), dimensions, logitsShapeInfo, false, false, block.getWorkspace());
// weights array can be single scalar or has the same rank as loss, and must be broadcastable to loss
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || shape::rank(weightsShapeInfo) == shape::rank(lossShapeInfo), 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: weights array should be scalar or have the same rank as loss array, but got %i and %i correspondingly!", shape::rank(weightsShapeInfo), shape::rank(lossShapeInfo));
// check whether broadcast operation is possible for weights array
REQUIRE_TRUE(shape::isScalar(weightsShapeInfo) || ShapeUtils::areShapesBroadcastable(weightsShapeInfo, lossShapeInfo), 0, "SOFTMAX_CROSS_ENTROPY_LOSS_GRAD OP: shapes of weights and loss arrays should be broadcastable, but got weights = %s and loss = %s instead!", ShapeUtils::shapeAsString(weightsShapeInfo).c_str(), ShapeUtils::shapeAsString(lossShapeInfo).c_str());
auto outType = DataTypeUtils::pickFloatingType(ArrayOptions::dataType(logitsShapeInfo));
auto dLdpShapeInfo = ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(outType, shape::order(logitsShapeInfo), shape::shapeOf(logitsShapeInfo), shape::rank(logitsShapeInfo)));
auto dLdwShapeInfo = ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(outType, shape::order(weightsShapeInfo), shape::shapeOf(weightsShapeInfo), shape::rank(weightsShapeInfo)));
auto dLdlShapeInfo = ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(outType, shape::order(labelsShapeInfo), shape::shapeOf(labelsShapeInfo), shape::rank(labelsShapeInfo)));
return SHAPELIST(dLdpShapeInfo, dLdwShapeInfo, dLdlShapeInfo);
}
}
}
#endif