cavis/libnd4j/include/ops/declarable/generic/nn/batchnorm.cpp

360 lines
15 KiB
C++
Raw Normal View History

2019-06-06 14:21:15 +02:00
/*******************************************************************************
* Copyright (c) 2015-2018 Skymind, Inc.
* Copyright (c) 2019 Konduit K.K.
2019-06-06 14:21:15 +02:00
*
* 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 raver119@gmail.com, created on 29/10/17.
// @author Yurii Shyrma (iuriish@yahoo.com)
//
#include <system/op_boilerplate.h>
2019-06-06 14:21:15 +02:00
#if NOT_EXCLUDED(OP_batchnorm)
#include <ops/declarable/CustomOperations.h>
#include<ops/declarable/helpers/batchnorm.h>
namespace sd {
2019-06-06 14:21:15 +02:00
namespace ops {
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(batchnorm, 3, 1, false, 1, 2) {
2019-06-06 14:21:15 +02:00
auto input = INPUT_VARIABLE(0);
auto mean = INPUT_VARIABLE(1);
auto variance = INPUT_VARIABLE(2);
NDArray* gamma = nullptr;
NDArray* beta = nullptr;
auto output = OUTPUT_VARIABLE(0);
const bool applyScale = (bool)INT_ARG(0);
const bool applyOffset = (bool)INT_ARG(1);
const double epsilon = T_ARG(0);
if(applyScale)
gamma = INPUT_VARIABLE(3);
if(applyOffset)
beta = INPUT_VARIABLE(3 + (int)applyScale);
2019-06-06 14:21:15 +02:00
const int numOfIntArgs = block.getIArguments()->size();
const int inRank = input->rankOf();
// get axes args to normalize input array over
std::vector<int> axes;
if(numOfIntArgs > 2)
for(int i = 2; i < numOfIntArgs; ++i)
axes.push_back(INT_ARG(i));
else
axes.push_back(inRank-1); // default dimension to reduce along is last dimension
const uint numOfAxes = axes.size();
REQUIRE_TRUE(numOfAxes <= inRank, 0, "BATCHNORM op: too big number of input axes to normalize over, expected number should be less or equal to rank of input array, but got %i and %i correspondingly !", numOfAxes, inRank);
2019-06-06 14:21:15 +02:00
// evaluate expected shape for mean, variance and gamma. These 3 arrays should have identical shapes
// for example if input shape is {2,3,4,5,6} and axes = {1,3}, then expected shape would be {1,3,1,5,1}, and if axes = {3}, then expected shape would be {5}
std::vector<Nd4jLong> expShape;
if(numOfAxes == 1)
expShape.push_back(input->sizeAt(axes[0]));
else { // get, for example, something like {1, inputDim1, 1, inputDim3, 1} if axes = {1, 3}
expShape = std::vector<Nd4jLong>(inRank, 1);
for(uint i = 0; i < numOfAxes; ++i)
expShape[axes[i]] = input->sizeAt(axes[i]);
}
2019-06-06 14:21:15 +02:00
REQUIRE_TRUE(mean->isSameShape(expShape) , 0, "BATCHNORM op: wrong shape of mean array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(mean).c_str());
REQUIRE_TRUE(variance->isSameShape(expShape), 0, "BATCHNORM op: wrong shape of variance array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(variance).c_str());
2019-06-06 14:21:15 +02:00
if(gamma)
REQUIRE_TRUE(gamma->isSameShape(expShape), 0, "BATCHNORM op: wrong shape of gamma array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(gamma).c_str());
2019-06-06 14:21:15 +02:00
if(beta)
REQUIRE_TRUE(beta->isSameShape(expShape), 0, "BATCHNORM op: wrong shape of beta array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(beta).c_str());
2019-06-06 14:21:15 +02:00
// types of all input arrays should be the same
for(unsigned long i = 1; i < block.width(); ++i)
REQUIRE_TRUE(INPUT_VARIABLE(0)->dataType() == INPUT_VARIABLE(i)->dataType(), 0, "BATCHNORM op: types of all input arrays should be the same !");
2019-06-06 14:21:15 +02:00
nd4j_debug("MKL-DNN is not used for batchnorm!\n", 0);
2019-06-06 14:21:15 +02:00
// formula: output = gamma * ((input - mean) / sqrt(variance + epsilon)) + beta
// auto v = input->varianceAlongDimension(variance::SummaryStatsVariance, false, ShapeUtils::evalDimsToExclude(input->rankOf(), axes));
// auto m = input->reduceAlongDimension(sd::reduce::Mean, ShapeUtils::evalDimsToExclude(input->rankOf(), axes));
2019-06-06 14:21:15 +02:00
helpers::batchnorm(input, mean, variance, gamma, beta, output, axes, epsilon);
// NDArray stdInv = *v + epsilon;
// stdInv.applyTransform(transform::Reciprocal); // 1 / (variance + epsilon)
// stdInv.applyTransform(transform::Sqrt); // 1 / (variance + epsilon)^0.5
// if(applyScale)
// stdInv *= *gamma;
// // empty array with same shape as input
// input->applyBroadcast(sd::broadcast::Subtract, axes, m, output);
// output->applyBroadcast(sd::broadcast::Multiply, axes, &stdInv);
// if(applyOffset)
// output->applyBroadcast(sd::broadcast::Add, axes, beta);
// delete v;
// delete m;
2019-06-06 14:21:15 +02:00
return Status::OK();
}
DECLARE_TYPES(batchnorm) {
2019-06-06 14:21:15 +02:00
getOpDescriptor()->setAllowedInputTypes({ALL_FLOATS})->setSameMode(true);
}
DECLARE_SHAPE_FN(batchnorm) {
2019-06-06 14:21:15 +02:00
auto inShapeInfo = inputShape->at(0);
DataType outType = DataTypeUtils::pickFloatingType(ArrayOptions::dataType(inShapeInfo));
2019-06-06 14:21:15 +02:00
auto outShapeInfo = ShapeBuilders::copyShapeInfoAndType(inShapeInfo, outType, false, block.getWorkspace()); // output shape is identical to input shape
return SHAPELIST(CONSTANT(outShapeInfo));
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(batchnorm_bp, 4, 3, false, 1, 2) {
NDArray* input = INPUT_VARIABLE(0);
NDArray* mean = INPUT_VARIABLE(1);
NDArray* variance = INPUT_VARIABLE(2);
NDArray* gamma = nullptr;
NDArray* beta = nullptr;
NDArray* dLdO = INPUT_VARIABLE(block.width() - 1); // next epsilon
2019-06-06 14:21:15 +02:00
NDArray* dLdI = OUTPUT_VARIABLE(0);
NDArray* dLdM = OUTPUT_VARIABLE(1);
NDArray* dLdV = OUTPUT_VARIABLE(2);
NDArray* dLdG = nullptr;
NDArray* dLdB = nullptr;
2019-06-06 14:21:15 +02:00
const bool applyScale = (bool)INT_ARG(0);
const bool applyOffset = (bool)INT_ARG(1);
const float epsilon = T_ARG(0);
2019-06-06 14:21:15 +02:00
if(applyScale) {
gamma = INPUT_VARIABLE(3);
2019-06-06 14:21:15 +02:00
dLdG = OUTPUT_VARIABLE(3);
}
if(applyOffset) {
beta = INPUT_VARIABLE(3 + (int)applyScale);
dLdB = OUTPUT_VARIABLE(3 + (int)applyScale);
2019-06-06 14:21:15 +02:00
}
const int numOfIntArgs = block.getIArguments()->size();
const int inRank = input->rankOf();
// get axes args to normalize input array over
std::vector<int> axes;
if(numOfIntArgs > 2)
for(int i = 2; i < numOfIntArgs; ++i)
axes.push_back(INT_ARG(i));
else
axes.push_back(inRank-1); // default dimension to reduce along is last dimension
const uint numOfAxes = axes.size();
REQUIRE_TRUE(numOfAxes <= inRank, 0, "BATCHNORM_BP op: too big number of input axes to normalize over, expected number should be less or equal to rank of input array, but got %i and %i correspondingly !", numOfAxes, inRank);
// evaluate expected shape for mean, variance and gamma. These 3 arrays should have identical shapes
// for example if input shape is {2,3,4,5,6} and axes = {1,3}, then expected shape would be {1,3,1,5,1}, and if axes = {3}, then expected shape would be {5}
std::vector<Nd4jLong> expShape;
if(numOfAxes == 1)
expShape.push_back(input->sizeAt(axes[0]));
else { // get, for example, something like {1, inputDim1, 1, inputDim3, 1} if axes = {1, 3}
expShape = std::vector<Nd4jLong>(inRank, 1);
for(uint i = 0; i < numOfAxes; ++i)
expShape[axes[i]] = input->sizeAt(axes[i]);
}
REQUIRE_TRUE(mean->isSameShape(expShape), 0, "BATCHNORM_BP op: wrong shape of mean array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(mean).c_str());
REQUIRE_TRUE(variance->isSameShape(expShape), 0, "BATCHNORM_BP op: wrong shape of variance array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(variance).c_str());
if(gamma)
REQUIRE_TRUE(gamma->isSameShape(expShape), 0, "BATCHNORM_BP op: wrong shape of gamma array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(gamma).c_str());
if(beta)
REQUIRE_TRUE(beta->isSameShape(expShape), 0, "BATCHNORM_BP op: wrong shape of beta array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(expShape).c_str(), ShapeUtils::shapeAsString(beta).c_str());
REQUIRE_TRUE(input->isSameShape(dLdO), 0, "BATCHNORM_BP op: wrong shape of output gradients array, expected is %s, but got %s instead !", ShapeUtils::shapeAsString(input).c_str(), ShapeUtils::shapeAsString(dLdO).c_str());
// types of all input arrays should be the same (except dLdO)
for(unsigned long i = 1; i < block.width() - 2; ++i)
REQUIRE_TRUE(INPUT_VARIABLE(0)->dataType() == INPUT_VARIABLE(i)->dataType(), 0, "BATCHNORM_BP op: types of arrays (input, mean, variance, gamma, beta) should be the same !");
2019-06-06 14:21:15 +02:00
// ***** calculations ***** //
// notations:
// f = g * (gamma * ((x - m) / (v + eps)^0.5) + beta) -> means dLdO * ff_output, g = dLdO
// stdInv = 1 / (v + eps)^0.5
// N - batch size (product of spatial dimensions)
// derivatives:
// dLdI = dfdx + dfdm*dmdx + dfdv*(dvdm*dmdx + dvdx)
// dfdx = gamma*stdInv*g;
// dfdm = -gamma*stdInv*g_sum;
// dmdx = 1/N;
// dvdx = 2 * (x - m) / N
// dvdm = -2 * [(x - m)]_sum / N
// dfdv = -0.5 * [g*(x - m)]_sum * stdInv^3, drop gamma here for calc convenience
// finally:
// dLdI = gamma * ( stdInv * (g - g_sum/N) + (2/N) * dfdv * (dvdm/2 + (x - m)) )
// dLdG = (g * (x - m))_sum * stdInv
// dLdB = g_sum
2019-06-06 14:21:15 +02:00
// variance = input->varianceAlongDimension(variance::SummaryStatsVariance, false, ShapeUtils::evalDimsToExclude(input->rankOf(), axes));
// mean = input->reduceAlongDimension(sd::reduce::Mean, ShapeUtils::evalDimsToExclude(input->rankOf(), axes));
const auto excludedAxes = ShapeUtils::evalDimsToExclude(inRank, axes);
const bool keepUnitiesInShape = inRank == mean->rankOf();
2019-06-06 14:21:15 +02:00
// inverse batch size 1/N
const float Ninv = 1.f * shape::tadLength(input->getShapeInfo(), axes.data(), axes.size()) / input->lengthOf();
// input - mean
NDArray xMinusMean(input); // empty array with same shape as input
input->applyBroadcast(sd::broadcast::Subtract, axes, *mean, xMinusMean);
2019-06-06 14:21:15 +02:00
// stdInv
NDArray stdInv = *variance + epsilon;
Shyrma temp (#131) * - specifying template instantiation for certain types in float16 and bloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - polishing bfloat16 and float16 member functions template specialization Signed-off-by: Yurii <iuriish@yahoo.com> * - rewrite and overload array +-*/ scalar and scalar +-*/ arr in NDAray class Signed-off-by: Yurii <iuriish@yahoo.com> * - make corrections which have to do with and rvalue lvalue conversions Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantic in NDArray operators array +-/* array Signed-off-by: Yurii <iuriish@yahoo.com> * float16/bfloat16 tweaks Signed-off-by: raver119 <raver119@gmail.com> * one more tweak Signed-off-by: raver119 <raver119@gmail.com> * - make float16 and bfloat16 to compile successfully on cuda Signed-off-by: Yurii <iuriish@yahoo.com> * - do not use resources of view-like arrays when move semantics is applied Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of pointers in signatures NDArray methods 1 Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::dup method Signed-off-by: Yurii <iuriish@yahoo.com> * - correction of signature of NDArray::reduceAlongDimension method Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyIndexReduce and applyTrueBroadcast methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyReduce3 and varianceAlongDimension methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tensorsAlongDimension and diagonal methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::allTensorsAlongDimension Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduceAlongDimension 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyPairwiseTransform 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyTrueBroadcast 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::applyScalar and applyScalarArr Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::lambda methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::reduce3 methods 2 Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of following NDArray methods: add/sub/mul/div row/column and fillAsTriangular Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::tileToShape methods Signed-off-by: Yurii <iuriish@yahoo.com> * - signature correction of NDArray::isShapeSameStrict method Signed-off-by: Yurii <iuriish@yahoo.com> * minor corrections in tests Signed-off-by: Yurii <iuriish@yahoo.com> * - replace reduce op in batchnorm mkldnn Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit templates instantiations for operator+(NDArray&&. const scalar) Signed-off-by: Yurii <iuriish@yahoo.com> * - corrections of casts in float16/bfloat16 Signed-off-by: Yurii <iuriish@yahoo.com> * - provide move semantics in following NDArray methods: transform, applyTrueBroadcast, transpose, reshape, permute Signed-off-by: Yurii <iuriish@yahoo.com> * - get rid of input array A duplicate in svd cuda op Signed-off-by: Yurii <iuriish@yahoo.com> * - avoid available bug in svd cuda API Signed-off-by: Yurii <iuriish@yahoo.com> * - add temporary global memory buffer in svd cuda when calcUV = false and m != n Signed-off-by: Yurii <iuriish@yahoo.com> * - remove test with blfoat16 type for betainC Signed-off-by: Yurii <iuriish@yahoo.com> * - resolve conflicts after master has been merged in Signed-off-by: Yurii <iuriish@yahoo.com> * - changed type of affected input array in fused_batch_norm Signed-off-by: Yurii <iuriish@yahoo.com> * - add several explicit type castings Signed-off-by: Yurii <iuriish@yahoo.com> * - add ND4J_EXPORT to operators Signed-off-by: Yurii <iuriish@yahoo.com> * - add explicit template types in instantiations of template arithm operators of NDArray class Signed-off-by: Yurii <iuriish@yahoo.com> * - one more test fix Signed-off-by: Yurii <iuriish@yahoo.com> Co-authored-by: raver119 <raver119@gmail.com>
2019-12-20 20:35:39 +01:00
stdInv.applyTransform(transform::Reciprocal, stdInv); // 1 / (variance + epsilon)
stdInv.applyTransform(transform::Sqrt, stdInv); // 1 / (variance + epsilon)^0.5
// dvdm (use dLdM as storage for dvdm)
xMinusMean.reduceAlongDimension(sd::reduce::Sum, *dLdM, excludedAxes, keepUnitiesInShape);
*dLdM *= -Ninv;
// g_sum
auto gSum = dLdO->reduceAlongDimension(sd::reduce::Sum, excludedAxes, keepUnitiesInShape);
2019-06-06 14:21:15 +02:00
// dLdB
if(applyOffset)
dLdB->assign(gSum);
// stdInv * (g - g_sum/N) (use dLdI as storage for this expression)
gSum *= Ninv;
dLdO->applyBroadcast(sd::broadcast::Subtract, axes, gSum, *dLdI);
dLdI->applyBroadcast(sd::broadcast::Multiply, axes, stdInv, *dLdI);
// dLdV <- [g*(x - m)]_sum
(xMinusMean * *dLdO).reduceAlongDimension(sd::reduce::Sum, *dLdV, excludedAxes, keepUnitiesInShape);
2019-06-06 14:21:15 +02:00
// dLdG
*dLdV *= stdInv;
if(applyScale)
dLdG->assign(dLdV);
2019-06-06 14:21:15 +02:00
// (2 / N) * dfdv (use dLdV as storage for dfdv)
*dLdV *= stdInv*stdInv; // dLdV*stdInv * stdInv^2
*dLdV *= -Ninv; // -0.5f * (2 / N);
// dfdv * (dvdm + (x - m)) (use xMinusMean as storage for this expression)
xMinusMean.applyBroadcast(sd::broadcast::Add, axes, *dLdM, xMinusMean);
xMinusMean.applyBroadcast(sd::broadcast::Multiply, axes, *dLdV, xMinusMean);
// dLdI
*dLdI += xMinusMean;
if(applyScale)
dLdI->applyBroadcast(sd::broadcast::Multiply, axes, *gamma, *dLdI);
*dLdM = 0; // put zeros so far
*dLdV = 0; // put zeros so far
2019-06-06 14:21:15 +02:00
// java code
// NDArray std = *variance + epsilon;
// std.applyTransform(transform::Reciprocal); // 1 / (variance + epsilon)
// std.applyTransform(transform::Sqrt); // 1 / (variance + epsilon)^0.5
// NDArray xMu(input);
// input->applyBroadcast(sd::broadcast::Subtract, axes, mean, &xMu);
// NDArray xHat(input);
// xMu.applyBroadcast(sd::broadcast::Multiply, axes, &std, &xHat);
// NDArray dxhat(input);
// dLdO->applyBroadcast(sd::broadcast::Multiply, axes, gamma, &dxhat);
// NDArray temp = dxhat*xMu;
// temp.reduceAlongDimension(reduce::Sum, dLdV, excludedAxes, keepUnitiesInShape);
// *dLdV *= -0.5f * std*std*std;
// NDArray* dxmu1 = dxhat.reduceAlongDimension(reduce::Sum, excludedAxes, keepUnitiesInShape);
// *dxmu1 *= -std;
// NDArray* dxmu2 = xMu.reduceAlongDimension(reduce::Sum, excludedAxes, keepUnitiesInShape);
// *dxmu2 *= *dLdV * (-2.f/N);
// NDArray dLdmu = *dxmu1 + *dxmu2;
// dLdmu *= (1.f /N);
// *dLdV *= (2.f/N);
// dxhat.applyBroadcast(sd::broadcast::Multiply, axes, &std);
// xMu.applyBroadcast(sd::broadcast::Multiply, axes, dLdV);
// dxhat += xMu;
// dxhat.applyBroadcast(sd::broadcast::Add, axes, &dLdmu, dLdI);
// delete dxmu1;
// delete dxmu2;
// xHat *= *dLdO;
// xHat.reduceAlongDimension(reduce::Sum, dLdG, excludedAxes, keepUnitiesInShape);
2019-06-06 14:21:15 +02:00
return Status::OK();
}
DECLARE_TYPES(batchnorm_bp) {
getOpDescriptor()
->setAllowedInputTypes(0, sd::DataType::ANY)
->setAllowedInputTypes(1, sd::DataType::ANY)
->setAllowedInputTypes(2, sd::DataType::ANY)
->setAllowedInputTypes(3, {ALL_FLOATS})
->setAllowedInputTypes(4, sd::DataType::ANY)
->setAllowedInputTypes(5, sd::DataType::ANY)
->setAllowedOutputTypes({ALL_FLOATS});
}
2019-06-06 14:21:15 +02:00
//////////////////////////////////////////////////////////////////////////
DECLARE_SHAPE_FN(batchnorm_bp) {
Nd4jLong* inShapeInfo = inputShape->at(0);
Nd4jLong* meanShapeInfo = inputShape->at(1);
2019-06-06 14:21:15 +02:00
const bool applyScale = (bool)INT_ARG(0);
const bool applyOffset = (bool)INT_ARG(1);
DataType outType = DataTypeUtils::pickFloatingType(ArrayOptions::dataType(inShapeInfo));
2019-06-06 14:21:15 +02:00
auto shapes = SHAPELIST();
2019-06-06 14:21:15 +02:00
// dLdI shapeInfo
shapes->push_back(ConstantShapeHelper::getInstance()->createShapeInfo(outType, inShapeInfo));
2019-06-06 14:21:15 +02:00
// dLdM shapeInfo
shapes->push_back(ConstantShapeHelper::getInstance()->createShapeInfo(outType, meanShapeInfo));
2019-06-06 14:21:15 +02:00
// dLdV shapeInfo (same as dLdM)
shapes->push_back(shapes->at(shapes->size()-1));
2019-06-06 14:21:15 +02:00
// dLdG shapeInfo (same as dLdM)
if(applyScale)
shapes->push_back(shapes->at(shapes->size()-1));
2019-06-06 14:21:15 +02:00
// dLdB shapeInfo (same as dLdM)
if(applyOffset)
shapes->push_back(shapes->at(shapes->size()-1));
2019-06-06 14:21:15 +02:00
return shapes;
2019-06-06 14:21:15 +02:00
}
}
}
#endif