Brutex
/
cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
cavis/libnd4j/include/ops/declarable/generic/nn/batchnorm.cpp

603 lines
28 KiB
C++
Raw Blame History

/*******************************************************************************
* 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 raver119@gmail.com, created on 29/10/17.
// @author Yurii Shyrma (iuriish@yahoo.com)
//
#include <op_boilerplate.h>
#if NOT_EXCLUDED(OP_batchnorm)
#include <ops/declarable/CustomOperations.h>
#include<ops/declarable/helpers/batchnorm.h>
namespace nd4j {
namespace ops {
#ifdef HAVE_MKLDNN
using namespace mkldnn;
static void getMKLDNNMemoryDescBatchNorm(const NDArray* src, const NDArray* diff_src, const NDArray* dst,
mkldnn::memory::desc* batchnorm_src_md, mkldnn::memory::desc* batchnorm_diff_src_md, mkldnn::memory::desc* batchnorm_dst_md,
mkldnn::memory::desc* user_src_md, mkldnn::memory::desc* user_diff_src_md, mkldnn::memory::desc* user_dst_md, int axis) {
const Nd4jLong* shape = src->getShapeInfo();
Nd4jLong rank = shape[0];
Nd4jLong dim1 = axis; // MKL-DNN supports only 1 axis, which has to be the "channel" one
Nd4jLong dim2 = axis >= 2 ? 1 : 2;
Nd4jLong dim3 = axis >= 3 ? 2 : 3;
mkldnn::memory::dims batchnorm_src_tz = { (int)shape[1], (int)shape[dim1 + 1], rank > 2 ? (int)shape[dim2 + 1] : 1, rank > 3 ? (int)shape[dim3 + 1] : 1};
auto type = mkldnn::memory::data_type::f32;
auto format = mkldnn::memory::format::nchw;
auto supposed_to_be_any_format = mkldnn::memory::format::nChw8c; // doesn't work with "any"
if (src != nullptr && src->getBuffer() != nullptr && batchnorm_src_md != nullptr) {
*batchnorm_src_md = mkldnn::memory::desc({ batchnorm_src_tz }, type, supposed_to_be_any_format);
*user_src_md = mkldnn::memory::desc({ batchnorm_src_tz }, type, format);
user_src_md->data.format = mkldnn_blocked; // overrides format
user_src_md->data.layout_desc.blocking.strides[0][0] = src->stridesOf()[0];
user_src_md->data.layout_desc.blocking.strides[0][1] = src->stridesOf()[dim1];
user_src_md->data.layout_desc.blocking.strides[0][2] = rank > 2 ? src->stridesOf()[dim2] : 1;
user_src_md->data.layout_desc.blocking.strides[0][3] = rank > 3 ? src->stridesOf()[dim3] : 1;
}
if (diff_src != nullptr && diff_src->getBuffer() != nullptr && batchnorm_diff_src_md != nullptr) {
*batchnorm_diff_src_md = mkldnn::memory::desc({ batchnorm_src_tz }, type, supposed_to_be_any_format);
*user_diff_src_md = mkldnn::memory::desc({ batchnorm_src_tz }, type, format);
user_diff_src_md->data.format = mkldnn_blocked; // overrides format
user_diff_src_md->data.layout_desc.blocking.strides[0][0] = diff_src->stridesOf()[0];
user_diff_src_md->data.layout_desc.blocking.strides[0][1] = diff_src->stridesOf()[dim1];
user_diff_src_md->data.layout_desc.blocking.strides[0][2] = rank > 2 ? diff_src->stridesOf()[dim2] : 1;
user_diff_src_md->data.layout_desc.blocking.strides[0][3] = rank > 3 ? diff_src->stridesOf()[dim3] : 1;
}
if (dst != nullptr && dst->getBuffer() != nullptr && batchnorm_dst_md != nullptr) {
*batchnorm_dst_md = mkldnn::memory::desc({ batchnorm_src_tz }, type, supposed_to_be_any_format);
*user_dst_md = mkldnn::memory::desc({ batchnorm_src_tz }, type, format);
user_dst_md->data.format = mkldnn_blocked; // overrides format
user_dst_md->data.layout_desc.blocking.strides[0][0] = dst->stridesOf()[0];
user_dst_md->data.layout_desc.blocking.strides[0][1] = dst->stridesOf()[dim1];
user_dst_md->data.layout_desc.blocking.strides[0][2] = rank > 2 ? dst->stridesOf()[dim2] : 1;
user_dst_md->data.layout_desc.blocking.strides[0][3] = rank > 3 ? dst->stridesOf()[dim3] : 1;
}
}
#endif
CUSTOM_OP_IMPL(batchnorm, 3, 1, false, 1, 2) {
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);
// FIXME: double?
const double epsilon = T_ARG(0);
if(applyScale)
gamma = INPUT_VARIABLE(3);
if(applyOffset)
beta = INPUT_VARIABLE(3 + static_cast<int>(applyScale));
std::vector<const NDArray*> inArrs(block.width());
for(int i = 0; i < block.width(); ++i)
inArrs[i] = INPUT_VARIABLE(i);
// check whether all input shapes are mutually broadcastable
Nd4jLong* outShapeInfo = nullptr;
const bool areShapesOk = ShapeUtils::evalCommonBroadcastShapeInfo(inArrs, outShapeInfo, block.getWorkspace());
REQUIRE_TRUE(areShapesOk, 0, "BATCHNORM op: the shapes of input arrays are not mutually broadcastable !");
// normalized output = gamma * ((input - mean) / sqrt(variance + epsilon)) + beta
auto sigmaInvGam = (*variance + epsilon).transform(transform::RSqrt);
if(applyScale)
sigmaInvGam *= *gamma;
NDArray inputMinusMean;
if(!input->isSameShape(output) && !mean->isSameShape(output)) {
auto inputTiled = NDArray(output, false, block.launchContext());
input->tile(inputTiled);
inputMinusMean = inputTiled - *mean;
}
else
inputMinusMean = *input - *mean;
if (applyOffset)
output->assign(inputMinusMean * sigmaInvGam + *beta);
else
output->assign(inputMinusMean * sigmaInvGam);
return Status::OK();
}
DECLARE_TYPES(batchnorm) {
getOpDescriptor()
->setAllowedInputTypes(nd4j::DataType::ANY)
->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
DECLARE_SHAPE_FN(batchnorm) {
std::vector<const NDArray*> inArrs(block.width());
auto in = inputShape->at(0);
for(int i = 0; i < block.width(); ++i)
inArrs[i] = INPUT_VARIABLE(i);
// check whether all input shapes are mutually broadcastable
Nd4jLong* outShapeInfo = nullptr;
const bool areShapesOk = ShapeUtils::evalCommonBroadcastShapeInfo(inArrs, outShapeInfo, block.getWorkspace());
REQUIRE_TRUE(areShapesOk, 0, "BATCHNORM op: the shapes of input arrays are not mutually broadcastable !");
auto result = ConstantShapeHelper::getInstance()->createShapeInfo(ShapeDescriptor(outShapeInfo, DataTypeUtils::pickFloatingType(ArrayOptions::dataType(in))));
return SHAPELIST(result);
}
//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(batchnorm_new, 3, 1, false, 1, 2) {
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 + static_cast<int>(applyScale));
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 int numOfAxes = axes.size();
REQUIRE_TRUE(numOfAxes <= inRank, 0, "BATCHNORM_NEW 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);
// get, for example, something like {1, inDim1, 1, inDim3, 1} if axes = {1, 3}
std::vector<Nd4jLong> expShapeWithUnities(inRank, 1);
for(int i = 0; i < numOfAxes; ++i)
expShapeWithUnities[axes[i]] = input->sizeAt(axes[i]);
// 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 = numOfAxes == 1 ? std::vector<Nd4jLong>(1, input->sizeAt(axes[0])) : expShapeWithUnities;
std::string expShapeStr = ShapeUtils::shapeAsString(expShape);
REQUIRE_TRUE(ShapeUtils::shapeAsString(mean) == expShapeStr, 0, "BATCHNORM_NEW op: wrong shape of mean array, expected is %s, but got %s instead !", expShapeStr.c_str(), ShapeUtils::shapeAsString(mean).c_str());
REQUIRE_TRUE(ShapeUtils::shapeAsString(variance) == expShapeStr, 0, "BATCHNORM_NEW op: wrong shape of variance array, expected is %s, but got %s instead !", expShapeStr.c_str(), ShapeUtils::shapeAsString(variance).c_str());
if(gamma)
REQUIRE_TRUE(ShapeUtils::shapeAsString(gamma) == expShapeStr, 0, "BATCHNORM_NEW op: wrong shape of gamma array, expected is %s, but got %s instead !", expShapeStr.c_str(), ShapeUtils::shapeAsString(gamma).c_str());
if(beta)
REQUIRE_TRUE(ShapeUtils::shapeAsString(beta) == expShapeStr, 0, "BATCHNORM_NEW op: wrong shape of beta array, expected is %s, but got %s instead !", expShapeStr.c_str(), ShapeUtils::shapeAsString(beta).c_str());
// types of all input arrays should be the same
for(int i = 1; i < block.width(); ++i)
REQUIRE_TRUE(INPUT_VARIABLE(0)->dataType() == INPUT_VARIABLE(i)->dataType(), 0, "BATCHNORM_NEW op: types of all input arrays should be the same !");
#ifdef HAVE_MKLDNN
if (block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported({input, mean, variance, gamma, beta, output}) && numOfAxes == 1) {
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
if (streams.empty()) {
streams.push_back(MKLDNNStream("batchnorm_new"));
}
std::vector<Nd4jLong> shape({2, mean->lengthOf()});
NDArray weights = NDArrayFactory::create<float>('c', shape, block.launchContext());
weights({0, 1, 0, 0}).assign(1.0f);
weights({1, 2, 0, 0}).assign(0.0f);
if (streams[0].checkAndReset({input, mean, variance, gamma, beta}, {output}, {(float)epsilon}, axes)) {
mkldnn_memory_desc_t empty;
mkldnn::memory::desc batchnorm_src_md(empty), batchnorm_dst_md(empty), user_src_md(empty), user_dst_md(empty);
getMKLDNNMemoryDescBatchNorm(input, nullptr, output,
&batchnorm_src_md, nullptr, &batchnorm_dst_md,
&user_src_md, nullptr, &user_dst_md, axes[0]);
auto batchnorm_desc = batch_normalization_forward::desc(prop_kind::forward_inference, batchnorm_src_md, epsilon,
use_global_stats | (applyScale || applyOffset ? use_scale_shift : 0));
auto engine = streams[0].getEngine();
auto batchnorm_prim_desc = batch_normalization_forward::primitive_desc(batchnorm_desc, engine);
auto user_src_memory = mkldnn::memory({user_src_md, engine}, input->buffer());
auto user_dst_memory = mkldnn::memory({user_dst_md, engine}, output->buffer());
auto batchnorm_mean_memory = mkldnn::memory(batchnorm_prim_desc.mean_primitive_desc(), mean->buffer());
auto batchnorm_variance_memory = mkldnn::memory(batchnorm_prim_desc.variance_primitive_desc(), variance->buffer());
auto batchnorm_src_memory = user_src_memory;
streams[0].addMemory(user_src_memory);
if (mkldnn::memory::primitive_desc({batchnorm_src_md, engine})
!= user_src_memory.get_primitive_desc()) {
batchnorm_src_memory = mkldnn::memory({batchnorm_src_md, engine});
streams[0].addMemory(batchnorm_src_memory);
streams[0].addOperation(reorder(user_src_memory, batchnorm_src_memory));
}
auto batchnorm_dst_memory = user_dst_memory;
streams[0].addMemory(user_dst_memory);
if (mkldnn::memory::primitive_desc(batchnorm_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
batchnorm_dst_memory = mkldnn::memory(batchnorm_prim_desc.dst_primitive_desc());
streams[0].addMemory(batchnorm_dst_memory);
}
streams[0].addMemory(batchnorm_mean_memory);
streams[0].addMemory(batchnorm_variance_memory);
if (applyScale || applyOffset) {
auto batchnorm_weights_memory = mkldnn::memory(batchnorm_prim_desc.weights_primitive_desc(), weights.buffer());
streams[0].addMemory(batchnorm_weights_memory);
streams[0].addOperation(batch_normalization_forward(batchnorm_prim_desc, (mkldnn::primitive::at)batchnorm_src_memory,
(mkldnn::primitive::at)batchnorm_mean_memory, (mkldnn::primitive::at)batchnorm_variance_memory, (mkldnn::primitive::at)batchnorm_weights_memory, batchnorm_dst_memory));
} else {
streams[0].addOperation(batch_normalization_forward(batchnorm_prim_desc, (mkldnn::primitive::at)batchnorm_src_memory,
(mkldnn::primitive::at)batchnorm_mean_memory, (mkldnn::primitive::at)batchnorm_variance_memory, batchnorm_dst_memory));
}
if (mkldnn::memory::primitive_desc(batchnorm_prim_desc.dst_primitive_desc())
!= user_dst_memory.get_primitive_desc()) {
streams[0].addOperation(reorder(batchnorm_dst_memory, user_dst_memory));
}
}
if (applyScale || applyOffset) {
if (gamma != nullptr) {
weights({0, 1, 0, 0}).assign(gamma);
}
if (beta != nullptr) {
weights({1, 2, 0, 0}).assign(beta);
}
}
streams[0].submitAndWait();
return Status::OK();
}
#endif
nd4j_debug("MKL-DNN is not used for batchnorm_new!\n", 0);
// formula: output = gamma * ((input - mean) / sqrt(variance + epsilon)) + beta
helpers::batchnorm(input, mean, variance, gamma, beta, output, axes, epsilon);
return Status::OK();
}
DECLARE_TYPES(batchnorm_new) {
getOpDescriptor()->setAllowedInputTypes({ALL_FLOATS})->setSameMode(true);
}
DECLARE_SHAPE_FN(batchnorm_new) {
auto inShapeInfo = inputShape->at(0);
DataType outType = DataTypeUtils::pickFloatingType(ArrayOptions::dataType(inShapeInfo));
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) {
auto input = INPUT_VARIABLE(0);
auto mean = INPUT_VARIABLE(1);
auto variance = INPUT_VARIABLE(2);
NDArray *gamma = nullptr;
NDArray *beta = nullptr;
NDArray *dLdO = nullptr; // next epsilon
auto dLdI = OUTPUT_VARIABLE(0);
auto dLdM = OUTPUT_VARIABLE(1);
auto dLdV = OUTPUT_VARIABLE(2);
NDArray *dLdG = nullptr;
NDArray *dLdB = nullptr;
const bool applyScale = (bool)INT_ARG(0);
const bool applyOffset = (bool)INT_ARG(1);
// FIXME: double?
const double epsilon = T_ARG(0);
const int dLdONum = static_cast<int>(applyScale) + static_cast<int>(applyOffset);
if(applyScale) {
gamma = INPUT_VARIABLE(3);
dLdG = OUTPUT_VARIABLE(3);
}
if(applyOffset) {
beta = INPUT_VARIABLE(3 + static_cast<int>(applyScale));
dLdB = OUTPUT_VARIABLE(3 + static_cast<int>(applyScale));
}
dLdO = INPUT_VARIABLE(3 + dLdONum);
std::vector<const NDArray*> inArrs(block.width());
for(int i = 0; i < 4 + dLdONum; ++i)
inArrs[i] = INPUT_VARIABLE(i);
// check whether all input shapes are mutually broadcastable
Nd4jLong* outShapeInfo = nullptr;
const bool areShapesOk = ShapeUtils::evalCommonBroadcastShapeInfo(inArrs, outShapeInfo, block.getWorkspace());
REQUIRE_TRUE(areShapesOk, 0, "BATCHNORM_BP op: the shapes of input arrays are not mutually broadcastable !");
// ***** calculations ***** //
auto sigmaInv = (*variance + epsilon).transform(transform::RSqrt);
NDArray sigmaInvGamdLdO = -sigmaInv * *dLdO;
if(applyScale)
sigmaInvGamdLdO *= *gamma;
NDArray inputMinusMean;
if(!input->isSameShape(dLdO) && !mean->isSameShape(dLdO)) {
auto inputTiled = NDArray(dLdO, false, block.launchContext());
input->tile(inputTiled);
inputMinusMean = inputTiled - *mean;
}
else
inputMinusMean = *input - *mean;
// dLdI
if(!dLdI->isSameShape(dLdO))
dLdI->assign( (-sigmaInvGamdLdO).reduceAlongDims(reduce::Sum, ShapeUtils::evalBroadcastBackwardAxis(dLdI->getShapeInfo(), dLdO->getShapeInfo())) );
else
dLdI->assign(-sigmaInvGamdLdO);
// dLdM
if(!dLdM->isSameShape(dLdO))
dLdM->assign( sigmaInvGamdLdO.reduceAlongDims(reduce::Sum, ShapeUtils::evalBroadcastBackwardAxis(dLdM->getShapeInfo(), dLdO->getShapeInfo())) );
else
dLdM->assign(sigmaInvGamdLdO);
// dLdV
if(!dLdV->isSameShape(dLdO)) {
dLdV->assign( (sigmaInv * sigmaInv * sigmaInvGamdLdO * inputMinusMean * 0.5f).reduceAlongDims(reduce::Sum, ShapeUtils::evalBroadcastBackwardAxis(dLdV->getShapeInfo(), dLdO->getShapeInfo())) );
}
else
dLdV->assign(sigmaInv * sigmaInv * sigmaInvGamdLdO * inputMinusMean * 0.5f);
// dLdG
if(applyScale) {
if(!dLdG->isSameShape(dLdO))
dLdG->assign( (sigmaInv * inputMinusMean * *dLdO).reduceAlongDims(reduce::Sum, ShapeUtils::evalBroadcastBackwardAxis(dLdG->getShapeInfo(), dLdO->getShapeInfo())) );
else
dLdG->assign(sigmaInv * inputMinusMean * *dLdO);
}
// dLdB
if(applyOffset) {
if(!dLdB->isSameShape(dLdO))
dLdB->assign(dLdO->reduceAlongDims(reduce::Sum, ShapeUtils::evalBroadcastBackwardAxis(dLdB->getShapeInfo(), dLdO->getShapeInfo())) );
else
dLdB->assign(dLdO);
}
return Status::OK();
}
DECLARE_TYPES(batchnorm_bp) {
getOpDescriptor()
->setAllowedInputTypes(0, nd4j::DataType::ANY)
->setAllowedInputTypes(1, nd4j::DataType::ANY)
->setAllowedInputTypes(2, nd4j::DataType::ANY)
->setAllowedInputTypes(3, nd4j::DataType::ANY)
->setAllowedInputTypes(4, nd4j::DataType::ANY)
->setAllowedInputTypes(5, {ALL_FLOATS})
->setAllowedOutputTypes({ALL_FLOATS});
}
//////////////////////////////////////////////////////////////////////////
DECLARE_SHAPE_FN(batchnorm_bp) {
const bool applyScale = (bool)INT_ARG(0);
const bool applyOffset = (bool)INT_ARG(1);
const int dLdONum = static_cast<int>(applyScale) + static_cast<int>(applyOffset);
std::vector<const NDArray*> inArrs(block.width());
for(int i = 0; i < 4 + dLdONum; ++i)
inArrs[i] = INPUT_VARIABLE(i);
// check whether all input shapes are mutually broadcastable
Nd4jLong* outShapeInfo = nullptr;
const bool areShapesOk = ShapeUtils::evalCommonBroadcastShapeInfo(inArrs, outShapeInfo, block.getWorkspace());
REQUIRE_TRUE(areShapesOk, 0, "BATCHNORM_BP op: the shapes of input arrays are not mutually broadcastable !");
Nd4jLong* dLdIShapeInfo(nullptr), *dLdMShapeInfo(nullptr), *dLdVShapeInfo(nullptr), *dLdGShapeInfo(nullptr), *dLdBShapeInfo(nullptr);
COPY_SHAPE(inputShape->at(0), dLdIShapeInfo);
COPY_SHAPE(inputShape->at(1), dLdMShapeInfo);
COPY_SHAPE(inputShape->at(2), dLdVShapeInfo);
if(applyScale) {
COPY_SHAPE(inputShape->at(3), dLdGShapeInfo);
}
if(applyOffset){
COPY_SHAPE(inputShape->at(3 + static_cast<int>(applyScale)), dLdBShapeInfo);
}
if(!applyScale && !applyOffset)
return SHAPELIST(CONSTANT(dLdIShapeInfo), CONSTANT(dLdMShapeInfo), CONSTANT(dLdVShapeInfo));
if(applyScale && !applyOffset)
return SHAPELIST(CONSTANT(dLdIShapeInfo), CONSTANT(dLdMShapeInfo), CONSTANT(dLdVShapeInfo), CONSTANT(dLdGShapeInfo));
if(!applyScale && applyOffset)
return SHAPELIST(CONSTANT(dLdIShapeInfo), CONSTANT(dLdMShapeInfo), CONSTANT(dLdVShapeInfo), CONSTANT(dLdBShapeInfo));
return SHAPELIST(CONSTANT(dLdIShapeInfo), CONSTANT(dLdMShapeInfo), CONSTANT(dLdVShapeInfo), CONSTANT(dLdGShapeInfo), CONSTANT(dLdBShapeInfo));
}
// //////////////////////////////////////////////////////////////////////////
// CONFIGURABLE_OP_IMPL(batchnorm_bp, 5, 1, true, 0, 1) {
// NDArray<T>* input = INPUT_VARIABLE(0);
// NDArray<T>* epsilon = INPUT_VARIABLE(1);
// NDArray<T>* gamma = INPUT_VARIABLE(2);
// NDArray<T>* dGlobalMeanView = INPUT_VARIABLE(3);
// NDArray<T>* dGlobalVarView = INPUT_VARIABLE(4);
// NDArray<T>* outEpsilon = this->getZ(block);
// std::vector<int> argI = *(block.getIArguments());
// const int bS = epsilon->sizeAt(0);
// bool isLockGammaBeta = (bool)argI[0];
// const int* epsilonShape = epsilon->getShapeInfo() + 1;
// const T eps = (T)1e-5;
// int rank = epsilon->rankOf();
// std::initializer_list<int> dimensions;
// int effectiveBatchSize;
// if (rank == 2) {
// dimensions = {0};
// effectiveBatchSize = bS;
// }
// else if (rank == 4) {
// dimensions = {0, 2, 3};
// effectiveBatchSize = input->sizeAt(0)*input->sizeAt(2)*input->sizeAt(3);
// }
// else
// throw "Graph operation batchnorm_bp: the epsilon rank must be equal to 2 or 4 !";
// NDArray<T> *mean(nullptr), *var(nullptr), *dBeta(nullptr), *dGamma(nullptr), *dLdVar(nullptr), *dxmu1(nullptr), *dxmu2(nullptr);
// mean = input->template reduceAlongDimension<simdOps::Mean<T>>(dimensions);
// var = input->template varianceAlongDimension<simdOps::SummaryStatsVariance<T>>(false, dimensions);
// var->template applyScalar<simdOps::Add<T>>(eps, nullptr);
// auto std = new NDArray<T>(var->getShapeInfo(), block.getWorkspace());
// var->template applyTransform<simdOps::Sqrt<T>>(std, nullptr);
// auto xMu = new NDArray<T>(input->getShapeInfo(), block.getWorkspace());
// auto xHat = new NDArray<T>(input->getShapeInfo(), block.getWorkspace());
// auto temp1 = new NDArray<T>(epsilon->getShapeInfo(), block.getWorkspace());
// auto temp2 = new NDArray<T>(std->getShapeInfo(), block.getWorkspace());
// auto dGammaView = new NDArray<T>('c', {1, epsilonShape[1]}, block.getWorkspace());
// auto dBetaView = new NDArray<T>('c', {1, epsilonShape[1]}, block.getWorkspace());
// auto dxhat = new NDArray<T>(epsilon->getShapeInfo(), block.getWorkspace());
// if (rank == 2) {
// input->subRowVector(mean, xMu);
// xMu->divRowVector(std, xHat);
// }
// else {
// input->template applyBroadcast<simdOps::Subtract<T>>({1}, mean, xMu, nullptr);
// xMu->template applyBroadcast<simdOps::Divide<T>>({1}, std, xHat, nullptr);
// }
// dBeta = epsilon->sum(dimensions); // dL/dBeta = sum_examples dL/dOut
// epsilon->template applyPairwiseTransform<simdOps::Multiply<T>>(xHat, temp1, nullptr); //dL/dGamma = sum_examples dL/dOut .* xHat
// dGamma = temp1->sum(dimensions); //dL/dGamma = sum_examples dL/dOut .* xHat
// if (isLockGammaBeta)
// epsilon->template applyPairwiseTransform<simdOps::Multiply<T>>(gamma, dxhat, nullptr);
// else {// Standard case
// if(rank == 2)
// epsilon->mulRowVector(gamma, dxhat); //dL/dxHat = dL/dOut . gamma Shape: [minibatchSize, nOut]
// else
// epsilon->template applyBroadcast<simdOps::Multiply<T>>({1}, gamma, dxhat, nullptr);
// }
// // dLdVar - dL/dVariance, shape: [1, miniBatch]
// dxhat->template applyPairwiseTransform<simdOps::Multiply<T>>(xMu, temp1, nullptr);
// dLdVar = temp1->sum(dimensions);
// dLdVar->template applyScalar<simdOps::Multiply<T>>((T)-0.5, nullptr);
// T powParams[] = {(T)(-3.)};
// std->template applyTransform<simdOps::Pow<T>>(temp2, powParams);
// dLdVar->template applyPairwiseTransform<simdOps::Multiply<T>>(temp2, nullptr);
// //dL/dmu
// dxmu1 = dxhat->sum(dimensions);
// dxmu1->template applyPairwiseTransform<simdOps::Divide<T>>(std, nullptr);
// dxmu1->template applyTransform<simdOps::Neg<T>>();
// dxmu2 = xMu->sum(dimensions);
// dxmu2->template applyScalar<simdOps::Multiply<T>>((T)(-2.)/effectiveBatchSize);
// dxmu2->template applyPairwiseTransform<simdOps::Multiply<T>>(dLdVar, nullptr);
// dxmu1->template applyPairwiseTransform<simdOps::Add<T>>(dxmu2, nullptr);
// NDArray<T>* dLdmu = dxmu1; // = dL/dmu Shape: [1, nOut]
// //Note the array reuse here: dxhat, xMu, dLdVar, dLdmu - all are invalid after this line (but aren't used later anyway)
// NDArray<T>* dLdx = dxhat;
// dLdVar->template applyScalar<simdOps::Multiply<T>>((T)(2.)/effectiveBatchSize);
// dLdmu->template applyScalar<simdOps::Multiply<T>>((T)(1.)/effectiveBatchSize);
// if(rank == 2) {
// dLdx->divRowVector(std, dLdx);
// xMu->mulRowVector(dLdVar, xMu);
// }
// else {
// dLdx->template applyBroadcast<simdOps::Divide<T>>({1}, std, dLdx, nullptr);
// xMu->template applyBroadcast<simdOps::Multiply<T>>({1}, dLdVar, xMu, nullptr);
// }
// dLdx->template applyPairwiseTransform<simdOps::Add<T>>(xMu, nullptr);
// if(rank == 2)
// dLdx->addRowVector(dLdmu, dLdx);
// else
// dLdx->template applyBroadcast<simdOps::Add<T>>({1}, dLdmu, dLdx, nullptr);
// *outEpsilon = *dLdx;
// //TODO rework this to avoid the assign here
// // dGammaView->assign(dGamma);
// // dBetaView->assign(dBeta);
// // dGlobalMeanView->assign((T)0.);
// // dGlobalVarView->assign((T)0.);
// // retGradient.setGradientFor(BatchNormalizationParamInitializer.GAMMA, dGammaView);
// // retGradient.setGradientFor(BatchNormalizationParamInitializer.BETA, dBetaView);
// // retGradient.setGradientFor(BatchNormalizationParamInitializer.GLOBAL_MEAN, dGlobalMeanView);
// // retGradient.setGradientFor(BatchNormalizationParamInitializer.GLOBAL_VAR, dGlobalVarView);
// delete std;
// delete xMu;
// delete xHat;
// delete mean;
// delete var;
// delete dBeta;
// delete dGamma;
// delete dLdVar;
// delete dxmu1;
// delete dxmu2;
// delete temp1;
// delete temp2;
// delete dxhat;
// delete dGammaView;
// delete dBetaView;
// return ND4J_STATUS_OK;
// }
}
}
#endif
Reference in New Issue View Git Blame Copy Permalink