提交 7b8229d6 编写于 作者: M mindspore-ci-bot 提交者: Gitee

!4913 transform bn to scale

Merge pull request !4913 from zhengjun10/master
......@@ -16,7 +16,7 @@ tracking
mtk_isface
mtk_landmark
mtk_pose_tuku
mtk_face_recognition_v1
# mtk_face_recognition_v1
mtk_2012_ATLANTA_10class_20190614_v41
mtk_detect-deeper-halfdeeper-mbv1-lastearlySSD-shortcut-400-400_nopostprocess_simplified
detect-deeper-halfdeeper-mbv1-shortcut-400-400_nopostprocess_simplified
......@@ -28,16 +28,16 @@ ml_hand_detection
ml_ocr_cn
ml_ocr_sfz_detect_0325
ml_hardware_liveness
ml_liveness_detect_landmark
# ml_liveness_detect_landmark
ml_face_contour
2012_ATLANTA_1class_20190621_v4.x_nomean
ml_handpose
ml_ocr_sfz_add_final_0325
ml_hardware_pose
# ml_hardware_pose
ml_bank_recog
2012_ATLANTA_10class_20190131_v4.0
mnet
recognition
# recognition
ml_face_landmark
model_hebing_3branch
hiai_cv_focusShootOCRModel_07
......@@ -50,7 +50,7 @@ hiai_cpu_face_hat
hiai_video_seg
hiai_semantic_seg
hiai_human_seg
hiai_face_recognition_1
# hiai_face_recognition_1
hiai_cpu_face_detect
hiai_cpu_face_attr
hiai_face_attr1
......
......@@ -27,6 +27,7 @@
#include "tools/converter/legacy_optimizer/fusion/format_trans_fusion_pass.h"
#include "tools/converter/legacy_optimizer/fusion/format_trans_transpose_fusion_pass.h"
#include "tools/converter/legacy_optimizer/fusion/quant_cast_fusion_pass.h"
#include "tools/converter/legacy_optimizer/fusion/batchnorm_convert_scale_pass.h"
#include "tools/converter/legacy_optimizer/graph/weight_format_hardcode_pass.h"
#include "tools/converter/legacy_optimizer/graph/weight_format_transform_pass.h"
#include "tools/converter/legacy_optimizer/graph/format_trans_pass.h"
......@@ -126,6 +127,17 @@ int GraphDefTransform::Transform(const converter::Flags &ctx) {
}
}
// postconvert pass
{
Optimizer fusionOptimizer;
fusionOptimizer.AddPass(new (std::nothrow) BatchNormConvertScalePass());
fusionOptimizer.AddPass(new (std::nothrow) IsolatedNodeRemovePass());
status = fusionOptimizer.Run(graphDefT);
if (status != RET_OK && status != RET_NO_CHANGE) {
MS_LOG(ERROR) << "Run fusionOptimizer BatchNormConvertScalePass Failed";
return status;
}
}
// format transform
if (ctx.formatTrans) {
Optimizer formatTransOptimizer;
......@@ -187,6 +199,7 @@ int GraphDefTransform::Transform(const converter::Flags &ctx) {
return status;
}
}
// topological sorting
{
Optimizer topologicalOptimizer;
......
......@@ -6,6 +6,7 @@ add_library(fusion_mid OBJECT
${CMAKE_CURRENT_SOURCE_DIR}/batchnorm_fold_fusion_pass.cc
${CMAKE_CURRENT_SOURCE_DIR}/format_trans_fusion_pass.cc
${CMAKE_CURRENT_SOURCE_DIR}/format_trans_transpose_fusion_pass.cc
${CMAKE_CURRENT_SOURCE_DIR}/batchnorm_convert_scale_pass.cc
)
target_link_libraries(fusion_mid securec)
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://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.
*/
#include "tools/converter/legacy_optimizer/fusion/batchnorm_convert_scale_pass.h"
#include <cfloat>
#include <memory>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include "utils/log_adapter.h"
#include "tools/common/graph_util.h"
#include "tools/common/tensor_util.h"
#include "include/errorcode.h"
#include "schema/inner/model_generated.h"
#include "src/common/op_utils.h"
namespace mindspore {
namespace lite {
#define CAFFE_BATCHNORM_OP_WEIGHT_NUM 2
#define TF_BATCHNORM_OP_WEIGHT_NUM 4
#define CAFFE_BATCHNORM_MEAN_INDEX 0
#define CAFFE_BATCHNORM_VARIANCE_INDEX 1
#define TF_BATCHNORM_SCALE_INDEX 0
#define TF_BATCHNORM_BIAS_INDEX 1
#define TF_BATCHNORM_MEAN_INDEX 2
#define TF_BATCHNORM_VARIANCE_INDEX 3
namespace {
constexpr const float EPS = 1e-8;
constexpr const float EPS_DEFAULT_FLOAT = 1e-5;
constexpr const float POW_NUM = 0.5;
constexpr const int32_t NCHW_DIM_C = 1;
}
STATUS BatchNormConvertScalePass::Run(MetaGraphT *graph) { return FusionPass::Run(graph); }
STATUS BatchNormConvertScalePass::DefinePattern() {
// with preNode
{
auto inputOp = std::make_shared<PatternOp>();
inputOp->id = inputOpName;
inputOp->types = {schema::PrimitiveType_NONE};
inputOp->isPlaceHold = true;
auto bnOp = std::make_shared<PatternOp>();
bnOp->id = bnOpName;
bnOp->types = {schema::PrimitiveType_FusedBatchNorm, schema::PrimitiveType_BatchNorm};
bnOp->left = inputOp;
std::unique_ptr<FusionPattern> fusionPattern(new(std::nothrow) FusionPattern(bnPatternName));
if (fusionPattern == nullptr) {
MS_LOG(ERROR) << "new fusionPattern failed";
return RET_ERROR;
}
fusionPattern->AddPatternOp(inputOp);
fusionPattern->AddPatternOp(bnOp);
fusionPattern->Finish();
this->patterns.emplace_back(fusionPattern.release());
}
return RET_OK;
}
STATUS BatchNormConvertScalePass::DoFusion(MetaGraphT *graph, const std::string &patternName,
std::unordered_map<std::string, std::shared_ptr<Path>> &matchedPath) {
MS_ASSERT(graph != nullptr);
if (patternName != bnPatternName) {
MS_LOG(ERROR) << "BatchNormConvertScale-Fusion match failed";
return RET_PARAM_INVALID;
}
auto status = FindNodes(graph, matchedPath);
if (status != RET_OK) {
MS_LOG(ERROR) << "FindNodes failed: " << status;
return status;
}
auto type = bnNode->primitive->value.type;
if (type != schema::PrimitiveType_FusedBatchNorm && type != schema::PrimitiveType_BatchNorm) {
return RET_OK;
}
auto bnPath = matchedPath.at(bnOpName);
status = GetTransParam(graph, bnPath);
if (status != RET_OK) {
MS_LOG(ERROR) << "GetTransParam failed: " << status;
return status;
}
status = GenNewScaleTensor(graph, bnPath);
if (status != RET_OK) {
MS_LOG(ERROR) << "GenNewScaleTensor failed: " << status;
return status;
}
status = ConvertBNToScale(graph, bnPath);
if (status != RET_OK) {
MS_LOG(ERROR) << "GenNewScaleTensor failed: " << status;
return status;
}
return RET_OK;
}
STATUS BatchNormConvertScalePass::ConvertBNToScale(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath) {
auto scaleNode = std::unique_ptr<CNodeT>(new(std::nothrow) CNodeT);
if (scaleNode == nullptr) {
MS_LOG(ERROR) << "new TransNode failed";
return RET_ERROR;
}
scaleNode->name = bnNode->name;
scaleNode->primitive = std::make_unique<schema::PrimitiveT>();
if (scaleNode->primitive == nullptr) {
MS_LOG(ERROR) << "op->primitive is null";
return RET_NULL_PTR;
}
scaleNode->primitive->value.type = schema::PrimitiveType_Scale;
std::unique_ptr<ScaleT> scaleParam(new ScaleT());
if (scaleParam == nullptr) {
MS_LOG(ERROR) << "new transposeParam failed";
return RET_ERROR;
}
scaleParam->axis = NCHW_DIM_C;
scaleNode->primitive->value.value = scaleParam.release();
auto scaleIter = graph->nodes.begin() + bnPath->nodeIdx;
STATUS errorCode = RET_OK;
scaleIter =
InsertNode(graph, scaleIter, kBefore, 0, std::move(scaleNode), &errorCode, ScaleOpCopyer);
if (errorCode != RET_OK) {
MS_LOG(ERROR) << "InsertNode failed: %d"; // errorCode);
return errorCode;
}
auto &newScaleNode = *(scaleIter - 1);
graph->allTensors.emplace_back(std::move(newScaleWeightTensor));
auto weightTensorIdx = graph->allTensors.size() - 1;
graph->allTensors.emplace_back(std::move(newScaleBiasTensor));
auto biasTensorIdx = graph->allTensors.size() - 1;
newScaleNode->inputIndex.push_back(weightTensorIdx);
newScaleNode->inputIndex.push_back(biasTensorIdx);
// delete bn node
auto status = IsolateOneWayNode(graph, bnPath->nodeIdx + 1, true);
if (status != RET_OK) {
MS_LOG(ERROR) << "IsolateOneWayNode " << bnNode->name.c_str() << " failed, error: " << status;
return status;
}
return RET_OK;
}
STATUS BatchNormConvertScalePass::GenNewScaleTensor(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath) {
MS_ASSERT(graph != nullptr);
GetTransParam(graph, bnPath);
newScaleWeightTensor = std::unique_ptr<TensorT>(new(std::nothrow) TensorT);
if (newScaleWeightTensor == nullptr) {
MS_LOG(ERROR) << "new weightTensor failed";
return RET_ERROR;
}
newScaleWeightTensor->dataType = bnMeanTensor->dataType;
newScaleWeightTensor->format = bnMeanTensor->format;
newScaleWeightTensor->refCount = schema::NodeType_ValueNode;
newScaleWeightTensor->dims = bnMeanTensor->dims;
auto weightShapeSize = GetShapeSize(*bnMeanTensor);
newScaleWeightTensor->data.resize(weightShapeSize * sizeof(float));
auto ret = memcpy_s(newScaleWeightTensor->data.data(), weightShapeSize * sizeof(float), transScale,
weightShapeSize * sizeof(float));
if (ret != RET_OK) {
delete transScale;
MS_LOG(ERROR) << "memcpy error: " << ret;
return RET_ERROR;
}
newScaleBiasTensor = std::unique_ptr<TensorT>(new(std::nothrow) TensorT);
if (newScaleBiasTensor == nullptr) {
MS_LOG(ERROR) << "new weightTensor failed";
return RET_ERROR;
}
newScaleBiasTensor->dataType = bnMeanTensor->dataType;
newScaleBiasTensor->format = bnMeanTensor->format;
newScaleBiasTensor->refCount = schema::NodeType_ValueNode;
newScaleBiasTensor->dims = bnMeanTensor->dims;
weightShapeSize = GetShapeSize(*bnMeanTensor);
newScaleBiasTensor->data.resize(weightShapeSize * sizeof(float));
ret = memcpy_s(newScaleBiasTensor->data.data(), weightShapeSize * sizeof(float), transBias,
weightShapeSize * sizeof(float));
if (ret != RET_OK) {
delete transBias;
MS_LOG(ERROR) << "memcpy error: " << ret;
return RET_ERROR;
}
return RET_OK;
}
STATUS BatchNormConvertScalePass::FindNodes(MetaGraphT *graph,
const std::unordered_map<std::string, std::shared_ptr<Path>> &matchedPath) {
MS_ASSERT(graph != nullptr);
auto inputPath = matchedPath.at(inputOpName);
auto bnPath = matchedPath.at(bnOpName);
MS_ASSERT(inputPath != nullptr);
MS_ASSERT(bnPath != nullptr);
if (inputPath->subGraphIdx != bnPath->subGraphIdx) {
MS_LOG(ERROR) << "matched nodes should from same subGraph";
return RET_ERROR;
}
MS_ASSERT(graph->nodes.size() > inputPath->nodeIdx);
MS_ASSERT(graph->nodes.size() > bnPath->nodeIdx);
inputNode = graph->nodes.at(inputPath->nodeIdx).get();
bnNode = graph->nodes.at(bnPath->nodeIdx).get();
MS_ASSERT(inputNode != nullptr);
MS_ASSERT(bnNode != nullptr);
return RET_OK;
}
STATUS BatchNormConvertScalePass::GetTransParam(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath) {
MS_ASSERT(graph != nullptr);
MS_ASSERT(bnPath != nullptr);
BNWeightTensors bnWeightTensors;
auto status = GetBnWeightTensors(graph, bnPath, &bnWeightTensors);
if (status != RET_OK) {
MS_LOG(ERROR) << "GetBnWeightTensors error";
return status;
}
auto *meanTensor = bnWeightTensors.meanTensor;
auto *varianceTensor = bnWeightTensors.varianceTensor;
auto *scaleTensor = bnWeightTensors.scaleTensor;
auto *biasTensor = bnWeightTensors.biasTensor;
auto *meanData = reinterpret_cast<float *>(meanTensor->data.data());
auto *varianceData = reinterpret_cast<float *>(varianceTensor->data.data());
eps = EPS_DEFAULT_FLOAT;
status = GetBnEpsilon(graph);
if (status != RET_OK) {
MS_LOG(ERROR) << "GetBnEpsilon failed";
return status;
}
this->transScale = new(std::nothrow) float[bnChannel];
this->transBias = new(std::nothrow) float[bnChannel];
// cal transScale, tf : scale/sqrt(variance + eps); caffe : 1/sqrt(variance + eps)
if (memcpy_s(transScale, bnChannel * sizeof(float), varianceData, bnChannel * sizeof(float)) != 0) {
MS_LOG(ERROR) << "memcpy_s transScale error";
return RET_ERROR;
}
// 1/sqrt(variance + eps)
for (uint32_t i = 0; i < bnChannel; i++) {
float tmp = transScale[i] + eps;
tmp = pow(tmp, POW_NUM);
transScale[i] = 1 / tmp;
}
if (scaleTensor != nullptr) {
auto *scaleData = reinterpret_cast<float *>(scaleTensor->data.data());
// scale/sqrt(variance + eps)
for (uint32_t i = 0; i < bnChannel; i++) {
transScale[i] *= scaleData[i];
}
}
// cal transBias, tf : -scale*mean/sqrt(variance + eps) + bias; caffe : -mean/sqrt(variance + eps)
// -mean/sqrt(variance + eps)
for (uint32_t i = 0; i < bnChannel; i++) {
transBias[i] = -meanData[i] * transScale[i];
}
if (biasTensor != nullptr) {
auto *biasData = reinterpret_cast<float *>(biasTensor->data.data());
// -scale*mean/sqrt(variance + eps) + bias
for (uint32_t i = 0; i < bnChannel; i++) {
transBias[i] += biasData[i];
}
}
return RET_OK;
}
// BatchNorm weight Tensor definition:
// caffe
// estimated_mean --0
// estimated_variance --1
// tensorflow
// scale -- 0
// bias --1
// estimated_mean --2
// estimated_variance --3
STATUS BatchNormConvertScalePass::GetBnWeightTensors(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath,
BNWeightTensors* bnWeightTensors) {
if (graph == nullptr || bnPath == nullptr) {
MS_LOG(ERROR) << "null pointer dereferencing.";
return RET_NULL_PTR;
}
MS_ASSERT(graph->allTensors.size() > bnNode->inputIndex.at(1));
auto bnWeightTensorIdxes = bnNode->inputIndex;
bnWeightTensorIdxes.erase(bnWeightTensorIdxes.begin());
if (bnWeightTensorIdxes.size() == CAFFE_BATCHNORM_OP_WEIGHT_NUM) {
bnWeightTensors->meanTensor = graph->allTensors.at(bnWeightTensorIdxes[CAFFE_BATCHNORM_MEAN_INDEX]).get();
bnWeightTensors->varianceTensor = graph->allTensors.at(bnWeightTensorIdxes[CAFFE_BATCHNORM_VARIANCE_INDEX]).get();
} else if (bnWeightTensorIdxes.size() == TF_BATCHNORM_OP_WEIGHT_NUM) {
bnWeightTensors->scaleTensor = graph->allTensors.at(bnWeightTensorIdxes[TF_BATCHNORM_SCALE_INDEX]).get();
bnWeightTensors->biasTensor = graph->allTensors.at(bnWeightTensorIdxes[TF_BATCHNORM_BIAS_INDEX]).get();
bnWeightTensors->meanTensor = graph->allTensors.at(bnWeightTensorIdxes[TF_BATCHNORM_MEAN_INDEX]).get();
bnWeightTensors->varianceTensor = graph->allTensors.at(bnWeightTensorIdxes[TF_BATCHNORM_VARIANCE_INDEX]).get();
} else {
MS_LOG(ERROR) << "BatchNorm should has 2 or 4 weight tensors, current number of weight tensors: "
<< bnWeightTensorIdxes.size();
return RET_ERROR;
}
if (bnWeightTensors->meanTensor == nullptr) {
MS_LOG(ERROR) << "BatchNorm's mean tensor is nullptr";
return RET_ERROR;
}
if (bnWeightTensors->varianceTensor == nullptr) {
MS_LOG(ERROR) << "BatchNorm's variance tensor is nullptr";
return RET_ERROR;
}
bnChannel = bnWeightTensors->meanTensor->data.size() * sizeof(uint8_t) / sizeof(float);
if (bnChannel <= 0) {
MS_LOG(ERROR) << "BatchNorm's channel less or equal 0";
return RET_ERROR;
}
bnMeanTensor = bnWeightTensors->meanTensor;
if (bnChannel != bnWeightTensors->varianceTensor->data.size() * sizeof(uint8_t) / sizeof(float)) {
MS_LOG(ERROR) << "conv kernel num expected to be equal to variance size";
return RET_ERROR;
}
if (bnWeightTensors->scaleTensor != nullptr) {
if (bnChannel != bnWeightTensors->scaleTensor->data.size() * sizeof(uint8_t) / sizeof(float)) {
MS_LOG(ERROR) << "conv kernel num expected to be equal to scale size";
return RET_ERROR;
}
}
if (bnWeightTensors->biasTensor != nullptr) {
if (bnChannel != bnWeightTensors->biasTensor->data.size() * sizeof(uint8_t) / sizeof(float)) {
MS_LOG(ERROR) << "conv kernel num expected to be equal to bias size";
return RET_ERROR;
}
}
return RET_OK;
}
STATUS BatchNormConvertScalePass::GetBnEpsilon(MetaGraphT *graph) {
if (graph == nullptr) {
MS_LOG(ERROR) << "null pointer dereferencing.";
return RET_NULL_PTR;
}
if (bnNode == nullptr) {
MS_LOG(ERROR) << "null pointer dereferencing.";
return RET_NULL_PTR;
}
if (bnNode->primitive->value.type == schema::PrimitiveType_FusedBatchNorm) {
eps = bnNode->primitive->value.AsFusedBatchNorm()->epsilon;
} else if (bnNode->primitive->value.type == schema::PrimitiveType_BatchNorm) {
eps = bnNode->primitive->value.AsBatchNorm()->epsilon;
} else {
MS_LOG(ERROR) << "match pattern has error, not BatchNorm node";
return RET_ERROR;
}
if (eps < EPS) {
eps = EPS_DEFAULT_FLOAT;
}
return RET_OK;
}
BatchNormConvertScalePass::~BatchNormConvertScalePass() {
if (this->transScale != nullptr) {
delete (this->transScale);
}
if (this->transBias != nullptr) {
delete (this->transBias);
}
}
} // namespace lite
} // namespace mindspore
/**
* Copyright 2020 Huawei Technologies Co., Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://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.
*/
#ifndef MINDSPORE_PREDICT_BATCHNORM_CONVERT_SCALE_PASS_H
#define MINDSPORE_PREDICT_BATCHNORM_CONVERT_SCALE_PASS_H
#include <unordered_map>
#include <memory>
#include <string>
#include <utility>
#include "tools/converter/legacy_optimizer/fusion/fusion_pass.h"
#include "tools/common/graph_util.h"
namespace mindspore {
namespace lite {
struct BNWeightTensors {
TensorT *meanTensor = nullptr;
TensorT *varianceTensor = nullptr;
TensorT *scaleTensor = nullptr;
TensorT *biasTensor = nullptr;
};
class BatchNormConvertScalePass : public FusionPass {
public:
BatchNormConvertScalePass() = default;
~BatchNormConvertScalePass() override;
STATUS DefinePattern() override;
STATUS DoFusion(MetaGraphT *graph, const std::string &patternName,
std::unordered_map<std::string, std::shared_ptr<Path>> &matchedPath) override;
STATUS Run(MetaGraphT *graph) override;
protected:
STATUS GetTransParam(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath);
// Get and check BNNode weight tensor
STATUS GetBnWeightTensors(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath, BNWeightTensors* bnWeightTensors);
STATUS GetBnEpsilon(MetaGraphT *graph);
STATUS FindNodes(MetaGraphT *graph, const std::unordered_map<std::string, std::shared_ptr<Path>> &matchedPath);
STATUS GenNewScaleTensor(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath);
STATUS ConvertBNToScale(MetaGraphT *graph, const std::shared_ptr<Path> &bnPath);
CNodeT *inputNode = nullptr;
CNodeT *bnNode = nullptr;
std::string inputOpName = "Input";
std::string bnOpName = "BatchNorm";
std::string bnPatternName = "BnToScaleFusion";
uint32_t bnChannel = 0;
float eps = 0;
TensorT *bnMeanTensor = nullptr;
float *transScale = nullptr;
float *transBias = nullptr;
std::unique_ptr<TensorT> newScaleWeightTensor = nullptr;
std::unique_ptr<TensorT> newScaleBiasTensor = nullptr;
OpDefCopyer ScaleOpCopyer = [](CNodeT *inOpDef) -> std::unique_ptr<CNodeT> {
std::unique_ptr<CNodeT> newOpDef(new(std::nothrow) CNodeT);
if (newOpDef == nullptr) {
MS_LOG(ERROR) << "new OpDefT failed";
return nullptr;
}
newOpDef->name = inOpDef->name;
newOpDef->quantType = inOpDef->quantType;
newOpDef->primitive = std::make_unique<schema::PrimitiveT>();
newOpDef->primitive->value.type = schema::PrimitiveType_Scale;
auto scaleParam = new(std::nothrow) ScaleT;
if (scaleParam == nullptr) {
MS_LOG(ERROR) << "new scaleParam failed";
return nullptr;
}
auto inParam = inOpDef->primitive->value.AsScale();
MS_ASSERT(inParam != nullptr);
scaleParam->axis = inParam->axis;
newOpDef->primitive->value.value = scaleParam;
return std::move(newOpDef);
};
};
} // namespace lite
} // namespace mindspore
#endif // MINDSPORE_PREDICT_BATCHNORM_CONVERT_SCALE_PASS_H
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册