From 5b5346ecc74fdd804ccb26c9de35bdd8d9a9a187 Mon Sep 17 00:00:00 2001 From: yangyaming Date: Tue, 13 Jun 2017 14:18:19 +0800 Subject: [PATCH] Follow comments. --- paddle/gserver/layers/DetectionUtil.cpp | 83 ++++++++++++++----------- paddle/gserver/layers/DetectionUtil.h | 28 +++++---- 2 files changed, 61 insertions(+), 50 deletions(-) diff --git a/paddle/gserver/layers/DetectionUtil.cpp b/paddle/gserver/layers/DetectionUtil.cpp index f301d2d33..e9295edb6 100644 --- a/paddle/gserver/layers/DetectionUtil.cpp +++ b/paddle/gserver/layers/DetectionUtil.cpp @@ -16,29 +16,33 @@ limitations under the License. */ namespace paddle { -size_t appendWithPermute(const MatrixPtr inMatrix, +size_t appendWithPermute(const Matrix& inMatrix, size_t height, size_t width, size_t outTotalSize, size_t outOffset, size_t batchSize, - MatrixPtr outMatrix, + Matrix& outMatrix, PermMode permMode, bool useGpu) { - if (permMode == NCHWTONHWC) { - size_t inElementCnt = inMatrix->getElementCnt(); + if (permMode == kNCHWToNHWC) { + size_t inElementCnt = inMatrix.getElementCnt(); size_t channels = inElementCnt / (height * width * batchSize); size_t imgSize = height * width; for (size_t i = 0; i < batchSize; ++i) { size_t offset = i * (outTotalSize / batchSize) + outOffset; - const MatrixPtr inTmp = - Matrix::create(inMatrix->getData() + i * channels * imgSize, - channels, + const MatrixPtr inTmp = Matrix::create( + const_cast(inMatrix.getData()) + i * channels * imgSize, + channels, + imgSize, + false, + useGpu); + MatrixPtr outTmp = + Matrix::create(const_cast(outMatrix.getData()) + offset, imgSize, + channels, false, useGpu); - MatrixPtr outTmp = Matrix::create( - outMatrix->getData() + offset, imgSize, channels, false, useGpu); inTmp->transpose(outTmp, false); } return channels * imgSize; @@ -47,29 +51,33 @@ size_t appendWithPermute(const MatrixPtr inMatrix, } } -size_t decomposeWithPermute(const MatrixPtr inMatrix, +size_t decomposeWithPermute(const Matrix& inMatrix, size_t height, size_t width, size_t inTotalSize, size_t inOffset, size_t batchSize, - MatrixPtr outMatrix, + Matrix& outMatrix, PermMode permMode, bool useGpu) { - if (permMode == NHWCTONCHW) { - size_t outElementCnt = outMatrix->getElementCnt(); + if (permMode == kNHWCToNCHW) { + size_t outElementCnt = outMatrix.getElementCnt(); size_t channels = outElementCnt / (height * width * batchSize); size_t imgSize = height * width; for (size_t i = 0; i < batchSize; ++i) { size_t offset = i * (inTotalSize / batchSize) + inOffset; - const MatrixPtr inTmp = Matrix::create( - inMatrix->getData() + offset, imgSize, channels, false, useGpu); - MatrixPtr outTmp = - Matrix::create(outMatrix->getData() + i * channels * imgSize, - channels, + const MatrixPtr inTmp = + Matrix::create(const_cast(inMatrix.getData()) + offset, imgSize, + channels, false, useGpu); + MatrixPtr outTmp = Matrix::create( + const_cast(outMatrix.getData()) + i * channels * imgSize, + channels, + imgSize, + false, + useGpu); inTmp->transpose(outTmp, false); } return channels * imgSize; @@ -99,9 +107,10 @@ real jaccardOverlap(const NormalizedBBox& bbox1, const NormalizedBBox& bbox2) { } } -vector encodeBBoxWithVar(const NormalizedBBox& priorBBox, - const vector priorBBoxVar, - const NormalizedBBox& gtBBox) { +void encodeBBoxWithVar(const NormalizedBBox& priorBBox, + const vector priorBBoxVar, + const NormalizedBBox& gtBBox, + vector& outVec) { real priorBBoxWidth = priorBBox.getWidth(); real priorBBoxHeight = priorBBox.getHeight(); real priorBBoxCenterX = priorBBox.getCenterX(); @@ -112,17 +121,15 @@ vector encodeBBoxWithVar(const NormalizedBBox& priorBBox, real gtBBoxCenterX = gtBBox.getCenterX(); real gtBBoxCenterY = gtBBox.getCenterY(); - vector offsetParam; - offsetParam.push_back((gtBBoxCenterX - priorBBoxCenterX) / priorBBoxWidth / - priorBBoxVar[0]); - offsetParam.push_back((gtBBoxCenterY - priorBBoxCenterY) / priorBBoxHeight / - priorBBoxVar[1]); - offsetParam.push_back(std::log(std::fabs(gtBBoxWidth / priorBBoxWidth)) / - priorBBoxVar[2]); - offsetParam.push_back(std::log(std::fabs(gtBBoxHeight / priorBBoxHeight)) / - priorBBoxVar[3]); - - return offsetParam; + outVec.clear(); + outVec.push_back((gtBBoxCenterX - priorBBoxCenterX) / priorBBoxWidth / + priorBBoxVar[0]); + outVec.push_back((gtBBoxCenterY - priorBBoxCenterY) / priorBBoxHeight / + priorBBoxVar[1]); + outVec.push_back(std::log(std::fabs(gtBBoxWidth / priorBBoxWidth)) / + priorBBoxVar[2]); + outVec.push_back(std::log(std::fabs(gtBBoxHeight / priorBBoxHeight)) / + priorBBoxVar[3]); } NormalizedBBox decodeBBoxWithVar(const NormalizedBBox& priorBBox, @@ -318,9 +325,9 @@ void matchBBox(const vector& priorBBoxes, } pair generateMatchIndices( - const MatrixPtr priorValue, + const Matrix& priorValue, const size_t numPriorBBoxes, - const MatrixPtr gtValue, + const Matrix& gtValue, const int* gtStartPosPtr, const size_t seqNum, const vector>& maxConfScore, @@ -331,7 +338,7 @@ pair generateMatchIndices( vector>* matchIndicesVecPtr, vector>* negIndicesVecPtr) { vector priorBBoxes; // share same prior bboxes - getBBoxFromPriorData(priorValue->getData(), numPriorBBoxes, priorBBoxes); + getBBoxFromPriorData(priorValue.getData(), numPriorBBoxes, priorBBoxes); size_t totalPos = 0; size_t totalNeg = 0; for (size_t n = 0; n < batchSize; ++n) { @@ -349,7 +356,7 @@ pair generateMatchIndices( } vector gtBBoxes; getBBoxFromLabelData( - gtValue->getData() + gtStartPosPtr[n] * 6, numGTBBoxes, gtBBoxes); + gtValue.getData() + gtStartPosPtr[n] * 6, numGTBBoxes, gtBBoxes); matchBBox( priorBBoxes, gtBBoxes, overlapThreshold, &matchIndices, &matchOverlaps); @@ -523,7 +530,7 @@ void getDetectionOutput(const real* confData, const size_t batchSize, const vector>>& allIndices, const vector>& allDecodedBBoxes, - MatrixPtr out) { + Matrix& out) { MatrixPtr outBuffer; Matrix::resizeOrCreate(outBuffer, numKept, 7, false, false); real* bufferData = outBuffer->getData(); @@ -550,7 +557,7 @@ void getDetectionOutput(const real* confData, } } } - out->copyFrom(bufferData, numKept * 7); + out.copyFrom(bufferData, numKept * 7); } NormalizedBBox clipBBox(const NormalizedBBox& bbox) { diff --git a/paddle/gserver/layers/DetectionUtil.h b/paddle/gserver/layers/DetectionUtil.h index e30cfa8f0..263dea001 100644 --- a/paddle/gserver/layers/DetectionUtil.h +++ b/paddle/gserver/layers/DetectionUtil.h @@ -42,10 +42,12 @@ struct BBoxBase { T getSize() const { return getWidth() * getHeight(); } + // coordinate of bounding box T xMin; T yMin; T xMax; T yMax; + // whether difficult object (e.g. object with heavy occlusion is difficult) bool isDifficult; }; @@ -53,31 +55,31 @@ struct NormalizedBBox : BBoxBase { NormalizedBBox() : BBoxBase() {} }; -enum PermMode { NCHWTONHWC, NHWCTONCHW }; +enum PermMode { kNCHWToNHWC, kNHWCToNCHW }; /** * @brief First permute input maxtrix then append to output matrix */ -size_t appendWithPermute(const MatrixPtr inMatrix, +size_t appendWithPermute(const Matrix& inMatrix, size_t height, size_t width, size_t outTotalSize, size_t outOffset, size_t batchSize, - MatrixPtr outMatrix, + Matrix& outMatrix, PermMode permMode, bool useGpu); /** * @brief First permute input maxtrix then decompose to output */ -size_t decomposeWithPermute(const MatrixPtr inMatrix, +size_t decomposeWithPermute(const Matrix& inMatrix, size_t height, size_t width, size_t totalSize, size_t offset, size_t batchSize, - MatrixPtr outMatrix, + Matrix& outMatrix, PermMode permMode, bool useGpu); @@ -89,15 +91,17 @@ size_t decomposeWithPermute(const MatrixPtr inMatrix, real jaccardOverlap(const NormalizedBBox& bbox1, const NormalizedBBox& bbox2); /** - * @brief Compute offset parameters between prior bbox and groundtruth bbox + * @brief Compute offset parameters between prior bbox and ground truth bbox * and variances of prior bbox are considered * @param priorBBox Input prior bbox * @param priorBBoxVar Variance parameters of prior bbox * @param gtBBox Groundtruth bbox + * @param outVec Output vector */ -vector encodeBBoxWithVar(const NormalizedBBox& priorBBox, - const vector priorBBoxVar, - const NormalizedBBox& gtBBox); +void encodeBBoxWithVar(const NormalizedBBox& priorBBox, + const vector priorBBoxVar, + const NormalizedBBox& gtBBox, + vector& outVec); /** * @brief Decode prior bbox with offset parameters @@ -194,9 +198,9 @@ negative examples * @param negIndicesVecPtr Save indices of negative prior bbox */ pair generateMatchIndices( - const MatrixPtr priorValue, + const Matrix& priorValue, const size_t numPriorBBoxes, - const MatrixPtr gtValue, + const Matrix& gtValue, const int* gtStartPosPtr, const size_t seqNum, const vector>& maxConfScore, @@ -298,7 +302,7 @@ void getDetectionOutput(const real* confData, const size_t batchSize, const vector>>& allIndices, const vector>& allDecodedBBoxes, - MatrixPtr out); + Matrix& out); NormalizedBBox clipBBox(const NormalizedBBox& bbox); -- GitLab