基于Caffe改造的深度学习框架，为Caffe添加基于C++的网络构建、友好的Train和Inference框架

#include <algorithm> #include <csignal> #include <ctime> #include <functional> #include <map> #include <set> #include <string> #include <utility> #include <vector> #include "boost/iterator/counting_iterator.hpp" #include "caffe/util/bbox_util.hpp" namespace caffe { bool SortBBoxAscend(const NormalizedBBox& bbox1, const NormalizedBBox& bbox2) { return bbox1.score() < bbox2.score(); } bool SortBBoxDescend(const NormalizedBBox& bbox1, const NormalizedBBox& bbox2) { return bbox1.score() > bbox2.score(); } template <typename T> bool SortScorePairAscend(const pair<float, T>& pair1, const pair<float, T>& pair2) { return pair1.first < pair2.first; } // Explicit initialization. template bool SortScorePairAscend(const pair<float, int>& pair1, const pair<float, int>& pair2); template bool SortScorePairAscend(const pair<float, pair<int, int> >& pair1, const pair<float, pair<int, int> >& pair2); template <typename T> bool SortScorePairDescend(const pair<float, T>& pair1, const pair<float, T>& pair2) { return pair1.first > pair2.first; } // Explicit initialization. template bool SortScorePairDescend(const pair<float, int>& pair1, const pair<float, int>& pair2); template bool SortScorePairDescend(const pair<float, pair<int, int> >& pair1, const pair<float, pair<int, int> >& pair2); NormalizedBBox UnitBBox() { NormalizedBBox unit_bbox; unit_bbox.set_xmin(0.); unit_bbox.set_ymin(0.); unit_bbox.set_xmax(1.); unit_bbox.set_ymax(1.); return unit_bbox; } bool IsCrossBoundaryBBox(const NormalizedBBox& bbox) { return bbox.xmin() < 0 || bbox.xmin() > 1 || bbox.ymin() < 0 || bbox.ymin() > 1 || bbox.xmax() < 0 || bbox.xmax() > 1 || bbox.ymax() < 0 || bbox.ymax() > 1; } void IntersectBBox(const NormalizedBBox& bbox1, const NormalizedBBox& bbox2, NormalizedBBox* intersect_bbox) { if (bbox2.xmin() > bbox1.xmax() || bbox2.xmax() < bbox1.xmin() || bbox2.ymin() > bbox1.ymax() || bbox2.ymax() < bbox1.ymin()) { // Return [0, 0, 0, 0] if there is no intersection. intersect_bbox->set_xmin(0); intersect_bbox->set_ymin(0); intersect_bbox->set_xmax(0); intersect_bbox->set_ymax(0); } else { intersect_bbox->set_xmin(std::max(bbox1.xmin(), bbox2.xmin())); intersect_bbox->set_ymin(std::max(bbox1.ymin(), bbox2.ymin())); intersect_bbox->set_xmax(std::min(bbox1.xmax(), bbox2.xmax())); intersect_bbox->set_ymax(std::min(bbox1.ymax(), bbox2.ymax())); } } float BBoxSize(const NormalizedBBox& bbox, const bool normalized) { if (bbox.xmax() < bbox.xmin() || bbox.ymax() < bbox.ymin()) { // If bbox is invalid (e.g. xmax < xmin or ymax < ymin), return 0. return 0; } else { if (bbox.has_size()) { return bbox.size(); } else { float width = bbox.xmax() - bbox.xmin(); float height = bbox.ymax() - bbox.ymin(); if (normalized) { return width * height; } else { // If bbox is not within range [0, 1]. return (width + 1) * (height + 1); } } } } template <typename Dtype> Dtype BBoxSize(const Dtype* bbox, const bool normalized) { if (bbox[2] < bbox[0] || bbox[3] < bbox[1]) { // If bbox is invalid (e.g. xmax < xmin or ymax < ymin), return 0. return Dtype(0.); } else { const Dtype width = bbox[2] - bbox[0]; const Dtype height = bbox[3] - bbox[1]; if (normalized) { return width * height; } else { // If bbox is not within range [0, 1]. return (width + 1) * (height + 1); } } } template float BBoxSize(const float* bbox, const bool normalized); template double BBoxSize(const double* bbox, const bool normalized); void ClipBBox(const NormalizedBBox& bbox, NormalizedBBox* clip_bbox) { clip_bbox->set_xmin(std::max(std::min(bbox.xmin(), 1.f), 0.f)); clip_bbox->set_ymin(std::max(std::min(bbox.ymin(), 1.f), 0.f)); clip_bbox->set_xmax(std::max(std::min(bbox.xmax(), 1.f), 0.f)); clip_bbox->set_ymax(std::max(std::min(bbox.ymax(), 1.f), 0.f)); clip_bbox->clear_size(); clip_bbox->set_size(BBoxSize(*clip_bbox)); clip_bbox->set_difficult(bbox.difficult()); } void ClipBBox(const NormalizedBBox& bbox, const float height, const float width, NormalizedBBox* clip_bbox) { clip_bbox->set_xmin(std::max(std::min(bbox.xmin(), width), 0.f)); clip_bbox->set_ymin(std::max(std::min(bbox.ymin(), height), 0.f)); clip_bbox->set_xmax(std::max(std::min(bbox.xmax(), width), 0.f)); clip_bbox->set_ymax(std::max(std::min(bbox.ymax(), height), 0.f)); clip_bbox->clear_size(); clip_bbox->set_size(BBoxSize(*clip_bbox)); clip_bbox->set_difficult(bbox.difficult()); } void ScaleBBox(const NormalizedBBox& bbox, const int height, const int width, NormalizedBBox* scale_bbox) { scale_bbox->set_xmin(bbox.xmin() * width); scale_bbox->set_ymin(bbox.ymin() * height); scale_bbox->set_xmax(bbox.xmax() * width); scale_bbox->set_ymax(bbox.ymax() * height); scale_bbox->clear_size(); bool normalized = !(width > 1 || height > 1); scale_bbox->set_size(BBoxSize(*scale_bbox, normalized)); scale_bbox->set_difficult(bbox.difficult()); } void OutputBBox(const NormalizedBBox& bbox, const pair<int, int>& img_size, const bool has_resize, const ResizeParameter& resize_param, NormalizedBBox* out_bbox) { const int height = img_size.first; const int width = img_size.second; NormalizedBBox temp_bbox = bbox; if (has_resize && resize_param.resize_mode()) { float resize_height = resize_param.height(); CHECK_GT(resize_height, 0); float resize_width = resize_param.width(); CHECK_GT(resize_width, 0); float resize_aspect = resize_width / resize_height; int height_scale = resize_param.height_scale(); int width_scale = resize_param.width_scale(); float aspect = static_cast<float>(width) / height; float padding; NormalizedBBox source_bbox; switch (resize_param.resize_mode()) { case ResizeParameter_Resize_mode_WARP: ClipBBox(temp_bbox, &temp_bbox); ScaleBBox(temp_bbox, height, width, out_bbox); break; case ResizeParameter_Resize_mode_FIT_LARGE_SIZE_AND_PAD: if (aspect > resize_aspect) { padding = (resize_height - resize_width / aspect) / 2; source_bbox.set_xmin(0.); source_bbox.set_ymin(padding / resize_height); source_bbox.set_xmax(1.); source_bbox.set_ymax(1. - padding / resize_height); } else { padding = (resize_width - resize_height * aspect) / 2; source_bbox.set_xmin(padding / resize_width); source_bbox.set_ymin(0.); source_bbox.set_xmax(1. - padding / resize_width); source_bbox.set_ymax(1.); } ProjectBBox(source_bbox, bbox, &temp_bbox); ClipBBox(temp_bbox, &temp_bbox); ScaleBBox(temp_bbox, height, width, out_bbox); break; case ResizeParameter_Resize_mode_FIT_SMALL_SIZE: if (height_scale == 0 || width_scale == 0) { ClipBBox(temp_bbox, &temp_bbox); ScaleBBox(temp_bbox, height, width, out_bbox); } else { ScaleBBox(temp_bbox, height_scale, width_scale, out_bbox); ClipBBox(*out_bbox, height, width, out_bbox); } break; default: LOG(FATAL) << "Unknown resize mode."; } } else { // Clip the normalized bbox first. ClipBBox(temp_bbox, &temp_bbox); // Scale the bbox according to the original image size. ScaleBBox(temp_bbox, height, width, out_bbox); } } void LocateBBox(const NormalizedBBox& src_bbox, const NormalizedBBox& bbox, NormalizedBBox* loc_bbox) { float src_width = src_bbox.xmax() - src_bbox.xmin(); float src_height = src_bbox.ymax() - src_bbox.ymin(); loc_bbox->set_xmin(src_bbox