-
目录
- faster rcnn论文备注
- caffe代码框架简介
- faster rcnn代码分析
- 后记
-
faster rcnn论文备注
- 引言
faster rcnn paper是Ross Girshick在基于CNN生成region proposal提速识别方案, 主要体现在复用前面卷积后的feature map和多框一次出, feature map一路生成框结合另一路做分类.尤其是测试时计算出proposal时间消耗极小(By sharing convolutions at test-time, the marginal cost for computing proposals is small e.g., 10ms per image). -
主要组件
这个图摘自faster rcnn的论文
重要包含如下几个组件:- 输入层,仅在训练时有用.每次按照配置从一个epoch的图片拿一个批次的图片,最短边缩放到600像素.每次一个epoch完成后shuffle图片排序
- CNN层, 接收resize的图片,经过卷积和池化,通过加pad使得每次卷积后大小不变,池化后减半,最后feature map和输入图成比例关系,被后面RPN(Region Proposal Network)和ROI层复用
- RPN层(Region Proposal Network), 输入是一个feature map n×n的滑窗(论文中n = 3),输出是一组框和对应框的得分,对应VGG16网络结构一个滑窗可以覆盖228像素区域辅助上锚点(Anchor),可以翻译成9个区域.这层拆出2个loss,将框送入ROI层
- ROI层,接收RPN的输入和CNN的输入获取proposal的feature map的输入送入分类器
- 分类层,接收ROI层的feature输入给出分类的结果,这层有两个loss一个是分类的loss一个是框的loss
-
CNN层,卷基层的网络接口如下:faster RCNN卷积
共有13个卷积层后置一个relu的激活, 4个池化.这是CNN部分的caffe prototxt
可以看出每次卷积核(kernel)大小是3,垫置(pad)大小是1,从cs231n#conv中可以看出卷积后大小关系: (W - 3 + 2)/1 + 1 = W,卷积的输入宽高和输出层的宽高大小不变.池化层的参数kernel size = 2, stride = 2以极大值池化,每次池化宽高减半layer { name: "conv1_1" type: "Convolution" bottom: "data" top: "conv1_1" param { lr_mult: 0 decay_mult: 0 } param { lr_mult: 0 decay_mult: 0 } convolution_param { num_output: 64 pad: 1 kernel_size: 3 } } layer { name: "relu1_1" type: "ReLU" bottom: "conv1_1" top: "conv1_1" } layer { name: "conv1_2" type: "Convolution" bottom: "conv1_1" top: "conv1_2" param { lr_mult: 0 decay_mult: 0 } param { lr_mult: 0 decay_mult: 0 } convolution_param { num_output: 64 pad: 1 kernel_size: 3 } } layer { name: "relu1_2" type: "ReLU" bottom: "conv1_2" top: "conv1_2" } layer { name: "pool1" type: "Pooling" bottom: "conv1_2" top: "pool1" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } #中间层此处省略 # layer { name: "conv5_3" type: "Convolution" bottom: "conv5_2" top: "conv5_3" param { lr_mult: 1 } param { lr_mult: 2 } convolution_param { num_output: 512 pad: 1 kernel_size: 3 } } layer { name: "relu5_3" type: "ReLU" bottom: "conv5_3" top: "conv5_3" }
总共4个池化,最后卷积输出的通道数512(VGG16),feature map大小和输入的缩放图映射对应比例是1/16,卷基层的最终输出是'conv5_3',输入一路送入RPN算出对应的框,一路送入ROI算出对应feature map进行分类 -
Region Proposal Networks(RPN)
模型中负责生成'框'的网络, 输入是CNN中feature map中n×n的一个滑窗,输出是认为有物体的框和对应得分.一个滑窗的有效覆盖范围是228x228,经过锚点的映射后(缺省scale 和radio都是[0.5:1, 1:1, 2:1])成为9个框,下图出资论文原图针对VGG
可以看出anchor给出的框大小和横纵的适应性,通常一幅图像滑动feature map滑动窗大小是2400,anchor的总数约为20K左右(For a convolutional feature map of a size W � H (typically �2,400), there are WHk anchors intotal.) anchor设计是一个关键点,不用每次将图片resize到不同大小重新计算特征值,所有anchor的预测都是基于同一份feature(The design of multiscale anchors is a key component for sharing features without extra cost for addressing scales.)
RPN接收一个512xHxW的feature map,经过一次卷积之后甩出2路,一路用于生成K个框(2值cls, FG和BG得分),一路生成对应得分(4值bbox标识矩形框),网络结构如下:
假设原始训练图片的shape(3,h_origianl,w_origianl),每个批次一张图片,经过resize后==>(1, 3, h_resized,w_resized)经过CNN卷积池化操作之后==>(1,512,h_conv,w_conv) w_resized/16 = w_conv,h_resized/16 = h_conv ,经过'rpn_conv/3x3'(F = 3, P = 1, S = 1)后大小依然不变==>(1,512,h_conv,w_conv)但是内容已经图像卷积的feature map运算为RPN的基值(适应RPN loss从CNN的feature map做了一次转化),滑动窗的个数就等于w_conv×h_conv,所有anchor的数目是w_conv×h_conv×k(9)也就说一次rpn的卷积就完成了对全图的feature map生成proposal的过程借助GPU的并行运算能力非常省时,'rpn_conv/3x3'的输出作为'rpn_bbox_pred'和'rpn_cls_score'的输入,'rpn_cls_score'输出shape(1, 18, w_conv, h_conv), 18对应9个anchor的2个得分,因为输入blob shape(N, C, H, W)中NxHxW要等于预测/label的个数,所以这里要reshape一下(参数是shape { dim: 0 dim: 2 dim: -1 dim: 0 } ),再计算cls loss和softmax之前shape变为(1,2,9×h_conv,w_conv)可以参见softmax_loss_layer.cpp的解释:layer { name: "rpn_conv/3x3" type: "Convolution" bottom: "conv5_3" top: "rpn/output" param { lr_mult: 1.0 } param { lr_mult: 2.0 } convolution_param { num_output: 512 kernel_size: 3 pad: 1 stride: 1 weight_filler { type: "gaussian" std: 0.01 } bias_filler { type: "constant" value: 0 } } } layer { name: "rpn_relu/3x3" type: "ReLU" bottom: "rpn/output" top: "rpn/output" }得出图形所有的anchor scores一路送入计算loss一路走softmax算出FG和BG的概率.'rpn_cls_prob'输出是(1,2, 9*h_conv, w_conv),再reshape回(1,18,h_conv, w_conv)每一个window的9个anchor的概率就出来了,结合对应框送入proposal层;'rpn_conv/3x3'的另一路输出送入了'rpn_bbox_pred'算出对应的框(1,36, h_conv, w_conv),'rpn_bbox_pred'一路计算框的loss另一路送入proposal层;proposal层集合输入的概率和框生成proposal送入ROI层,整体流程如下:RPN network
这块儿比较容易乱,尤其里面层的实现还是基于python的层和c++实现的loss,对照prototxt图理解起来好很多 - Loss计算和训练
RPN loss包含两部分: score的loss和bbox的loss,引子原文
L({pi,ti}) = 1/Ncls×ΣLcls(pi, pi) + λ×1/Lreg×Σpi×Lreg(ti,ti), 其中i在mini-batch中anchor的序号,pi是第i个anchor预测是物体的概率,pi = 1 if ith anchor is ground true else 0.ti是预测的正例中矩形4值.Lcls是2值的log loss, Lreg(ti,ti) = R(ti,ti)其中R代表的是RobustLoss, λ是用于平衡两个loss的参数默认是10.其中矩形框的回归应用:ti预测框展开 tx = (x - xa)=wa; ty = (y - ya)=ha; tw = log(w/wa); th = log(h/ha); ti ground true矩形展开是 tx = (x�- xa)=wa; ty = (y* - ya)=ha;tw = log(w/wa); th = log(h/ha),其中x,y标识矩形中心坐标,w,h表示宽高,x标识预测坐标,xa标识anchor的坐标,x标识ground true的坐标,y,w,h类似.如论文所属这样的目的是'This can be thought of as bounding-box regression from an
anchor box to a nearby ground-truth box.' bounding-box regression基于同一份feature map,每个scale和radio不共享参数,独立回归一个对应的框.基于不同大小比例和横纵比的原始框和regressors卷积后得到k(9)近似ground true的框.原文如下:
图片中的anchor图像多数都是反例,造成数据不平衡,还有20k左右的anchor数目太多,随机128正例anchor和128反例,假如正例数目不够128用反例填充.论文中和代码中用的是每次训练一张图,用SGD训练.训练可以是RPN和RCNN交替训练迭代往复,也可以是合成一个大网络各自计算各自的loss,作者实验表明使用大网络训练在准确度差不多的情况下快1~1.5倍.
再有就是剔除anchor越出图片边界的,对于同一个ground true区域多个anchor都有覆盖交集(IoU)阈值设置为0.7,在采用非极大值抑制(NMS)一个图剩下的anchor大约还有2k,作者有提到NMS没有显著影响准确率而显著提升了效率.后面作者给出了切割实验给出了每一个point的效果,比如RPN和RCNN是否共享卷积层影响对比实验,再比如RPN的效果验证,把RPN替换成SS后面接上ZF/VGG16看准确率.这种类似可插拔式的实验组装思路非常好,可以验证每一个点实际cover的作用,但是往往改造起来切割实验的实现成本比较大.论文只是给了思路和点,实际在工程中具体细节还是要看代码.
- 引言
-
caffe代码框架简介
-
caffe整体结构
要了解faster rcnn的实现细节就要了解caffe的结构,以及如何定制自己的层(layer)
源码结构
主要目录结构如下:- include目录是暴露的cpp接口&class
- python是python的接口,基于封装的python和boost python将python调用翻译成cpp调用
- matlab是matlab接口层
-
src是caffe的实现层
结构如下:
-
Solver和Net的构造
Solver是一个基础类,封装caffe对外的训练和测试操作,类似tensorflow的optimizer,上面架着sgd,adam等等solver,反向传播更新参数时有些差异,除了直接构造SGDSolver类也可以通过python来创建: self.solver = caffe.SGDSolver(solver_prototxt),公共的基础操作都维护在Solver类中
以一个SGDSolver的构造过程看一下里面的结构和操作SGDSolver的构造器实现直接放进了头文件里,主要是清理一下历史,更新,临时备份的参数,主要工作都在Solver中完成template <typename Dtype> class SGDSolver : public Solver<Dtype> { public: explicit SGDSolver(const SolverParameter& param) : Solver<Dtype>(param) { PreSolve(); } explicit SGDSolver(const string& param_file) : Solver<Dtype>(param_file) { PreSolve(); } virtual inline const char* type() const { return "SGD"; } void SGDSolver<Dtype>::PreSolve() { // Initialize the history const vector<Blob<Dtype>*>& net_params = this->net_->learnable_params(); history_.clear(); update_.clear(); temp_.clear(); for (int i = 0; i < net_params.size(); ++i) { const vector<int>& shape = net_params[i]->shape(); history_.push_back(shared_ptr<Blob<Dtype> >(new Blob<Dtype>(shape))); update_.push_back(shared_ptr<Blob<Dtype> >(new Blob<Dtype>(shape))); temp_.push_back(shared_ptr<Blob<Dtype> >(new Blob<Dtype>(shape))); } } // history maintains the historical momentum data. // update maintains update related data and is not needed in snapshots. // temp maintains other information that might be needed in computation // of gradients/updates and is not needed in snapshots vector<shared_ptr<Blob<Dtype> > > history_, update_, temp_;再看Solver的构造, 默认root_solver = nullptr, void ReadSolverParamsFromTextFileOrDie(const string& param_file,SolverParameter* param) 主要是从proto反序列化为SolverParameter对象,针对历史版本做兼容,主要代码在Init中
Solver<Dtype>::Solver(const string& param_file, const Solver* root_solver) : net_(), callbacks_(), root_solver_(root_solver), requested_early_exit_(false) { SolverParameter param; ReadSolverParamsFromTextFileOrDie(param_file, ¶m); Init(param); }Init()中做了必要的初始化和检查,比如iter_和current_step_,两者关系是:this->current_step_ = this->iter_ / this->param_.stepsize();stepsize是在solver.prototxt中指定,关联学习率的修改
void Solver<Dtype>::Init(const SolverParameter& param) { CHECK(Caffe::root_solver() || root_solver_) << "root_solver_ needs to be set for all non-root solvers"; LOG_IF(INFO, Caffe::root_solver()) << "Initializing solver from parameters: " << std::endl << param.DebugString(); param_ = param; CHECK_GE(param_.average_loss(), 1) << "average_loss should be non-negative."; CheckSnapshotWritePermissions(); if (Caffe::root_solver() && param_.random_seed() >= 0) { Caffe::set_random_seed(param_.random_seed()); } // Scaffolding code InitTrainNet(); if (Caffe::root_solver()) { InitTestNets(); LOG(INFO) << "Solver scaffolding done."; } iter_ = 0; current_step_ = 0; }往下再看InitTrainNet()函数,这里写伪代码突出重点和流向,依照这log可以看出代码的流向:
solver.cpp:81] Creating training net from train_net file: models/pascal_voc/VGG16/faster_rcnn_end2end/train.prototxt
void Solver<Dtype>::InitTrainNet() { //训练部分参数的检查,包含有训练的网络参数,是否指定训练文件等等 deserialize train net file -> net_param net_.reset(new Net<Dtype>(net_param)); }重点部分在Net的初始化,抽取的伪代码如下:
至此solver -> net -> layer的初始化构造就完成了, 至于每一个layer定制的实现(卷积,池化,定制层)如何耦合进入框架稍后会有分析,整个过程图解如下:void Net<Dtype>::Init(const NetParameter& in_param) { //过滤参数 FilterNet(in_param, &filtered_param); // Create a copy of filtered_param with splits added where necessary. NetParameter param; InsertSplits(filtered_param, ¶m); memory_used_ = 0; // set the input blobs for (int input_id = 0; input_id < param.input_size(); ++input_id) { const int layer_id = -1; // inputs have fake layer ID -1,设置输入数据blob // Helper for Net::Init: add a new input or top blob to the net. (Inputs have // layer_id == -1, tops have layer_id >= 0.) //构造设置关键的变量,vector<shared_ptr<Blob<Dtype> > > blobs_( @brief the blobs storing intermediate results between the layer.) blob_names_, blob_need_backward_, net_input_blob_indices_, net_input_blobs_等等 AppendTop(param, layer_id, input_id, &available_blobs, &blob_name_to_idx); for (int layer_id = 0; layer_id < param.layer_size(); ++layer_id) { //构造每一层的layer, 这里使用类工厂的设计模型,通过宏来控制把构造函数放进注册中心,里面会设置blobs_,后面blobs_会伸出来在net以不同纬度共享引用 layers_.push_back(LayerRegistry<Dtype>::CreateLayer(layer_param)); // Figure out this layer's input and output for (int bottom_id = 0; bottom_id < layer_param.bottom_size(); ++bottom_id) { //构造每一层input blob,此处bottom_vecs_和blobs_通过指针共享blob对象 const int blob_id = AppendBottom(param, layer_id, bottom_id,&available_blobs, &blob_name_to_idx); // If a blob needs backward, this layer should provide it. need_backward |= blob_need_backward_[blob_id]; } //设置每一个layer的输出, top_vecs_和blobs_通过指针共享blob对象 for (int top_id = 0; top_id < num_top; ++top_id) { AppendTop(param, layer_id, top_id, &available_blobs,&blob_name_to_idx); } //根据网络设置layer->AutoTopBlobs(),创建自动输出的top的blob对象 //调用每一层的初始化函数 layers_[layer_id]->SetUp(bottom_vecs_[layer_id], top_vecs_[layer_id]); //根据每层内的参数是否设置了learning rate设置反向传播标致,构造每层的参数 for (int param_id = 0; param_id < num_param_blobs; ++param_id) { layers_[layer_id]->set_param_propagate_down(param_id, param_need_backward); AppendParam(param, layer_id, param_id); } } // Handle force_backward if needed. for (int layer_id = layers_.size() - 1; layer_id >= 0; --layer_id) { set layer_contributes_loss flag set layer_need_backward_ } // In the end, all remaining blobs are considered output blobs. for (set<string>::iterator it = available_blobs.begin(); it != available_blobs.end(); ++it) { net_output_blobs_.push_back(blobs_[blob_name_to_idx[*it]].get()); net_output_blob_indices_.push_back(blob_name_to_idx[*it]); } LOG_IF(INFO, Caffe::root_solver()) << "Network initialization done."; }SGDSolver构造 -
训练一次的step
网络构造完成后,就可以训练了, 一般的训练过程是:读入一批数据数据 -> 正向传播 -> 基于ground true计算loss ->反向求偏导映射到每个可以训练的layer上根据训练策略更新参数.while (cur < max_repeat){ data, result_group_true = read_data() result_calc = front_propagation(data); loss = calc_loss(result_calc, result_group_true); dws = compute_partial_derivative_4w(loss) update_w_by_strategy() }caffe把一次训练封装成一次step, SGDSolver直接调用Solver的step.抽取关键部分,代码如下:
void Solver<Dtype>::Step(int iters) { end_iter = cur + iters while (cur < end_iter){ clear_up() insert_test_if_need() hookup_before() Dtype loss = 0; for (int i = 0; i < param_.iter_size(); ++i) { loss += net_->ForwardBackward(bottom_vec); } loss /= param_.iter_size(); // average the loss across iterations for smoothed reporting,若average_loss为n：loss_容器里面就会存储前n个loss的值，而smooth_loss_相当于做了一个loss平均 UpdateSmoothedLoss(loss, start_iter, average_loss); hookup_after() ApplyUpdate(); take_snapshot_if_necessary() } }
显而易见重点就是net_的ForwardBackward(const vector<Blob<Dtype>* > & bottom)和ApplyUpdate().
首先看下Net的ForwardBackward(const vector<Blob<Dtype>* > & bottom),代码非常简单:Dtype ForwardBackward(const vector<Blob<Dtype>* > & bottom) { Dtype loss; Forward(bottom, &loss); Backward(); return loss; }这里有一个点有些奇怪, Step(int iter)中声明的vector<Blob<Dtype>*> bottom_vec;没有做任何输入直接传入了做正向传播,捋着代码看竟然把空的数据喂进了网络的输入blob 'net_input_blobs_'中,这里以faster rcnn训练网络为例, 网络里面包含了数据输入层(包括封装lmdb和做shuffle等等操作),看了下ForwardBackward()在所有测试用例里都没有额外的初始化.
net_input_blobs_等于啥都没放const vector<Blob<Dtype>*>& Net<Dtype>::Forward( const vector<Blob<Dtype>*> & bottom, Dtype* loss) { // Copy bottom to internal bottom for (int i = 0; i < bottom.size(); ++i) { net_input_blobs_[i]->CopyFrom(*bottom[i]); } return ForwardPrefilled(loss); }其中ForwardPrefilled(Dtype* loss)调用了ForwardFromTo(int start, int end),这里要做全网络的FP, 所以是*loss = ForwardFromTo(0, layers_.size() - 1);去除冗余的检查和debug信息后,代码非常凝练,这里就完成各个layer之间按照层级FG加loss的组织,各个层只要实现好自己Forward函数就好了
Dtype Net<Dtype>::ForwardFromTo(int start, int end) { for (int i = start; i <= end; ++i) { // LOG(ERROR) << "Forwarding " << layer_names_[i]; Dtype layer_loss = layers_[i]->Forward(bottom_vecs_[i], top_vecs_[i]); loss += layer_loss; } return loss; }在Forward(bottom, &loss);完成后接着进行反向传播Backward(),Backward()除了打了debug信息就调用了BackwardFromTo(layers_.size() - 1, 0);
void Net<Dtype>::BackwardFromTo(int start, int end) { for (int i = start; i >= end; --i) { if (layer_need_backward_[i]) { layers_[i]->Backward(top_vecs_[i], bottom_need_backward_[i], bottom_vecs_[i]); } } }每一层实现的函数原型是自己定制caffe layer Backward函数,从上面的loss偏导(error gradient)求出本层输入对应的偏导,propagate_down标识对应'bottom'是否计算loss偏导,标识函数原型如下:
/** * @brief Given the top blob error gradients, compute the bottom blob error * gradients. * * @param top * the output blobs, whose diff fields store the gradient of the error * with respect to themselves * @param propagate_down * a vector with equal length to bottom, with each index indicating * whether to propagate the error gradients down to the bottom blob at * the corresponding index * @param bottom * the input blobs, whose diff fields will store the gradient of the error * with respect to themselves after Backward is run * * The Backward wrapper calls the relevant device wrapper function * (Backward_cpu or Backward_gpu) to compute the bottom blob diffs given the * top blob diffs. * * Your layer should implement Backward_cpu and (optionally) Backward_gpu. */ inline void Backward(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);这样反向转一遍,bottom_vecs_中就保存着偏导信息.有一点值得注意,net_中包含全量信息(偏导,参数,中间的输入输出),bottom_vecs_指向的blobs_的某些块儿
/// @brief the blobs storing intermediate results between the layer. vector<shared_ptr<Blob<Dtype> > > blobs_; /// bottom_vecs stores the vectors containing the input for each layer. /// They don't actually host the blobs (blobs_ does), so we simply store /// pointers. vector<vector<Blob<Dtype>*> > bottom_vecs_; bottom_vecs_[layer_id].push_back(blobs_[blob_id].get());至此一次正向传播算loss,一次反向传播算error gradient就完成了,剩下的就是如何更新参数了,以简单的SGD为例
void SGDSolver<Dtype>::ApplyUpdate() { Dtype rate = GetLearningRate(); ClipGradients(); for (int param_id = 0; param_id < this->net_->learnable_params().size(); ++param_id) { Normalize(param_id); Regularize(param_id); ComputeUpdateValue(param_id, rate); } this->net_->Update(); }此处caffe里的clip gradient是什么意思？可以参考一下,大概的意思是限速,这不妨碍主流程.
对于每一个learnable的参数都是进行了一次Normalize, Regularize,然后更新参数.之前在Init时有在每一层AppendParam(net_param, layer_id, param_id);进行映射
更新参数时就是对learnable的那些blob进行axpy操作,一般在CPU模式下是调用BLAS的cblas_daxpy(N, alpha, X, 1, Y, 1),如果是GPU模式下是cublasSaxpy(Caffe::cublas_handle(), N, &alpha, X, 1, Y, 1).操作data = A*diff + data,完成参数更新:params_.push_back(layers_[layer_id]->blobs()[param_id]); if (xx condition){ ... const int learnable_param_id = learnable_params_.size(); learnable_params_.push_back(params_[net_param_id].get()); ... }blob基于error gradient更新参数至此一次迭代FG->loss&BG->update就大体清楚了
- caffe定制自己的层
- cpp定制层嵌入
之前将Solver Init的时候提到过Layer的实例化是通过类工厂里注册自己Layer的构造函数指针实现的,在Solver里只是通过一行layers_.push_back(LayerRegistry<Dtype>::CreateLayer(layer_param));就实现了
简单看下LayerRegistry的结构
LayerRegistry是注册条目,有LayerRegisterer管理,代码如下:class LayerRegistry { public: //函数指针类型定义 typedef shared_ptr<Layer<Dtype> > (*Creator)(const LayerParameter&); typedef std::map<string, Creator> CreatorRegistry; static CreatorRegistry& Registry() { //全局通过name找到构造layer函数指针 static CreatorRegistry* g_registry_ = new CreatorRegistry(); return *g_registry_; } // Adds a creator. 添加layer类型 static void AddCreator(const string& type, Creator creator) { //check exist ... registry[type] = creator; } // Get a layer using a LayerParameter.构造一个新的layer对象 static shared_ptr<Layer<Dtype> > CreateLayer(const LayerParameter& param) { //例行检查 return registry[type](param); } private: //确保单例 LayerRegistry() {} };
只要是调到了LayerRegisterer的构造器就LayerRegistry放入了类工厂,后面就可以实例化对象了.caffe就是通过宏动态生成的代码,把customer的层加入到框架里的,可以参考layer_factory.hpp的注释class LayerRegisterer { public: LayerRegisterer(const string& type, shared_ptr<Layer<Dtype> > (*creator)(const LayerParameter&)) { LayerRegistry<Dtype>::AddCreator(type, creator); } }; #define REGISTER_LAYER_CREATOR(type, creator) \ static LayerRegisterer<float> g_creator_f_##type(#type, creator<float>); \ static LayerRegisterer<double> g_creator_d_##type(#type, creator<double>) \ #define REGISTER_LAYER_CLASS(type) \ template <typename Dtype> \ shared_ptr<Layer<Dtype> > Creator_##type##Layer(const LayerParameter& param) \ { \ return shared_ptr<Layer<Dtype> >(new type##Layer<Dtype>(param)); \ } \ REGISTER_LAYER_CREATOR(type, Creator_##type##Layer)也就是在实现层cpp加入REGISTER_LAYER_CLASS宏就可以了,之前ngx build自己添加的plug in 指定cover那几个circle也是通过类似的宏手段控制编译的代码.layer_factory.hpproi_pooling_layer.cp
REGISTER_LAYER_CLASS(ROIPooling);翻译过来的代码:template <typename Dtype> shared_ptr<Layer<Dtype> > Creator_ROIPoolingLayer(const LayerParameter& param) { return shared_ptr<Layer<Dtype> >(new ROIPoolingLayer<Dtype>(param)); } //这里就调用了LayerRegisterer的构造器进而创建了LayerRegistry,这里创建一个float,一个double的 static LayerRegisterer<float> g_creator_f_ROIPooling(ROIPooling, creator<float>); static LayerRegisterer<double> g_creator_d_ROIPooling(ROIPooling, creator<double>)- 定制python层, caffe原生有一类的类型就'Python',为了方便python程序员定制自己的layer.实现的代码在PythonLayer中.通过boost python实现的,首先看一下faster rcnn中一个简单python层的定义:
以上就是一个加单的python层的定义,不涉及具体含义,先看下接口定义,和c++层一样需要实现forward,backward,setup,reshapelayer { name: 'input-data' #指定类型 type: 'Python' top: 'data' top: 'im_info' top: 'gt_boxes' python_param { #python文件 module: 'roi_data_layer.layer' #对应的class layer: 'RoIDataLayer' #传递给python的参数 param_str: "'num_classes': 21" } }
当然python层只能在cpu模式下运行,不能高效的使用GPU,使用中还是要做适当的trade offclass RoIDataLayer(caffe.Layer): def setup(self, bottom, top): """Setup the RoIDataLayer.""" layer_params = yaml.load(self.param_str_) #prototxt中定义参数传递到代码中 self._num_classes = layer_params['num_classes'] ... def forward(self, bottom, top): """Get blobs and copy them into this layer's top blob vector.""" blobs = self._get_next_minibatch() for blob_name, blob in blobs.iteritems(): top_ind = self._name_to_top_map[blob_name] # Reshape net's input blobs top[top_ind].reshape(*(blob.shape)) # Copy data into net's input blobs top[top_ind].data[...] = blob.astype(np.float32, copy=False) def backward(self, top, propagate_down, bottom): """This layer does not propagate gradients.""" pass def reshape(self, bottom, top): """Reshaping happens during the call to forward.""" pass - cpp定制层嵌入
-
-
faster rcnn代码分析
- 训练
把卷积层合并后,训练部分网络结构如下:总共loss有4部分组成RPN部分对应论文中的:L({pi,ti}) = 1/Ncls×ΣLcls(pi, pi) + λ×1/Lreg×Σpi×Lreg(ti,t*i),除了内置卷积,池化,relu激活,还有定制的python层和cpp层.
数据从input(python实现)层开始,读lmdb一个batch的图片,卷积后形成feature map一路送入RPN网络,一路送入ROI层(cpp定制实现),ROI层通RPN层送过来的proposal抽取对应proposal的feature map进行分类给出分类的loss和二次回归bbox的loss - 以python为入口的代码分析
faster rcnn训练分为stage交替训练和一个大网络统一训练,因为两者精度相仿而后者速度是前者1~1.5倍,所以本文都是一个大网络分析的.训练和测试方法在基于python+caffe的faster rcnn训练识别有过描述.首先看下训练过程是如何走进caffe的内部.训练的入口是faster_rcnn_end2end.sh脚本,主要代码如下:
训练入口在train_net.py中,测试入口在test_net.py中.抽取重要逻辑train_net.py中逻辑如下time ./tools/train_net.py --gpu ${GPU_ID} \ --solver models/${PT_DIR}/${NET}/faster_rcnn_end2end/ solver.prototxt \ --weights data/imagenet_models/${NET}.v2.caffemodel \ --imdb ${TRAIN_IMDB} \ --iters ${ITERS} \ --cfg experiments/cfgs/faster_rcnn_end2end.yml \ ${EXTRA_ARGS} time ./tools/test_net.py --gpu ${GPU_ID} \ --def models/${PT_DIR}/${NET}/faster_rcnn_end2end/t est.prototxt \ --net ${NET_FINAL} \ --imdb ${TEST_IMDB} \ --cfg experiments/cfgs/faster_rcnn_end2end.yml \ ${EXTRA_ARGS}
之前我们已经讲过了SGDSolver的初始化过程和Step流程.import caffe这一句已经包含所有需要的东西了,但是遍历caffe的python目录,也没有caffe.py这个文件, 其实import不仅可以import py文件也可以import目录,只要这个目录有__init__.py(不学习caffe还真不知道python有这个用法,可以参考下what-is-init-py-for)import caffe self.solver = caffe.SGDSolver(solver_prototxt) while self.solver.iter < max_iters: # Make one SGD update self.solver.step(1) take_snapshot_if_necessary() return model_pathspython/caffe的目录结构看下__init__.pypython/caffe
可以看出SGDSolver是从pycaffe中取得的from .pycaffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, RMSPropSolver, AdaDeltaSolver, AdamSolver from ._caffe import set_mode_cpu, set_mode_gpu, set_device, Layer, get_solver, layer_type_list, set_random_seed from ._caffe import __version__ from .proto.caffe_pb2 import TRAIN, TEST from .classifier import Classifier from .detector import Detector from . import io from .net_spec import layers, params, NetSpec, to_proto
_caffe.so是从_caffe.cpp编译出来的,看下caffe.cpp的代码,是基于boost python编译出来的python module将python函数&类映射成c++的函数&类,关键部分代码如下:from ._caffe import Net, SGDSolver, NesterovSolver, AdaGradSolver, \ RMSPropSolver, AdaDeltaSolver, AdamSolver
这样整个流程从python到c++的串联就完成了namespace bp = boost::python; // Selecting mode. void set_mode_gpu() { Caffe::set_mode(Caffe::GPU); } //所以编译出来是_caffe.so的python模块 BOOST_PYTHON_MODULE(_caffe) { //import的caffe模块属性映射 bp::scope().attr("__version__") = AS_STRING(CAFFE_VERSION); //函数映射 bp::def("set_mode_gpu", &set_mode_gpu); //类映射,python端使用默认构造器 bp::class_<Solver<Dtype>, shared_ptr<Solver<Dtype> >, boost::noncopyable>( "Solver", bp::no_init) //属性映射 .add_property("net", &Solver<Dtype>::net) .add_property("test_nets", bp::make_function(&Solver<Dtype>::test_nets, bp::return_internal_reference<>())) .add_property("iter", &Solver<Dtype>::iter) .def("solve", static_cast<void (Solver<Dtype>::*)(const char*)>( &Solver<Dtype>::Solve), SolveOverloads()) //关键函数 .def("step", &Solver<Dtype>::Step) .def("restore", &Solver<Dtype>::Restore) .def("snapshot", &Solver<Dtype>::Snapshot); //SGDSolver继承Solver,需要一个string参数构造器,explicit SGDSolver(const string& param_file) : Solver<Dtype>(param_file) { PreSolve(); } bp::class_<SGDSolver<Dtype>, bp::bases<Solver<Dtype> >, shared_ptr<SGDSolver<Dtype> >, boost::noncopyable>( "SGDSolver", bp::init<string>()); }
boost python的使用可以参考boost_python_tutorial - python layer部分
-
input-data层
这层的目的是读入数据,做预处理,输出:图片内容(index:0, name:'data');图像宽高,缩放比例(index:1, name:'im_info'); label和ground true框信息(index:2, name:'gt_box')如图所示
, data/im_info/gt_box送入'rpn-data'层出score的loss,data送入卷基层,gt_boxes送入'roi-data'层(集合proposal输出roi),im_info送入'proposal'层生成proposalinput输出
代码在roi_data_layer/layer.py中layer { name: 'input-data' type: 'Python' top: 'data' top: 'im_info' top: 'gt_boxes' python_param { module: 'roi_data_layer.layer' layer: 'RoIDataLayer' param_str: "'num_classes': N" } }
在_get_next_minibatch中,USE_PREFETCH默认是不开启的,作者发现没有太大作用('So far I haven't found this useful; likely more engineering work is required').当前拿的batch图片是否是一个新的epoch,如果是就shuffle一下,为了更好的性能shuffle的时候按照横图和纵图分组.拿到的是lmdb的项,minibatch.py中的get_minibatch获得完整数据, 这里有一个点需要注意一下, config.py和在脚本中指定的experiments/cfgs/faster_rcnn_end2end.yml融合成的配置,实际生效的配置需要再检查一下log('IMS_PER_BATCH': 1)def forward(self, bottom, top): """Get blobs and copy them into this layer's top blob vector.""" # 获得blob数据,key-value形式,按照name 设置top的输出顺序. blobs = self._get_next_minibatch() for blob_name, blob in blobs.iteritems(): top_ind = self._name_to_top_map[blob_name] # Reshape net's input blobs top[top_ind].reshape(*(blob.shape)) # Copy data into net's input blobs top[top_ind].data[...] = blob.astype(np.float32, copy=False)
这是基础输出的log辅助理解代码:def _get_next_minibatch_inds(self): """Return the roidb indices for the next minibatch.""" if self._cur + cfg.TRAIN.IMS_PER_BATCH >= len(self._roidb): self._shuffle_roidb_inds() #_perm保存的是排序的索引 db_inds = self._perm[self._cur:self._cur + cfg.TRAIN.IMS_PER_BATCH] self._cur += cfg.TRAIN.IMS_PER_BATCH return db_inds def _get_next_minibatch(self): """Return the blobs to be used for the next minibatch. If cfg.TRAIN.USE_PREFETCH is True, then blobs will be computed in a separate process and made available through self._blob_queue. """ if cfg.TRAIN.USE_PREFETCH: return self._blob_queue.get() else: #获得这个batch的lmdb索引 db_inds = self._get_next_minibatch_inds() #lmdb记录 minibatch_db = [self._roidb[i] for i in db_inds] #从对应lmdb记录转成图像数据输出,框信息 label信息,图片大小信息&缩放信息 return get_minibatch(minibatch_db, self._num_classes) def _shuffle_roidb_inds(self): """Randomly permute the training roidb.""" # Make minibatches from images that have similar aspect ratios (i.e. both tall and thin or both short and wide) in order to avoid wasting computation on zero-padding.通过横纵group避免zero padding if cfg.TRAIN.ASPECT_GROUPING: widths = np.array([r['width'] for r in self._roidb]) heights = np.array([r['height'] for r in self._roidb]) horz = (widths >= heights) vert = np.logical_not(horz) #横图 horz_inds = np.where(horz)[0] #纵图 vert_inds = np.where(vert)[0] inds = np.hstack(( np.random.permutation(horz_inds), np.random.permutation(vert_inds))) # 2个一组,绝大多数同一组的形状一致 inds = np.reshape(inds, (-1, 2)) row_perm = np.random.permutation(np.arange(inds.shape[0])) #以2个一组打算为单元重排,拉倒一层里,相邻的形状一致,之所以是两个一组,猜想是默认的__C.TRAIN.IMS_PER_BATCH = 2 inds = np.reshape(inds[row_perm, :], (-1,)) self._perm = inds else: self._perm = np.random.permutation(np.arange(len(self._roidb))) self._cur = 0
这样就返回了一batch的lmdb记录的索引,从_roi中找到对应lmdb记录,get_minibatch负责读取,以下是伪代码horz = [ True True True ..., True True True], horz = [False False False ..., False False False] horz_inds = [ 0 1 2 ..., 186205 186206 186207], vert_inds = [ 6 43 65 ..., 186176 186186 186194] inds = [163257 59770 49424 ..., 56475 31817 126653] inds = [[163257 59770] [ 49424 41168] [156295 1803] ..., [ 99367 20315] [142904 56475] [ 31817 126653]] row_perm = [77629 51661 58201 ..., 91810 47169 48787] inds = [118195 143322 121405 ..., 19415 18933 26468]
_get_image_blob在minibatch.py中, 处理缩放和把opencv imread的image数据转换成blobdef get_minibatch(roidb, num_classes): """Given a roidb, construct a minibatch sampled from it.""" num_images = len(roidb) # Sample random scales to use for each image in this batch #其实SCALES只有一个是600,这么写是为了支持缩放到多个尺寸 random_scale_inds = npr.randint(0, high=len(cfg.TRAIN.SCALES), size=num_images) #这里BATCH_SIZE = num_images, 在yml指定为1 rois_per_image = cfg.TRAIN.BATCH_SIZE / num_images fg_rois_per_image = np.round(cfg.TRAIN.FG_FRACTION * rois_per_image) # Get the input image blob, formatted for caffe # 传入lmdb记录和比例的索引 im_blob, im_scales = _get_image_blob(roidb, random_scale_inds) #数据 batch序号:C:H:W blobs = {'data': im_blob} #faster rcnn主要就是使用RPN if cfg.TRAIN.HAS_RPN: gt_inds = np.where(roidb[0]['gt_classes'] != 0)[0] gt_boxes = np.empty((len(gt_inds), 5), dtype=np.float32) #label框乘以缩放比例 = 统一缩放输入的框大小 gt_boxes[:, 0:4] = roidb[0]['boxes'][gt_inds, :] * im_scales[0] #对应分类一起赋值 gt_boxes[:, 4] = roidb[0]['gt_classes'][gt_inds] blobs['gt_boxes'] = gt_boxes #'im_info' = (H,W, im_scale) blobs['im_info'] = np.array( [[im_blob.shape[2], im_blob.shape[3], im_scales[0]]], dtype=np.float32)
prep_im_for_blob和im_list_to_blob都是util下blob的方法def _get_image_blob(roidb, scale_inds): """Builds an input blob from the images in the roidb at the specified scales. """ num_images = len(roidb) processed_ims = [] im_scales = [] for i in xrange(num_images): im = cv2.imread(roidb[i]['image']) #target_size = 600 target_size = cfg.TRAIN.SCALES[scale_inds[i]] #做缩放 返回图像&比例 im, im_scale = prep_im_for_blob(im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE) im_scales.append(im_scale) processed_ims.append(im) # Create a blob to hold the input images #做格式转换 blob = im_list_to_blob(processed_ims) return blob, im_scales
至此input层就大体清晰了,为什么之前看到前向传播时没有赋值input的blob(Dtype ForwardBackward(const vector<Blob<Dtype>* > & bottom)),因为在input python layer已经完成了read + shuffle + translate blob + scale + box info的处理def im_list_to_blob(ims): """Convert a list of images into a network input. Assumes images are already prepared (means subtracted, BGR order, ...). """ 图像的shape是H * W * 通道数, 取图像中最大的shape(np.array([(100, 5, 3), (110, 4, 3)]).max(axis=0) --> array([110, 5, 3])) max_shape = np.array([im.shape for im in ims]).max(axis=0) num_images = len(ims) blob = np.zeros((num_images, max_shape[0], max_shape[1], 3), dtype=np.float32) for i in xrange(num_images): im = ims[i] #序号:H:W:C blob[i, 0:im.shape[0], 0:im.shape[1], :] = im # Move channels (axis 3) to axis 1 # Axis order will become: (batch elem, channel, height, width) channel_swap = (0, 3, 1, 2) #交换shape的维度内的内容 blob = blob.transpose(channel_swap) return blob def prep_im_for_blob(im, pixel_means, target_size, max_size): """Mean subtract and scale an image for use in a blob.""" # type(im) = numpy array, uint8 -> float im = im.astype(np.float32, copy=False) # 减均值预处理 im -= pixel_means im_shape = im.shape im_size_min = np.min(im_shape[0:2]) im_size_max = np.max(im_shape[0:2]) #缩放比率 原图W/H * scale = 目标图像大小,短边缩放的600 im_scale = float(target_size) / float(im_size_min) # Prevent the biggest axis from being more than MAX_SIZE # 图像有最大限制,默认1000, 以上面的缩放比率是否超限,假如超限就用最大允许大小缩放 if np.round(im_scale * im_size_max) > max_size: im_scale = float(max_size) / float(im_size_max) im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale, interpolation=cv2.INTER_LINEAR) return im, im_scale- rpn-data层
rpn-data层接收的数据有:rpn_cls_score(来自rpn_cls_score层, 框的得分), gt_boxes(来自input层标注框信息), im_info(来自input层H*W,和原图缩放比例关系), proto和流向图如下:
layer { name: 'rpn-data' type: 'Python' bottom: 'rpn_cls_score' bottom: 'gt_boxes' bottom: 'im_info' bottom: 'data' top: 'rpn_labels' top: 'rpn_bbox_targets' top: 'rpn_bbox_inside_weights' top: 'rpn_bbox_outside_weights' python_param { module: 'rpn.anchor_target_layer' layer: 'AnchorTargetLayer' param_str: "'feat_stride': 16" } }rpn-data参数只有一个是步长, class是anchor_target_layer, 实现接口setup,forward, 这层是输出框和label,为下面计算loss所用,不可训练所以backward和reshape都是空实现,依次看setup代码如下:
其中generate_anchor在generate_anchor.py中,借助numpy完成def setup(self, bottom, top): layer_params = yaml.load(self.param_str_) # prototxt没指定, 默认的anchor缩放比例大小 anchor_scales = layer_params.get('scales', (8, 16, 32)) #对应一个卷积的K(9)个框, (左上坐标,右下坐标) self._anchors = generate_anchors(scales=np.array(anchor_scales)) self._num_anchors = self._anchors.shape[0] self._feat_stride = layer_params['feat_stride'] # allow boxes to sit over the edge by a small amount self._allowed_border = layer_params.get('allowed_border', 0) height, width = bottom[0].data.shape[-2:] A = self._num_anchors # labels top[0].reshape(1, 1, A * height, width) # bbox_targets top[1].reshape(1, A * 4, height, width) # bbox_inside_weights top[2].reshape(1, A * 4, height, width) # bbox_outside_weights top[3].reshape(1, A * 4, height, width)
接下来是forward,代码比较复杂,抽取伪代码看思路和方法.def generate_anchors(base_size=16, ratios=[0.5, 1, 2], scales=2**np.arange(3, 6)): """ Generate anchor (reference) windows by enumerating aspect ratios X scales wrt a reference (0, 0, 15, 15) window. """ # base anchor :np array [0,0, 15, 15] base_anchor = np.array([1, 1, base_size, base_size]) - 1 # 宽高比扩展:纵框,平框,横框 ratio_anchors = _ratio_enum(base_anchor, ratios) # 在base anchor大小的基础上针对大小扩展: x8, x16, x32 anchors = np.vstack([_scale_enum(ratio_anchors[i, :], scales) for i in xrange(ratio_anchors.shape[0])]) return anchors def _ratio_enum(anchor, ratios): """ Enumerate a set of anchors for each aspect ratio wrt an anchor. """ #转换成w,h,中心坐标 w, h, x_ctr, y_ctr = _whctrs(anchor) #原始面积 size = w * h #base anchor是一个正方形,假设边长为n, new w = n/(√radio), new h = n*√radio,新的边长具有如下特点:面积大体不变(忽略上下round的损失),w/h = radio,也就说这样计算完在面积大体不变的情况下:实现宽高按照raio设定的比例走,有点像拉长和压扁 size_ratios = size / ratios ws = np.round(np.sqrt(size_ratios)) hs = np.round(ws * ratios) #转成坐标形式,_whctrs的逆操作 anchors = _mkanchors(ws, hs, x_ctr, y_ctr) return anchors #按照面积比例扩展,实际是scales元素的平方扩展 def _scale_enum(anchor, scales): """ Enumerate a set of anchors for each scale wrt an anchor. """ w, h, x_ctr, y_ctr = _whctrs(anchor) ws = w * scales hs = h * scales anchors = _mkanchors(ws, hs, x_ctr, y_ctr) return anchorsdef forward(self, bottom, top): # Algorithm: # # for each (H, W) location i # generate 9 anchor boxes centered on cell i # apply predicted bbox deltas at cell i to each of the 9 anchors # filter out-of-image anchors # measure GT overlap assert bottom[0].data.shape[0] == 1, \ 'Only single item batches are supported' # map of shape (..., H, W),此处是框的得分,reshape = (1,18,H,W) height, width = bottom[0].data.shape[-2:] # GT boxes (x1, y1, x2, y2, label) gt_boxes = bottom[1].data # im_info im_info = bottom[2].data[0, :] # 1. Generate proposals from bbox deltas and shifted anchors # 这块的思路是生成一系列的shift, 然后每一个shift和9个anchor想加,迭代出每一个位置的9个框 shift_x = np.arange(0, width) * self._feat_stride shift_y = np.arange(0, height) * self._feat_stride shift_x, shift_y = np.meshgrid(shift_x, shift_y) #经过meshgrid shift_x = [[ 0 16 32 ..., 560 576 592] [ 0 16 32 ..., 560 576 592] [ 0 16 32 ..., 560 576 592] ..., [ 0 16 32 ..., 560 576 592] [ 0 16 32 ..., 560 576 592] [ 0 16 32 ..., 560 576 592]] #shift_y = [[ 0 0 0 ..., 0 0 0] [ 16 16 16 ..., 16 16 16] [ 32 32 32 ..., 32 32 32] ..., [560 560 560 ..., 560 560 560] [576 576 576 ..., 576 576 576] [592 592 592 ..., 592 592 592]] shifts = np.vstack((shift_x.ravel(), shift_y.ravel(), shift_x.ravel(), shift_y.ravel())).transpose() #转至之后形成所有位移 # add A anchors (1, A, 4) to # cell K shifts (K, 1, 4) to get # shift anchors (K, A, 4) # reshape to (K*A, 4) shifted anchors A = self._num_anchors K = shifts.shape[0] # numpy array + 操作_anchors中每一个anchor和每一个shift想加等出结果 all_anchors = (self._anchors.reshape((1, A, 4)) + shifts.reshape((1, K, 4)).transpose((1, 0, 2))) #K个位移,每个位移A个框 all_anchors = all_anchors.reshape((K * A, 4)) total_anchors = int(K * A) # only keep anchors inside the image,框在图片内 inds_inside = np.where( (all_anchors[:, 0] >= -self._allowed_border) & (all_anchors[:, 1] >= -self._allowed_border) & (all_anchors[:, 2] < im_info[1] + self._allowed_border) & # width (all_anchors[:, 3] < im_info[0] + self._allowed_border) # height )[0] # keep only inside anchors anchors = all_anchors[inds_inside, :] # label: 1 is positive, 0 is negative, -1 is dont care labels = np.empty((len(inds_inside), ), dtype=np.float32) labels.fill(-1) # overlaps between the anchors and the gt boxes # overlaps (ex, gt), 每个框对应每个box的重合面积,overlaps [anchor数目,box数目] overlaps = bbox_overlaps( np.ascontiguousarray(anchors, dtype=np.float), np.ascontiguousarray(gt_boxes, dtype=np.float)) # 针对每一个anchor内覆盖率最高的索引 argmax_overlaps = overlaps.argmax(axis=1) # 从索引取覆盖率, 每一个anchor覆盖最大的box的覆盖率 max_overlaps = overlaps[np.arange(len(inds_inside)), argmax_overlaps] # 从box出发覆盖最好的anchor的索引 gt_argmax_overlaps = overlaps.argmax(axis=0) #取覆盖最好的anchor全部box的覆盖值 gt_max_overlaps = overlaps[gt_argmax_overlaps, np.arange(overlaps.shape[1])] #match的anchor gt_argmax_overlaps = np.where(overlaps == gt_max_overlaps)[0] if not cfg.TRAIN.RPN_CLOBBER_POSITIVES: # assign bg labels first so that positive labels can clobber them labels[max_overlaps < cfg.TRAIN.RPN_NEGATIVE_OVERLAP] = 0 # fg label: for each gt, anchor with highest overlap labels[gt_argmax_overlaps] = 1 # fg label: above threshold IOU labels[max_overlaps >= cfg.TRAIN.RPN_POSITIVE_OVERLAP] = 1 if cfg.TRAIN.RPN_CLOBBER_POSITIVES: # assign bg labels last so that negative labels can clobber positives labels[max_overlaps < cfg.TRAIN.RPN_NEGATIVE_OVERLAP] = 0 # subsample positive labels if we have too many #最好是各FG,BG占一半,FG不足BG补充 num_fg = int(cfg.TRAIN.RPN_FG_FRACTION * cfg.TRAIN.RPN_BATCHSIZE) fg_inds = np.where(labels == 1)[0] if len(fg_inds) > num_fg: disable_inds = npr.choice( fg_inds, size=(len(fg_inds) - num_fg), replace=False) labels[disable_inds] = -1 # subsample negative labels if we have too many num_bg = cfg.TRAIN.RPN_BATCHSIZE - np.sum(labels == 1) bg_inds = np.where(labels == 0)[0] if len(bg_inds) > num_bg: disable_inds = npr.choice( bg_inds, size=(len(bg_inds) - num_bg), replace=False) labels[disable_inds] = -1 # 算出anchor和ground true box的dx,dy, dw,dh的偏差 bbox_targets = _compute_targets(anchors, gt_boxes[argmax_overlaps, :]) bbox_inside_weights = np.zeros((len(inds_inside), 4), dtype=np.float32) bbox_inside_weights[labels == 1, :] = np.array(cfg.TRAIN.RPN_BBOX_INSIDE_WEIGHTS) bbox_outside_weights = np.zeros((len(inds_inside), 4), dtype=np.float32) if cfg.TRAIN.RPN_POSITIVE_WEIGHT < 0: # uniform weighting of examples (given non-uniform sampling) num_examples = np.sum(labels >= 0) positive_weights = np.ones((1, 4)) * 1.0 / num_examples negative_weights = np.ones((1, 4)) * 1.0 / num_examples else: assert ((cfg.TRAIN.RPN_POSITIVE_WEIGHT > 0) & (cfg.TRAIN.RPN_POSITIVE_WEIGHT < 1)) positive_weights = (cfg.TRAIN.RPN_POSITIVE_WEIGHT / np.sum(labels == 1)) negative_weights = ((1.0 - cfg.TRAIN.RPN_POSITIVE_WEIGHT) / np.sum(labels == 0)) bbox_outside_weights[labels == 1, :] = positive_weights bbox_outside_weights[labels == 0, :] = negative_weights # map up to original set of anchors labels = _unmap(labels, total_anchors, inds_inside, fill=-1) bbox_targets = _unmap(bbox_targets, total_anchors, inds_inside, fill=0) bbox_inside_weights = _unmap(bbox_inside_weights, total_anchors, inds_inside, fill=0) bbox_outside_weights = _unmap(bbox_outside_weights, total_anchors, inds_inside, fill=0) # labels labels = labels.reshape((1, height, width, A)).transpose(0, 3, 1, 2) labels = labels.reshape((1, 1, A * height, width)) top[0].reshape(*labels.shape) top[0].data[...] = labels # bbox_targets bbox_targets = bbox_targets \ .reshape((1, height, width, A * 4)).transpose(0, 3, 1, 2) top[1].reshape(*bbox_targets.shape) top[1].data[...] = bbox_targets # bbox_inside_weights bbox_inside_weights = bbox_inside_weights \ .reshape((1, height, width, A * 4)).transpose(0, 3, 1, 2) assert bbox_inside_weights.shape[2] == height assert bbox_inside_weights.shape[3] == width top[2].reshape(*bbox_inside_weights.shape) top[2].data[...] = bbox_inside_weights # bbox_outside_weights bbox_outside_weights = bbox_outside_weights \ .reshape((1, height, width, A * 4)).transpose(0, 3, 1, 2) assert bbox_outside_weights.shape[2] == height assert bbox_outside_weights.shape[3] == width top[3].reshape(*bbox_outside_weights.shape) top[3].data[...] = bbox_outside_weights -
- 训练
* proposal层
* roi-data层
- c++ layer & loss(未完待续...)
- SmoothL1LossLayer层
- ROIPoolingLayer层
- 测试
-
后记
看到讲解faster rcnn的文章无一都要陌拜一下Ross Girshick大神,这里我也膜拜一下,确实厉害.论文写得非常有深度
该算法不是一蹴而就的,经历了rcnn -> fast rcnn ->faser rcnn. faster最大的特点是anchor的设计,不用resize基于相同feature map的regressor出不同,一次运算就出了所有的proposal.
在学习RL的时候就有点惊讶,他们那CNN出来的东西想让它是啥就是啥,然后用loss去修饰它,它就有了合理的解释,把网络拆分,不同部分有不同的含义,还是用不同loss去修饰它们
feature map从原图开始W,H在翻倍减小,维度在翻倍增加,然后map回头映射到输入点阵上,从输入图像上去预测框感觉有点玄妙,因为一个随便图可以有各式各样,给它合理的loss它就合理了
最后作者还给除了切割实验,把算法中的component替换验证其必要性着实严禁
也借着学习faster过程,窥探了一下caffe的结构,caffe代码框架清晰,比较干净不求大而全,代码也比较简洁,对有深度学习知识的人非常容易上手,这大概就是为啥Ross Girshick要基于caffe写faster rcnn的demo.初次学习一个陌生的框架还是要着眼全局不要过分计较一个局部的细节,全局通顺会带来更多的信息,信息的增多会细节的了解更加有帮助.
caffe在cpu环境下加速运算也是一个非常有意思而且有意义的问题,因为很多情况下GPU设置太大太贵在很多环境不合适
后面还有faster rcnn定制的python和cpp层的备注还没有写,抽空赶紧补上
