由于之前介绍过一次关于实现自己的网络层的文章,但是那篇文章偏难,这次我以最简单的对图像进行缩放的层为例进行实现。
在进行讲解之前,有一些必要条件你需要掌握,那就是你已经很了解怎么安装caffe,并且知道caffe里头的各个目录。
首先我们设计我们层所拥有的参数 out_height,即输出的图像的高度 out_width,即输出图像的宽度 visualize,是否需要将图像显示出来 那么可以在src/caffe/proto/caffe.proto文件中加入如下代码: [cpp] view plain copy message ImageScaleParameter { // Specify the output height and width optional uint32 out_height = 1; optional uint32 out_width = 2; // for debug you can see the source images and scaled images optional bool visualize = 3 [default = false]; } 这里就指定了参数的名称以及参数的类型,optional说明该参数是可选的可以出现也可以不出现,此外[default=false]表明该参数的默认值是false 每个参数都指定一个数字表明参数的标识。 接着,我们可以将我们设计好的参数放入 LayerParameter里头: [cpp] view plain copy optional HingeLossParameter hinge_loss_param = 114; optional ImageDataParameter image_data_param = 115; optional ImageScaleParameter image_scale_param = 147; optional InfogainLossParameter infogain_loss_param = 116; optional InnerProductParameter inner_product_param = 117; 注意加入的时候看一看 LayerParameter的注释,当你修改完毕了也要注意加入这样提示,这样方便后人更加方便地添加自定义层 // LayerParameter next available layer-specific ID: 148 (last added: image_scale_param) 接下来我们实现我们自己的层的头文件: (1)实现的首先需要设置不允许头文件重复加入的宏定义: [cpp] view plain copy #ifndef CAFFE_IMAGE_SCALE_LAYER_HPP_ #define CAFFE_IMAGE_SCALE_LAYER_HPP_ (2)加入必要的头文件 [cpp] view plain copy #include "caffe/blob.hpp" #include "caffe/layer.hpp" #include "caffe/proto/caffe.pb.h" #include "caffe/layer.hpp" (3)加入返回的层的类型字符串 [cpp] view plain copy virtual inline const char* type() const { return "ImageScale"; } (4)告诉caffe本层的输入有几个,输出有几个 [cpp] view plain copy virtual inline int ExactNumBottomBlobs() const { return 1; } virtual inline int ExactNumTopBlobs() const { return 1; } (5)由于本层实现是图像的缩放,所以不需要反传,因此直接写一个空的虚函数的实现 [cpp] view plain copy virtual void Backward_cpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}; (6)定义在使用过程中所使用的类中的成员变量,注意类的成员变量的命名最后是以下划线结束,这样能够保持与caffe的代码一致性 [cpp] view plain copy int out_height_; int out_width_; int height_; int width_; bool visualize_; int num_images_; int num_channels_; (7)最后别忘记加入endif这个宏,此外注意加入必要的注释,以表明这个endif所对应的开头是什么 [cpp] view plain copy #endif // CAFFE_IMAGE_SCALE_LAYER_HPP_ 下面给出详细的头文件代码: [cpp] view plain copy #ifndef CAFFE_IMAGE_SCALE_LAYER_HPP_ #define CAFFE_IMAGE_SCALE_LAYER_HPP_ #include "caffe/blob.hpp" #include "caffe/layer.hpp" #include "caffe/proto/caffe.pb.h" #include "caffe/layer.hpp" namespace caffe { // written by xizero00 2016/9/13 template <typename Dtype> class ImageScaleLayer : public Layer<Dtype> { public: explicit ImageScaleLayer(const LayerParameter& param) : Layer<Dtype>(param) {} virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top); virtual void Reshape(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top); virtual inline const char* type() const { return "ImageScale"; } virtual inline int ExactNumBottomBlobs() const { return 1; } virtual inline int ExactNumTopBlobs() const { return 1; } protected: /// @copydoc ImageScaleLayer virtual void Forward_cpu(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top); virtual void Backward_cpu(const vector<Blob<Dtype>*>& top, const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {}; int out_height_; int out_width_; int height_; int width_; bool visualize_; int num_images_; int num_channels_; }; } // namespace caffe #endif // CAFFE_IMAGE_SCALE_LAYER_HPP_ 接下来写具体的层的设置以及层的前传的实现: (8)加入必要的头文件 [cpp] view plain copy #include "caffe/layers/image_scale_layer.hpp" #include "caffe/util/math_functions.hpp" #include <opencv2/opencv.hpp> (9)实现层的设置函数 LayerSetUp,在该函数中将网络的配置参数读取到类中的成员变量中,便于前传的时候以及对层进行设置的时候使用,并且检查参数的合法性 [cpp] view plain copy template <typename Dtype> void ImageScaleLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) { // get parameters const ImageScaleParameter& param = this->layer_param_.image_scale_param(); // get the output size out_height_ = param.out_height(); out_width_ = param.out_width(); visualize_ = param.visualize(); // get the input size num_images_ = bottom[0]->num(); height_ = bottom[0]->height(); width_ = bottom[0]->width(); num_channels_ = bottom[0]->channels(); // check the channels must be images // channel must be 1 or 3, gray image or color image CHECK_EQ( (num_channels_==3) || (num_channels_ == 1), true); // check the output size CHECK_GT(out_height_, 0); CHECK_GT(out_height_, 0); } (10)实现层的Reshape函数,来设定该层的输出的大小,我们使用从网络配置文件中的参数类设置输出的大小 [cpp] view plain copy template <typename Dtype> void ImageScaleLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) { // reshape the outputs top[0]->Reshape(num_images_, num_channels_, out_height_, out_width_); } (11)实现前向传播函数 Forward_cpu,我实现的就是将图像一幅一幅地进行缩放到配置文件中所给的大小。 [cpp] view plain copy template <typename Dtype> void ImageScaleLayer<Dtype>::Forward_cpu( const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) { const Dtype* bottom_data = bottom[0]->cpu_data(); Dtype * top_data = top[0]->mutable_cpu_data(); cv::Mat srcimage, dstimage; // precompurte the index const int srcimagesize = width_ * height_; const int dstimagesize = out_width_ * out_height_; const int srcchimagesize = srcimagesize * num_channels_; const int dstchimagesize = dstimagesize * num_channels_; for ( int idx_img = 0; idx_img < num_images_; idx_img++ ) { // zeros source images and scaled images srcimage = cv::Mat::zeros(height_, width_, CV_32FC1); dstimage = cv::Mat::zeros(out_height_, out_width_, CV_32FC1); // read from bottom[0] for ( int idx_ch = 0; idx_ch < num_channels_; idx_ch++ ) { for (int i = 0; i < height_; i++) { for ( int j=0; j < width_; j++ ) { int image_idx = idx_img * srcchimagesize + srcimagesize * idx_ch + height_ *i + j; srcimage.at<float>(i,j) = (float)bottom_data[image_idx]; } } } // resize to specified size // here we use linear interpolation cv::resize(srcimage, dstimage, dstimage.size()); // store the resized image to top[0] for (int idx_ch = 0; idx_ch < num_channels_; idx_ch++) { for (int i = 0; i < out_height_; i++) { for (int j = 0; j < out_width_; j++) { int image_idx = idx_img * dstchimagesize + dstimagesize * idx_ch + out_height_*i + j; top_data[image_idx] = dstimage.at<float>(i,j); } } } if (visualize_) { cv::namedWindow("src image", CV_WINDOW_AUTOSIZE); cv::namedWindow("dst image", CV_WINDOW_AUTOSIZE); cv::imshow("src image", srcimage); cv::imshow("dst image", dstimage); cv::waitKey(0); } } } 最后给出完整的实现: [cpp] view plain copy #include "caffe/layers/image_scale_layer.hpp" #include "caffe/util/math_functions.hpp" #include <opencv2/opencv.hpp> namespace caffe { template <typename Dtype> void ImageScaleLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) { // get parameters const ImageScaleParameter& param = this->layer_param_.image_scale_param(); // get the output size out_height_ = param.out_height(); out_width_ = param.out_width(); visualize_ = param.visualize(); // get the input size num_images_ = bottom[0]->num(); height_ = bottom[0]->height(); width_ = bottom[0]->width(); num_channels_ = bottom[0]->channels(); // check the channels must be images // channel must be 1 or 3, gray image or color image CHECK_EQ( (num_channels_==3) || (num_channels_ == 1), true); // check the output size CHECK_GT(out_height_, 0); CHECK_GT(out_height_, 0); } template <typename Dtype> void ImageScaleLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) { // reshape the outputs top[0]->Reshape(num_images_, num_channels_, out_height_, out_width_); } template <typename Dtype> void ImageScaleLayer<Dtype>::Forward_cpu( const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) { const Dtype* bottom_data = bottom[0]->cpu_data(); Dtype * top_data = top[0]->mutable_cpu_data(); cv::Mat srcimage, dstimage; // precompurte the index const int srcimagesize = width_ * height_; const int dstimagesize = out_width_ * out_height_; const int srcchimagesize = srcimagesize * num_channels_; const int dstchimagesize = dstimagesize * num_channels_; for ( int idx_img = 0; idx_img < num_images_; idx_img++ ) { // zeros source images and scaled images srcimage = cv::Mat::zeros(height_, width_, CV_32FC1); dstimage = cv::Mat::zeros(out_height_, out_width_, CV_32FC1); // read from bottom[0] for ( int idx_ch = 0; idx_ch < num_channels_; idx_ch++ ) { for (int i = 0; i < height_; i++) { for ( int j=0; j < width_; j++ ) { int image_idx = idx_img * srcchimagesize + srcimagesize * idx_ch + height_ *i + j; srcimage.at<float>(i,j) = (float)bottom_data[image_idx]; } } } // resize to specified size // here we use linear interpolation cv::resize(srcimage, dstimage, dstimage.size()); // store the resized image to top[0] for (int idx_ch = 0; idx_ch < num_channels_; idx_ch++) { for (int i = 0; i < out_height_; i++) { for (int j = 0; j < out_width_; j++) { int image_idx = idx_img * dstchimagesize + dstimagesize * idx_ch + out_height_*i + j; top_data[image_idx] = dstimage.at<float>(i,j); } } } if (visualize_) { cv::namedWindow("src image", CV_WINDOW_AUTOSIZE); cv::namedWindow("dst image", CV_WINDOW_AUTOSIZE); cv::imshow("src image", srcimage); cv::imshow("dst image", dstimage); cv::waitKey(0); } } } #ifdef CPU_ONLY STUB_GPU(ImageScaleLayer); #endif INSTANTIATE_CLASS(ImageScaleLayer); REGISTER_LAYER_CLASS(ImageScale); } // namespace caffe 请把上述代码,保存为image_scale_layer.hpp和cpp。然后放入到对应的include和src/caffe/layers文件夹中。 那么在使用的时候可以进行如下配置 [cpp] view plain copy layer { name: "imagescaled" type: "ImageScale" bottom: "data" top: "imagescaled" image_scale_param { out_height: 128 out_width: 128 visualize: true } } 上述配置中 out_height和out_width就是经过缩放之后的图片的大小,而visualize表明是否显示的意思。 至此,我们就完成了一个很简单的caffe自定义层的实现,怎么样,很简单吧? 我测试的模型(我想你肯定知道怎么用caffe所听的工具将mnist数据集转换为lmdb吧)是: [cpp] view plain copy # Simple single-layer network to showcase editing model parameters. name: "sample" layer { name: "data" type: "Data" top: "data" include { phase: TRAIN } transform_param { scale: 0.0039215684 } data_param { source: "examples/mnist/mnist_train_lmdb" batch_size: 10 backend: LMDB } } layer { name: "imagescaled" type: "ImageScale" bottom: "data" top: "imagescaled" image_scale_param { out_height: 128 out_width: 128 visualize: true } } 测试所用的solver.prototxt [cpp] view plain copy net: "examples/imagescale/sample.prototxt" base_lr: 0.01 lr_policy: "step" gamma: 0.1 stepsize: 10000 display: 1 max_iter: 1 weight_decay: 0.0005 snapshot: 1 snapshot_prefix: "examples/imagescale/sample" momentum: 0.9 # solver mode: CPU or GPU solver_mode: GPU 然后运行的时候仅仅需要写个bash文件到caffe的目录: [plain] view plain copy #!/usr/bin/env sh set -e snap_dir="examples/imagescale/snapshots" mkdir -p $snap_dir TOOLS=./build/tools $TOOLS/caffe train \ --solver=examples/imagescale/solver.prototxt 2>&1 | tee -a $snap_dir/train.log 下面给出我的结果: 小的是输入的原始图像,大的是经过缩放之后的图像。 好了,到此结束。 代码打包下载,请戳这里 http://download.csdn.net/detail/xizero00/9629898