Caffe中用C++自定义新的layer
1.如何在caffe中增加新的layer
1)在./src/caffe/proto/caffe.proto 中增加对应layer的paramter message;
2)在./include/caffe/XXXlayers.hpp中增加该layer的类的声明,XXX表示有 common_layers.hpp, data_layers.hpp, neuron_layers.hpp, vision_layers.hpp 和loss_layers.hpp等;
3)在./src/caffe/layers/目录下新建.cpp和.cu文件,进行类实现;
4)在./src/caffe/gtest/中增加layer的测试代码,对所写的layer前传和反传进行测试,测试还包括速度。
Triplet loss原理以及梯度推导
最近,learning to rank 的思想逐渐被应用到很多领域,比如google用来做人脸识别(faceNet),微软Jingdong Wang 用来做 person-reid 等等。learning to rank中其中重要的一个步骤就是找到一个好的similarity function,而triplet loss是用的非常广泛的一种。
如上图所示,triplet是一个三元组,这个三元组是这样构成的:从训练数据集中随机选一个样本,该样本称为Anchor,然后再随机选取一个和 Anchor (记为x_a)属于同一类的样本和不同类的样本,这两个样本对应的称为Positive (记为$x_{p}$) 和Negative (记为$x_{n}$),由此构成一个(Anchor,Positive,Negative)三元组。
有了上面的triplet的概念, triplet loss就好理解了。针对三元组中的每个元素(样本),训练一个参数共享或者不共享的网络,得到三个元素的特征表达,分别记为:$f\left ( x_{i}^{a} \right )$,$f\left ( x_{i}^{p} \right )$,$f\left ( x_{i}^{n} \right )$。triplet loss的目的就是通过学习,让$x_{a}$和$x_{p}$特征表达之间的距离尽可能小,而$x_{a}$和$x_{n}$的特征表达之间的距离尽可能大,并且要让$x_{a}$与$x_{n}$之间的距离和$x_{a}$与$x_{p}$之间的距离之间有一个最小的间隔$\alpha$。公式化的表示就是:
对应的目标函数也就很清楚了:
这里距离用欧式距离度量,+表示[]内的值大于零的时候,取该值为损失,小于零的时候,损失为零。
由目标函数可以看出:
1)当$x_{a}$与$x_{n}$之间的距离小于$x_{a}$与$x_{p}$之间的距离加$\alpha$时,[]内的值大于零,就会产生损失。
2)当$x_{a}$与$x_{n}$之间的距离大于等于$x_{a}$与$x_{p}$之间的距离加$\alpha$时,损失为零。
triplet loss 梯度推导:
上述目标函数记为$L$。则当第$i$个triplet损失大于零的时候,仅就上述公式而言,有:
算法实现Trick:
可以看到,对$x_{p}$和$x_{n}$特征表达的梯度刚好利用了求损失时候的中间结果,给的启示就是,如果在CNN中实现 triplet loss layer, 如果能够在前向传播中存储着两个中间结果,反向传播的时候就能避免重复计算。这仅仅是算法实现时候的一个Trick。
在caffe中实现Triplet loss layer
caffe.proto中增加Triplet loss layer的定义
首先在message LayerParameter中追加 optional TripletLossParameter Triplet_loss_param = 137; 其中137是目前LayerParameter message中现有元素的个数,具体是多少,可以看LayerParameter message上面注释中的:
// NOTE
// Update the next available ID when you add a new LayerParameter field.
//
// LayerParameter next available layer-specific ID: 137 (last added: reduction_param)
然后增加Message:
message TripletLossParameter {
// margin for dissimilar pair
optional float margin = 1 [default = 1.0];
}
其中 margin就是定义Triplet loss原理以及梯度推导所讲的$\alpha$。
在./include/caffe/loss_layers.hpp中增加Triplet loss layer的类的声明
具体解释见注释,主要的是定义了一些变量,用来在前传中存储中间计算结果,以便在反传的时候避免重复计算。
/**
* @brief Computes the Tripletloss
*/
template <typename Dtype>
class TripletLossLayer : public LossLayer<Dtype> {
public:
explicit TripletLossLayer(const LayerParameter& param)
: LossLayer<Dtype>(param){}
virtual void LayerSetUp(const vector<Blob<Dtype>*>&bottom,
const vector<Blob<Dtype>*>& top);
virtual inline int ExactNumBottomBlobs() const { return 4; }
virtual inline const char* type() const { return "TripletLoss";}
/**
* Unlike most loss layers, in the TripletLossLayer we can backpropagate
* to the first three inputs.
*/
virtual inline bool AllowForceBackward(const int bottom_index) const {
return bottom_index != 3;
}
protected:
virtual void Forward_cpu(const vector<Blob<Dtype>*>&bottom,
const vector<Blob<Dtype>*>& top);
virtual void Forward_gpu(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);
virtual void Backward_gpu(const vector<Blob<Dtype>*>&top,
const vector<bool>&propagate_down, const vector<Blob<Dtype>*>& bottom);
Blob<Dtype> diff_ap_; //cached for backward pass
Blob<Dtype> diff_an_; //cached for backward pass
Blob<Dtype> diff_pn_; //cached for backward pass
Blob<Dtype> diff_sq_ap_; //cached for backward pass
Blob<Dtype> diff_sq_an_; //tmp storage for gpu forward pass
Blob<Dtype> dist_sq_ap_; //cached for backward pass
Blob<Dtype> dist_sq_an_; //cached for backward pass
Blob<Dtype> summer_vec_; //tmp storage for gpu forward pass
Blob<Dtype> dist_binary_; // tmp storage for gpu forward pass
};
在./src/caffe/layers/目录下新建Triplet_loss_layer.cpp,实现类
主要实现三个功能:
LayerSetUp:主要是做一些CHECK工作,然后根据bottom和top对类中的数据成员初始化。
Forward_cpu:前传,计算loss
Backward_cpu:反传,计算梯度。
/*
* Triplet_loss_layer.cpp
*/
#include <algorithm>
#include <vector>
#include "caffe/layer.hpp"
#include"caffe/loss_layers.hpp"
#include"caffe/util/io.hpp"
#include"caffe/util/math_functions.hpp"
namespace caffe {
template <typename Dtype>
void TripletLossLayer<Dtype>::LayerSetUp(
const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>&top) {
LossLayer<Dtype>::LayerSetUp(bottom, top);
CHECK_EQ(bottom[0]->num(), bottom[1]->num());
CHECK_EQ(bottom[1]->num(), bottom[2]->num());
CHECK_EQ(bottom[0]->channels(), bottom[1]->channels());
CHECK_EQ(bottom[1]->channels(), bottom[2]->channels());
CHECK_EQ(bottom[0]->height(), 1);
CHECK_EQ(bottom[0]->width(), 1);
CHECK_EQ(bottom[1]->height(), 1);
CHECK_EQ(bottom[1]->width(), 1);
CHECK_EQ(bottom[2]->height(), 1);
CHECK_EQ(bottom[2]->width(), 1);
CHECK_EQ(bottom[3]->channels(),1);
CHECK_EQ(bottom[3]->height(), 1);
CHECK_EQ(bottom[3]->width(), 1);
diff_ap_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
diff_an_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
diff_pn_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
diff_sq_ap_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1,1);
diff_sq_an_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1,1);
dist_sq_ap_.Reshape(bottom[0]->num(), 1, 1, 1);
dist_sq_an_.Reshape(bottom[0]->num(), 1, 1, 1);
// vector of ones used to sum along channels
summer_vec_.Reshape(bottom[0]->channels(), 1, 1, 1);
for (int i = 0; i < bottom[0]->channels(); ++i)
summer_vec_.mutable_cpu_data()[i] = Dtype(1);
dist_binary_.Reshape(bottom[0]->num(), 1, 1, 1);
for (int i = 0; i < bottom[0]->num();++i)
dist_binary_.mutable_cpu_data()[i]= Dtype(1);
}
template <typename Dtype>
void TripletLossLayer<Dtype>::Forward_cpu(
const vector<Blob<Dtype>*>&bottom,
const vector<Blob<Dtype>*>&top) {
int count = bottom[0]->count();
const Dtype* sampleW = bottom[3]->cpu_data();
caffe_sub(
count,
bottom[0]->cpu_data(), // a
bottom[1]->cpu_data(), // p
diff_ap_.mutable_cpu_data()); // a_i-p_i
caffe_sub(
count,
bottom[0]->cpu_data(), // a
bottom[2]->cpu_data(), // n
diff_an_.mutable_cpu_data()); // a_i-n_i
caffe_sub(
count,
bottom[1]->cpu_data(), // p
bottom[2]->cpu_data(), // n
diff_pn_.mutable_cpu_data()); // p_i-n_i
const int channels = bottom[0]->channels();
Dtype margin = this->layer_param_.triplet_loss_param().margin();
Dtype loss(0.0);
for (int i = 0; i < bottom[0]->num(); ++i) {
dist_sq_ap_.mutable_cpu_data()[i] =caffe_cpu_dot(channels,
diff_ap_.cpu_data() + (i*channels),diff_ap_.cpu_data() + (i*channels));
dist_sq_an_.mutable_cpu_data()[i] =caffe_cpu_dot(channels,
diff_an_.cpu_data() + (i*channels),diff_an_.cpu_data() + (i*channels));
Dtype mdist = sampleW[i]*std::max(margin +dist_sq_ap_.cpu_data()[i] - dist_sq_an_.cpu_data()[i], Dtype(0.0));
loss += mdist;
if(mdist==Dtype(0)){
//dist_binary_.mutable_cpu_data()[i]= Dtype(0);
//preparefor backward pass
caffe_set(channels,Dtype(0), diff_ap_.mutable_cpu_data() + (i*channels));
caffe_set(channels,Dtype(0), diff_an_.mutable_cpu_data() + (i*channels));
caffe_set(channels,Dtype(0), diff_pn_.mutable_cpu_data() + (i*channels));
}
}
loss = loss / static_cast<Dtype>(bottom[0]->num()) / Dtype(2);
top[0]->mutable_cpu_data()[0] = loss;
}
template <typename Dtype>
void TripletLossLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,
const vector<bool>&propagate_down, const vector<Blob<Dtype>*>& bottom) {
//Dtype margin =this->layer_param_.triplet_loss_param().margin();
const Dtype* sampleW = bottom[3]->cpu_data();
for (int i = 0; i < 3; ++i) {
if (propagate_down[i]) {
const Dtype sign = (i < 2) ? -1 : 1;
const Dtype alpha = sign *top[0]->cpu_diff()[0] /
static_cast<Dtype>(bottom[i]->num());
int num = bottom[i]->num();
int channels = bottom[i]->channels();
for (int j = 0; j < num; ++j) {
Dtype* bout =bottom[i]->mutable_cpu_diff();
if (i==0) { // a
//if(dist_binary_.cpu_data()[j]>Dtype(0)){
caffe_cpu_axpby(
channels,
alpha*sampleW[j],
diff_pn_.cpu_data() + (j*channels),
Dtype(0.0),
bout + (j*channels));
//}else{
// caffe_set(channels, Dtype(0), bout + (j*channels));
//}
} else if (i==1) { // p
//if(dist_binary_.cpu_data()[j]>Dtype(0)){
caffe_cpu_axpby(
channels,
alpha*sampleW[j],
diff_ap_.cpu_data() + (j*channels),
Dtype(0.0),
bout + (j*channels));
//}else{
// caffe_set(channels, Dtype(0), bout +(j*channels));
//}
}else if (i==2) { // n
//if(dist_binary_.cpu_data()[j]>Dtype(0)){
caffe_cpu_axpby(
channels,
alpha*sampleW[j],
diff_an_.cpu_data() + (j*channels),
Dtype(0.0),
bout + (j*channels));
//}else{
// caffe_set(channels, Dtype(0), bout + (j*channels));
//}
}
} // for num
} //if propagate_down[i]
} //for i
}
#ifdef CPU_ONLY
STUB_GPU(TripletLossLayer);
#endif
INSTANTIATE_CLASS(TripletLossLayer);
REGISTER_LAYER_CLASS(TripletLoss);
} // namespace caffe
在./src/caffe/layers/目录下新建Triplet_loss_layer.cu,实现GPU下的前传和反传,在GPU下实现前传和反传
/*
* Triplet_loss_layer.cu
*/
#include<algorithm>
#include<vector>
#include"caffe/layer.hpp"
#include"caffe/util/io.hpp"
#include"caffe/util/math_functions.hpp"
#include"caffe/vision_layers.hpp"
namespace caffe {
template <typename Dtype>
void TripletLossLayer<Dtype>::Forward_gpu(
const vector<Blob<Dtype>*>&bottom, const vector<Blob<Dtype>*>& top) {
const int count = bottom[0]->count();
caffe_gpu_sub(
count,
bottom[0]->gpu_data(), // a
bottom[1]->gpu_data(), // p
diff_ap_.mutable_gpu_data()); // a_i-p_i
caffe_gpu_sub(
count,
bottom[0]->gpu_data(), // a
bottom[2]->gpu_data(), // n
diff_an_.mutable_gpu_data()); //a_i-n_i
caffe_gpu_sub(
count,
bottom[1]->gpu_data(), // p
bottom[2]->gpu_data(), // n
diff_pn_.mutable_gpu_data()); // p_i-n_i
caffe_gpu_powx(
count,
diff_ap_.mutable_gpu_data(), // a_i-p_i
Dtype(2),
diff_sq_ap_.mutable_gpu_data()); // (a_i-p_i)^2
caffe_gpu_gemv(
CblasNoTrans,
bottom[0]->num(),
bottom[0]->channels(),
Dtype(1.0), //alpha
diff_sq_ap_.gpu_data(), // (a_i-p_i)^2 // A
summer_vec_.gpu_data(), // x
Dtype(0.0), //belta
dist_sq_ap_.mutable_gpu_data()); // \Sum (a_i-p_i)^2 //y
caffe_gpu_powx(
count,
diff_an_.mutable_gpu_data(), // a_i-n_i
Dtype(2),
diff_sq_an_.mutable_gpu_data()); // (a_i-n_i)^2
caffe_gpu_gemv(
CblasNoTrans,
bottom[0]->num(),
bottom[0]->channels(),
Dtype(1.0), //alpha
diff_sq_an_.gpu_data(), // (a_i-n_i)^2 // A
summer_vec_.gpu_data(), // x
Dtype(0.0), //belta
dist_sq_an_.mutable_gpu_data()); // \Sum (a_i-n_i)^2 //y
Dtype margin = this->layer_param_.triplet_loss_param().margin();
Dtype loss(0.0);
const Dtype* sampleW =bottom[3]->cpu_data();
for (int i = 0; i < bottom[0]->num();++i) {
loss += sampleW[i]*std::max(margin+dist_sq_ap_.cpu_data()[i]- dist_sq_an_.cpu_data()[i], Dtype(0.0));
}
loss = loss /static_cast<Dtype>(bottom[0]->num()) / Dtype(2);
top[0]->mutable_cpu_data()[0] = loss;
}
template <typename Dtype>
__global__ void CLLBackward(const int count, const int channels,
const Dtype margin, const Dtype alpha,const Dtype* sampleW,
const Dtype* diff, const Dtype*dist_sq_ap_, const Dtype* dist_sq_an_,
Dtype *bottom_diff) {
CUDA_KERNEL_LOOP(i, count) {
int n = i / channels; // the num index, to access dist_sq_ap_ anddist_sq_an_
Dtype mdist(0.0);
mdist = margin +dist_sq_ap_[n] -dist_sq_an_[n];
if (mdist > 0.0) {
bottom_diff[i] =alpha*sampleW[n]*diff[i];
} else {
bottom_diff[i] = 0;
}
}
}
template <typename Dtype>
void TripletLossLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>&propagate_down, const vector<Blob<Dtype>*>& bottom) {
Dtype margin = this->layer_param_.triplet_loss_param().margin();
const int count = bottom[0]->count();
const int channels =bottom[0]->channels();
for (int i = 0; i < 3; ++i) {
if (propagate_down[i]) {
const Dtype sign = (i < 2) ? -1 : 1;
const Dtype alpha = sign *top[0]->cpu_diff()[0] /
static_cast<Dtype>(bottom[0]->num());
if(i==0){
// NOLINT_NEXT_LINE(whitespace/operators)
CLLBackward<Dtype><<<CAFFE_GET_BLOCKS(count),CAFFE_CUDA_NUM_THREADS>>>(
count, channels, margin, alpha,
bottom[3]->gpu_data(),
diff_pn_.gpu_data(), // the cached eltwise difference between pand n
dist_sq_ap_.gpu_data(), // the cached square distance between a and p
dist_sq_an_.gpu_data(), // the cached square distance between a and n
bottom[i]->mutable_gpu_diff());
CUDA_POST_KERNEL_CHECK;
}else if(i==1){
// NOLINT_NEXT_LINE(whitespace/operators)
CLLBackward<Dtype><<<CAFFE_GET_BLOCKS(count),CAFFE_CUDA_NUM_THREADS>>>(
count, channels, margin, alpha,
bottom[3]->gpu_data(),
diff_ap_.gpu_data(), // the cached eltwise difference between aand p
dist_sq_ap_.gpu_data(), // the cached square distance between a and p
dist_sq_an_.gpu_data(), // the cached square distance between a and n
bottom[i]->mutable_gpu_diff());
CUDA_POST_KERNEL_CHECK;
}else if(i==2){
// NOLINT_NEXT_LINE(whitespace/operators)
CLLBackward<Dtype><<<CAFFE_GET_BLOCKS(count),CAFFE_CUDA_NUM_THREADS>>>(
count, channels, margin, alpha,
bottom[3]->gpu_data(),
diff_an_.gpu_data(), // the cached eltwise difference between aand n
dist_sq_ap_.gpu_data(), // the cached square distance between a and p
dist_sq_an_.gpu_data(), // the cached square distance between a and n
bottom[i]->mutable_gpu_diff());
CUDA_POST_KERNEL_CHECK;
}
}
}
}
INSTANTIATE_LAYER_GPU_FUNCS(TripletLossLayer);
} // namespace caffe
在./src/caffe/test/目录下增加test_Triplet_loss_layer.cpp
/*
* test_triplet_loss_layer.cpp
*/
#include<algorithm>
#include<cmath>
#include<cstdlib>
#include<cstring>
#include<vector>
#include"gtest/gtest.h"
#include"caffe/blob.hpp"
#include"caffe/common.hpp"
#include"caffe/filler.hpp"
#include"caffe/vision_layers.hpp"
#include "caffe/test/test_caffe_main.hpp"
#include"caffe/test/test_gradient_check_util.hpp"
namespace caffe {
template <typename TypeParam>
class TripletLossLayerTest: public MultiDeviceTest<TypeParam> {
typedef typename TypeParam::Dtype Dtype;
protected:
TripletLossLayerTest()
: blob_bottom_data_i_(new Blob<Dtype>(512, 2, 1, 1)),
blob_bottom_data_j_(new Blob<Dtype>(512, 2, 1, 1)),
blob_bottom_data_k_(new Blob<Dtype>(512, 2, 1, 1)),
blob_bottom_y_(new Blob<Dtype>(512, 1, 1, 1)),
blob_top_loss_(new Blob<Dtype>()){
// fill the values
FillerParameter filler_param;
filler_param.set_min(-1.0);
filler_param.set_max(1.0); // distances~=1.0 to test both sides ofmargin
UniformFiller<Dtype>filler(filler_param);
filler.Fill(this->blob_bottom_data_i_);
blob_bottom_vec_.push_back(blob_bottom_data_i_);
filler.Fill(this->blob_bottom_data_j_);
blob_bottom_vec_.push_back(blob_bottom_data_j_);
filler.Fill(this->blob_bottom_data_k_);
blob_bottom_vec_.push_back(blob_bottom_data_k_);
for (int i = 0; i <blob_bottom_y_->count(); ++i) {
blob_bottom_y_->mutable_cpu_data()[i] = caffe_rng_rand() % 2; // 0 or 1
}
blob_bottom_vec_.push_back(blob_bottom_y_);
blob_top_vec_.push_back(blob_top_loss_);
}
virtual ~TripletLossLayerTest() {
delete blob_bottom_data_i_;
delete blob_bottom_data_j_;
delete blob_bottom_data_k_;
delete blob_top_loss_;
}
Blob<Dtype>* const blob_bottom_data_i_;
Blob<Dtype>* const blob_bottom_data_j_;
Blob<Dtype>* const blob_bottom_data_k_;
Blob<Dtype>* const blob_bottom_y_;
Blob<Dtype>* const blob_top_loss_;
vector<Blob<Dtype>*>blob_bottom_vec_;
vector<Blob<Dtype>*>blob_top_vec_;
};
TYPED_TEST_CASE(TripletLossLayerTest,TestDtypesAndDevices);
TYPED_TEST(TripletLossLayerTest,TestForward) {
typedef typename TypeParam::Dtype Dtype;
LayerParameter layer_param;
TripletLossLayer<Dtype>layer(layer_param);
layer.SetUp(this->blob_bottom_vec_,this->blob_top_vec_);
layer.Forward(this->blob_bottom_vec_,this->blob_top_vec_);
// manually compute to compare
const Dtype margin = layer_param.triplet_loss_param().margin();
const int num =this->blob_bottom_data_i_->num();
const int channels =this->blob_bottom_data_i_->channels();
Dtype loss(0);
for (int i = 0; i < num; ++i) {
Dtype dist_sq_ij(0);
Dtype dist_sq_ik(0);
for (int j = 0; j < channels; ++j) {
Dtype diff_ij =this->blob_bottom_data_i_->cpu_data()[i*channels+j] -
this->blob_bottom_data_j_->cpu_data()[i*channels+j];
dist_sq_ij += diff_ij*diff_ij;
Dtype diff_ik =this->blob_bottom_data_i_->cpu_data()[i*channels+j] -
this->blob_bottom_data_k_->cpu_data()[i*channels+j];
dist_sq_ik += diff_ik*diff_ik;
}
loss += std::max(Dtype(0.0),margin+dist_sq_ij-dist_sq_ik);
}
loss /= static_cast<Dtype>(num) *Dtype(2);
EXPECT_NEAR(this->blob_top_loss_->cpu_data()[0], loss, 1e-6);
}
TYPED_TEST(TripletLossLayerTest,TestGradient) {
typedef typename TypeParam::Dtype Dtype;
LayerParameter layer_param;
TripletLossLayer<Dtype>layer(layer_param);
layer.SetUp(this->blob_bottom_vec_,this->blob_top_vec_);
GradientChecker<Dtype> checker(1e-2,1e-2, 1701);
// check the gradient for the first twobottom layers
checker.CheckGradientExhaustive(&layer,this->blob_bottom_vec_,
this->blob_top_vec_, 0);
checker.CheckGradientExhaustive(&layer,this->blob_bottom_vec_,
this->blob_top_vec_, 1);
}
} // namespace caffe
这个写法主要适合老版本的Caffe,新版本的Caffe每个layer实现都有一个.cpp与.hpp相对应,不过步骤都是一致的。
看我写的辛苦求打赏啊!!!有学术讨论和指点请加微信manutdzou,注明