关于triplet loss的原理。目标函数和梯度推导在上一篇博客中已经讲过了。详细见：triplet loss原理以及梯度推导。这篇博文主要是讲caffe下实现triplet loss。编程菜鸟。假设有写的不优化的地方，欢迎指出。
1.怎样在caffe中添加新的layer
新版的caffe中添加新的layer。变得轻松多了。概括说来。分四步：
1）在./src/caffe/proto/caffe.proto 中添加 相应layer的paramter message。
2）在./include/caffe/***layers.hpp中添加该layer的类的声明。***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/文件夹下新建.cpp和.cu文件，进行类实现。
4）在./src/caffe/gtest/中添加layer的測试代码。对所写的layer前传和反传进行測试，測试还包含速度。
最后一步非常多人省了，或者没意识到。可是为保证代码正确，建议还是严格进行測试，磨刀不误砍柴功。
2.caffe中实现triplet loss layer
1.caffe.proto中添加triplet loss layer的定义
首先在message LayerParameter中追加 optional TripletLossParameter triplet_loss_param = 138; 当中138是我眼下LayerParameter message中现有元素的个数，详细是多少。能够看LayerParameter message上面凝视中的：
//LayerParameter next available layer-specific ID: 134 (last added: reshape_param)message TripletLossParameter {
// margin for dissimilar pair
optional float margin = 1 [default = 1.0];
}2.在./include/caffe/loss_layers.hpp中添加triplet loss layer的类的声明
详细解释见凝视。基本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反复计算。
/**
* @brief Computes the triplet loss
*/
template
class TripletLossLayer : public LossLayer {
public:
explicit TripletLossLayer(const LayerParameter& param)
: LossLayer(param){}
virtual void LayerSetUp(const vector*>& bottom,
const vector*>& 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*>& bottom,
const vector*>& top);
virtual void Forward_gpu(const vector*>& bottom,
const vector*>& top);
virtual void Backward_cpu(const vector*>& top,
const vector& propagate_down, const vector*>& bottom);
virtual void Backward_gpu(const vector*>& top,
const vector& propagate_down, const vector*>& bottom);
Blob diff_ap_; // cached for backward pass
Blob diff_an_; // cached for backward pass
Blob diff_pn_; // cached for backward pass
Blob diff_sq_ap_; // cached for backward pass
Blob diff_sq_an_; // tmp storage for gpu forward pass
Blob dist_sq_ap_; // cached for backward pass
Blob dist_sq_an_; // cached for backward pass
Blob summer_vec_; // tmp storage for gpu forward pass
Blob dist_binary_; // tmp storage for gpu forward pass
}; 3. 在./src/caffe/layers/文件夹下新建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，然后依据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
*
* Created on: Jun 2, 2015
* Author: tangwei
*/
#include
#include
#include "caffe/layer.hpp"
#include "caffe/loss_layers.hpp"
#include "caffe/util/io.hpp"
#include "caffe/util/math_functions.hpp"
namespace caffe {
template
void TripletLossLayer::LayerSetUp(
const vector*>& bottom, const vector*>& top) {
LossLayer::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
void TripletLossLayer::Forward_cpu(
const vector*>& bottom,
const vector*>& 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);
//prepare for 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(bottom[0]->num()) / Dtype(2);
top[0]->mutable_cpu_data()[0] = loss;
}
template
void TripletLossLayer::Backward_cpu(const vector*>& top,
const vector& propagate_down, const vector*>& bottom) {
//Dtype margin = this->layer_param_.contrastive_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(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 4.在./src/caffe/layers/文件夹下新建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
*
* Created on: Jun 2, 2015
* Author: tangwei
*/
#include
#include
#include "caffe/layer.hpp"
#include "caffe/util/io.hpp"
#include "caffe/util/math_functions.hpp"
#include "caffe/vision_layers.hpp"
namespace caffe {
template
void TripletLossLayer::Forward_gpu(
const vector*>& bottom, const vector*>& 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(bottom[0]->num()) / Dtype(2);
top[0]->mutable_cpu_data()[0] = loss;
}
template
__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_ and dist_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
void TripletLossLayer::Backward_gpu(const vector*>& top,
const vector& propagate_down, const vector*>& 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(bottom[0]->num());
if(i==0){
// NOLINT_NEXT_LINE(whitespace/operators)
CLLBackward<<>>(
count, channels, margin, alpha,
bottom[3]->gpu_data(),
diff_pn_.gpu_data(), // the cached eltwise difference between p and 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<<>>(
count, channels, margin, alpha,
bottom[3]->gpu_data(),
diff_ap_.gpu_data(), // the cached eltwise difference between a and 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<<>>(
count, channels, margin, alpha,
bottom[3]->gpu_data(),
diff_an_.gpu_data(), // the cached eltwise difference between a and 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 5. 在./src/caffe/test/文件夹下添加test_triplet_loss_layer.cpp
/*
* test_triplet_loss_layer.cpp
*
* Created on: Jun 3, 2015
* Author: tangwei
*/
#include
#include
#include
#include
#include
#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
class TripletLossLayerTest : public MultiDeviceTest {
typedef typename TypeParam::Dtype Dtype;
protected:
TripletLossLayerTest()
: blob_bottom_data_i_(new Blob(512, 2, 1, 1)),
blob_bottom_data_j_(new Blob(512, 2, 1, 1)),
blob_bottom_data_k_(new Blob(512, 2, 1, 1)),
blob_bottom_y_(new Blob(512, 1, 1, 1)),
blob_top_loss_(new Blob()) {
// 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 of margin
UniformFiller 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* const blob_bottom_data_i_;
Blob* const blob_bottom_data_j_;
Blob* const blob_bottom_data_k_;
Blob* const blob_bottom_y_;
Blob* const blob_top_loss_;
vector*> blob_bottom_vec_;
vector*> blob_top_vec_;
};
TYPED_TEST_CASE(TripletLossLayerTest, TestDtypesAndDevices);
TYPED_TEST(TripletLossLayerTest, TestForward) {
typedef typename TypeParam::Dtype Dtype;
LayerParameter layer_param;
TripletLossLayer 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(); const Dtype *sampleW = this->blob_bottom_y_->cpu_data(); 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 += sampleW[i]*std::max(Dtype(0.0), margin+dist_sq_ij-dist_sq_ik);
}
loss /= static_cast(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 layer(layer_param);
layer.SetUp(this->blob_bottom_vec_, this->blob_top_vec_);
GradientChecker checker(1e-2, 1e-2, 1701);
// check the gradient for the first two bottom 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
3.编译測试
又一次 make all 假设出错，检查代码语法错误。
make test
make runtest 假设成功，全是绿色的OK 否则会给出红色提示。就得看看是不是实现逻辑上出错了。