在 caffe 中添加 FaceNet 中 Triplet Loss Layer
目录
- 目录
-
- 一Triplet Loss 定义
- 二创建Triplet Loss Layer
- 在caffeproto添加对应Message
- 创建triplet_loss_layerhpp
- 创建triplet_loss_layercpp
- 创建triplet_loss_layercu
- 创建test_triplet_loss_layercpp
-
一、Triplet Loss 定义
论文：FaceNet: A Unified Embedding for Face Recognition and Clustering 中有关于Triplet loss 详细定义及意义.
L=∑iN[||f(xai)−f(xpi)||22−||f(xai)−f(xni)||22+α]+
back propogation 需要用到的导数：
∂(L)∂f(xai)=2⋅(f(xni)−f(xpi))
∂(L)∂f(xpi)=−2⋅(f(xai)−f(xpi))
∂(L)∂f(xni)=2⋅(f(xai)−f(xni))
二、创建Triplet Loss Layer
1. 在caffe.proto添加对应Message
- 首先在LayerParameter中添加：
optional TripletLossParameter triplet_loss_param = 136;
//其中136是我自己装的caffe的当前参数 ID，记得每添加新的参数时，更新message首段注释中的最新ID，以方便下次添加- 由于我们需要设定的参数只有 margin: α ， 所以新建Message TripletLossParameter ：
message TripletLossParameter {
optional float margin = 1 [default = 1.0];
} 2.创建triplet_loss_layer.hpp
#ifndef CAFFE_TRIPLET_LOSS_LAYER_HPP_
#define CAFFE_TRIPLET_LOSS_LAYER_HPP_
#include {
public:
explicit TripletLossLayer (const LayerParameter& param)
: LossLayer(param), diff_np_(), diff_ap_(), diff_an_() {}
virtual void LayerSetUp(const vector diff_np_; // cached for backward pass: f(x_i^n) - f(x_i^p)
Blob diff_ap_; // cached for backward pass: f(x_i^a) - f(x_i^p)
Blob diff_an_; // cached for backward pass: f(x_i^a) - f(x_i^n)
Blob dist_ap_sq_; // cached for backward pass :||f(x_i^a)-f(x_i^p)||^2
Blob dist_an_sq_; // cached for backward pass :||f(x_i^a)-f(x_i^n)||^2
Blob diff_ap_sq_; // tmp storage for gpu forward pass: f(x_i^a)-f(x_i^p).^2
Blob diff_an_sq_; // tmp storage for gpu forward pass: f(x_i^a)-f(x_i^n).^2
Blob summer_vec_; // tmp storage for gpu forward pass: [ 1, 1, 1,...,1 ]
};
} // namespace caffe
#endif // CAFFE_TRIPLET_LOSS_LAYER_HPP_
3.创建triplet_loss_layer.cpp
#include ::LayerSetUp(
const vector::LayerSetUp(bottom, top);
// bottom[0] : f(x_i^a); bottom[1] : f(x_i^p); bottom[2] : f(x_i^n)
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);
diff_np_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
diff_ap_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
diff_an_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
dist_ap_sq_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
diff_an_sq_.Reshape(bottom[0]->num(), bottom[0]->channels(), 1, 1);
dist_ap_sq_.Reshape(bottom[0]->num(), 1, 1, 1);
dist_an_sq_.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);
}
template <typename Dtype>
void TripletLossLayer::Forward_cpu(
const vector::Backward_cpu(const vector 4.创建triplet_loss_layer.cu
#include ::Forward_gpu(
const vector::Backward_gpu(const vector<<>>(
count, channels, margin, alpha,
diff_np_.gpu_data(), // f(x_i^n)-f(x_i^p)
dist_ap_sq_.gpu_data(), // ||f(x_i^a)-f(x_i^p)||^2
dist_an_sq_.gpu_data(), // ||f(x_i^a)-f(x_i^n)||^2
bottom[i]->mutable_gpu_diff());
CUDA_POST_KERNEL_CHECK;
}else if(i==1){ // \frac{\partial(L)}{\partial f(x_i^p)}
// NOLINT_NEXT_LINE(whitespace/operators)
CLLBackward<<>>(
count, channels, margin, alpha,
diff_ap_.gpu_data(), // f(x_i^a)-f(x_i^p)
dist_ap_sq_.gpu_data(), // ||f(x_i^a)-f(x_i^p)||^2
dist_an_sq_.gpu_data(), // ||f(x_i^a)-f(x_i^n)||^2
bottom[i]->mutable_gpu_diff());
CUDA_POST_KERNEL_CHECK;
}else if(i==2){ // \frac{\partial(L)}{\partial f(x_i^n)}
// NOLINT_NEXT_LINE(whitespace/operators)
CLLBackward<<>>(
count, channels, margin, alpha,
diff_an_.gpu_data(), // f(x_i^a)-f(x_i^n)
dist_ap_sq_.gpu_data(), // ||f(x_i^a)-f(x_i^p)||^2
dist_an_sq_.gpu_data(), // ||f(x_i^a)-f(x_i^n)||^2
bottom[i]->mutable_gpu_diff());
CUDA_POST_KERNEL_CHECK;
}
}
}
}
INSTANTIATE_LAYER_GPU_FUNCS(TripletLossLayer);
} // namespace caffe
5.创建test_triplet_loss_layer.cpp
#include {
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_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_);
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_; // f(x_i^a)
Blob* const blob_bottom_data_j_; // f(x_i^p)
Blob* const blob_bottom_data_k_; // f(x_i^n)
Blob* const blob_top_loss_; // loss
vector 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_ap(0);
Dtype dist_sq_an(0);
for (int j = 0; j < channels; ++j) {
Dtype diff_ap = this->blob_bottom_data_i_->cpu_data()[i*channels+j] -
this->blob_bottom_data_j_->cpu_data()[i*channels+j];
dist_sq_ap += diff_ap * diff_ap;
Dtype diff_an = this->blob_bottom_data_i_->cpu_data()[i*channels+j] -
this->blob_bottom_data_k_->cpu_data()[i*channels+j];
dist_sq_an += diff_an * diff_an;
}
loss += std::max(Dtype(0.0), margin + dist_sq_ap - dist_sq_an);
}
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 three bottom layers
checker.CheckGradientExhaustive(&layer, this->blob_bottom_vec_,
this->blob_top_vec_, 0); // check gradient for f(x_i^a)
checker.CheckGradientExhaustive(&layer, this->blob_bottom_vec_,
this->blob_top_vec_, 1); // check gradient for f(x_i^p)
checker.CheckGradientExhaustive(&layer, this->blob_bottom_vec_,
this->blob_top_vec_, 2); // check gradient for f(x_i^n)
}
} // namespace caffe