matlab---triplet loss

**摘要:**triplet loss 可以提高特征匹配的性能，可用物体识别，人脸识别，检索等方面，本文用matlab实现triplet loss。

triplet loss 就是学习一个函数隐射$f(x)$, 从特征$x$ 映射到 $\mathbb{R}^{D}$, 有如下关系：$y=f(x)$. 在一个特征空间$\mathbb{R}^{D}$中，我们通过欧式距离度量两个特征向量的距离。 triplet loss 的目的在于使同一个类别在$\mathbb{R}^{D}$空间中靠近，不同类别在$\mathbb{R}^{D}$空间中远离，那么我们就可以抽象为如下优化函数：

$$\min_{w}\frac{1}{m}\sum_{i}^{m}||f(x_i^{a}-f(x_i^p))||_2^2-||f(x_i^{a}-f(x_i^n))||_2^2+\alpha$$

其中$f(x)=wx$, $x_i^{a}$是锚点，$x_i^{p}$是正样本点，它和$x_i^{a}$属于同一类别，$x_i^{n}$是负样本点，它和$x_i^{a}$不属于同一类别。
这样我们可以通过无约束优化来约束上面函数。

function demo_tripletloss
clear all
clc
data{1} = rand(600,300);
data{2} = rand(600,300);
data{3} = rand(600,300);
inputSize = 600;
hiddenSize = 400;
theta = initializeParameters(hiddenSize, inputSize);

addpath minFunc/
options.Method = 'cg'; % Here, we use L-BFGS to optimize our cost
                          % function. Generally, for minFunc to work, you
                          % need a function pointer with two outputs: the
                          % function value and the gradient. In our problem,
                          % sparseAutoencoderCost.m satisfies this.
options.maxIter = 400;      % Maximum number of iterations of L-BFGS to run 
options.display = 'on';

[opttheta, cost] = minFunc( @(p) tripletCost(p, inputSize, hiddenSize, data),theta, options);

tripletCost.m

function [cost,grad] = tripletCost(theta, inputSize, hiddenSize, data)
%================================================
%何凌霄 中科院自动化所
%创建时间：2016年5月25日
%================================================


% data: trainning sample for triplet loss
% W: transformer matrix
% cost: cost function
% grad: gradient direction
W = reshape(theta, hiddenSize, inputSize);
% Cost and gradient variables (your code needs to compute these values). 
% Here, We initialize them to zeros. 
cost = 0;
Wgrad = zeros(size(W)); 

% the gradient descent update to W1 Would be W1 := W1 - alpha * W1grad, and similarly for W2, b1, b2. 
% 

[n m] = size(data{1});%m: the number of samples，m:the dim of feature
bias = 0.2;
%forWard
% calc cost
for i = 1:m
   cost = cost + ((W*data{1}(:,i) - W*data{2}(:,i))'*(W *data{1}(:,i) - W*data{2}(:,i)) - (W*data{1}(:,i) - W*data{3}(:,i))'*(W *data{1}(:,i) - W*data{3}(:,i)) +bias)/m;
end

%calc gradient 

%计算W1grad 
Wgrad = (2*W*((data{1} - data{2})*(data{1} - data{2})' - (data{1} - data{3})*(data{1} - data{3})'))/m;

%-------------------------------------------------------------------
% After computing the cost and gradient, We Will convert the gradients back
% to a vector format (suitable for minFunc).  Specifically, We Will unroll
% your gradient matrices into a vector.

grad = Wgrad(:);

end

initializeParameters.m

function theta = initializeParameters(hiddenSize, inputSize)

% Initialize parameters randomly based on layer sizes.
r  = sqrt(6) / sqrt(hiddenSize+inputSize+1);   % we'll choose weights uniformly from the interval [-r, r]
W1 = rand(hiddenSize, inputSize) * 2 * r - r;

theta = W1(:);

end

另外需要一个优化包见资源.