mexTrainDL - SPAms

稀疏分解 $Y=D\alpha$ 分为两个步骤：

• sparse coding stage: means given D, to get sparse coefficients. (根据对稀疏系数向量的稀疏范数要求，分为算法OMP for $L_0$, 算法LARS for $L_1, L_2$)
• atoms updating stage: means given $\alpha$, to update atoms in dictionary $D$.

1) if param.mode=0

$$\min_{D \in C} \frac{1}{n} \sum_{i=1}^n \frac{1}{2} ||x_i -D\alpha_i||_2^2, \quad s.t. ||\alpha_i||_1 \leq \lambda$$
2) if param.mode=1
$$\min_{D \in C} \frac{1}{n} \sum_{i=1}^{n} ||\alpha_i||_1 \quad s.t.||x_i-D\alpha_i||_2^2\leq \lambda$$
3) if param.mode=2 (default)
$$\min_{D \in C} \frac{1}{n} \sum_{i=1}^n \{ \frac{1}{2} ||x_i-D\alpha_i||_2^2 + \lambda||\alpha_i||_1 + \lambda_2||\alpha_i||_2^2 \}$$
4) if param.mode=3
$$\min_{D \in C} \frac{1}{n} \sum_{i=1}^n \frac{1}{2}||x_i-D\alpha_i||_2^2 \quad s.t. ||\alpha_i||_0 \leq \lambda$$
5) if param.mode=4
$$\min_{D \in C} \frac{1}{n} \sum_{i=1}^n ||\alpha_i||_0 \quad s.t. ||x_i-D\alpha_i||_2^2 \leq \lambda$$
6) if param.mode=5
$$\min_{D \in C} \frac{1}{n} \sum_{i=1}^n \{ \frac{1}{2}||x_i-D\alpha_i||_2^2 +\lambda||\alpha_i||_0\}$$

When param.mode=1, 2, 可以考虑 mexTrainDL_Memory()
When param.mode=3, 4, 5, sparse coding stage is solved by OMP.

对D的约束：$\mathcal{C}$

$\mathcal{C}$ is a convex set verifing:

1) if param.modeD=0 (default)

$$\mathcal{C}=\{D\in \Re^{m\times p}, \quad s.t.\quad \forall j,\quad ||d_j||_2^2\leq1 \}$$
2) if param.modeD=1
$$\mathcal{C}=\{D \in \Re^{m\times p},\quad s.t.\quad \forall j, \quad ||d_j||_2^2+\gamma_1||d_j||_1\leq1\}$$
3) if param.modeD=2
$$\mathcal{C}=\{D \in \Re^{m\times p},\quad s.t.\quad \forall j, \quad ||d_j||_2^2+\gamma_1||d_j||_1+\gamma_2 FL(d_j)\leq1\}$$
4) if param.modeD=3
$$\mathcal{C}=\{D \in \Re^{m\times p},\quad s.t.\quad \forall j,\quad (1-\gamma_1)||d_j||_2^2+\gamma_1||d_j||_1\leq1\}$$

参数说明

parameter	说明
param.D
param.K
param.lamda
param.lamda2
param.iter
param.mode
param.posAlpha	系数矩阵 $\alpha$ 非负性(not compatible with param.mode =3,4))
param.modeD
param.posD	字典 D <script type="math/tex" id="MathJax-Element-61">D</script> 矩阵的非负性(not compatible with param.modeD=2)
param.gamma1
pamma.gamma2
param.batchsize	default 512
param.iter_updateD
param.modeParam	0, 1, 2
param.rho
param.t0
param.clean
param.verbose
param.numThreads	by default -1