Source Code for Module model_descriptions.full_gradient

  1  """ 
  2  This module contains non incremental (batch or full gradient) functions. 
  3  """ 
  4   
  5  import victor_utils, math 
  6   
  7  ######################## 
  8  ## Full Gradient Code 
  9  #  
 10  ######################## 
 11  ################# 
 12  # Ell2 gradient 
 13  ############### 
 14 -def ell2_full_gradient(w, cur_grad, x, y): 
 15     """ 
 16     Calculates ell2 gradient. 
 17      
 18     @type w: vector 
 19     @param w: model vector 
 20     @type cur_grad: vector 
 21     @param cur_grad: current gradient 
 22     @type x: vector 
 23     @param x: feature vector of the example 
 24     @type y: number 
 25     @param y: label of the example 
 26     """ 
 27     wx  = victor_utils.dot(w, x) 
 28     err = (wx - y) 
 29     victor_utils.scale_and_add_i(cur_grad, x, err) 
 30      
 31  # ell2 gradient step 
 32 -def ell2_gradient_step(w, grad, stepsize, mu, B): 
 33     """ 
 34     Computes w as a step 
 35     @type w: vector 
 36     @param w: model vector 
 37     @type grad: vector 
 38     @param grad: gradient 
 39     @type stepsize: double 
 40     @param stepsize: step size 
 41     @type mu: double 
 42     @param mu: mu 
 43     @type B: double 
 44     @param B: B 
 45     """ 
 46     # compute w as a step 
 47     victor_utils.scale_and_add_i(w, grad, -stepsize) 
 48     victor_utils.l2_project(w, B) 
 49   
 50  ################ 
 51  # SVM + L2 Prox 
 52  ############## 
 53 -def svm_full_gradient(w, cur_grad, x,y): 
 54     """ 
 55     Calculates svm gradient. 
 56      
 57     @type w: vector 
 58     @param w: model vector 
 59     @type cur_grad: vector 
 60     @param cur_grad: current gradient 
 61     @type x: vector 
 62     @param x: feature vector of the example 
 63     @type y: number 
 64     @param y: label of the example 
 65     """ 
 66     wx  = victor_utils.dot(w, x) 
 67     if(y*wx < 1): 
 68        victor_utils.scale_and_add_i(cur_grad, x, y) 
 69   
 70 -def ell2_prox_term(w, cur_grad, stepsize, mu, B): 
 71     """ 
 72     Computes w as a step 
 73     @type w: vector 
 74     @param w: model vector 
 75     @type cur_grad: vector 
 76     @param cur_grad: gradient 
 77     @type stepsize: double 
 78     @param stepsize: step size 
 79     @type mu: double 
 80     @param mu: mu 
 81     @type B: double 
 82     @param B: B 
 83     """ 
 84     victor_utils.scale_and_add_i(w, cur_grad, -stepsize) 
 85     victor_utils.scale(w, 1.0/(1.0 + stepsize * mu)) 
 86   
 87   
 88  ################ 
 89  # Logit helpers 
 90  ############## 
 91 -def logit_l1_sparse(w, cur_grad, x_indexes, x_vectors, y): 
 92     """ 
 93     Calculates logit l1 sparse 
 94     @type w: vector 
 95     @param w: model vector 
 96     @type cur_grad: vector 
 97     @param cur_grad: current gradient 
 98     @type x_indexes: vector 
 99     @param x_indexes: indexes vector of the example 
100     @type x_vectors: vector 
101     @param x_vectors: values of the feature vector of the example 
102     @type y: number 
103     @param y: label of the example 
104     """ 
105     wx      = victor_utils.dot_dss(w,x_indexes,x_vectors) 
106     sig     = victor_utils.sigma(-wx*y) 
107     victor_utils.scale_and_add_dss(cur_grad, x_indexes, x_vectors, -y*sig) 
108   
109 -def l1_shrink_prox(w, cur_grad, stepsize, mu, B): 
110     """ 
111     Computes l1 shrink prox 
112     @type w: vector 
113     @param w: model vector 
114     @type cur_grad: vector 
115     @param cur_grad: gradient 
116     @type stepsize: double 
117     @param stepsize: step size 
118     @type mu: double 
119     @param mu: mu 
120     @type B: double 
121     @param B: B 
122     """ 
123     victor_utils.scale_and_add_di(w,cur_grad, -stepsize) 
124     victor_utils.l1_shrink_mask(w, mu*step) 
125   
126 -def logit_loss(w, x_indexes, x_values, y):    
127     """ 
128     Returns logit loss 
129     @type w: vector 
130     @param w: model vector 
131     @type x_indexes: vector 
132     @param x_indexes: indexes vector of the example 
133     @type x_values: vector 
134     @param x_values: values of the feature vector of the example 
135     @rtype: double 
136     @return: logit loss of the given model and the example 
137     """ 
138     wx      = victor_utils.dot_dss(w,x_indexes,x_values) 
139     return  math.log(1 + math.exp(-y*wx)) 
140