Problems/synthetic_cosh.py

   1
   2 import numpy as np
   3
   4
   5
   6 class synthetic(): ### minimize log(cosh(cx-e) s.t. ax-b<=0)
   7     def __init__(self, seed, num_nodes, dim):
   8         np.random.seed(seed)
   9         self.n = num_nodes
  10         self.d = dim
  11         self.A = np.random.randn(num_nodes, dim)
  12         self.e = np.random.randn(num_nodes)
  13         self.C = np.random.randn(num_nodes, dim)
  14         self.b = np.matmul(self.A, np.random.randn(dim)) + 1*(np.random.rand(num_nodes))
  15
  16         self.N = 1
  17         self.X_train = None
  18         self.Y_train = None
  19         self.p = self.A.shape[1]
  20         self.dim = self.A.shape[1]
  21
  22     def F_val(self, theta):
  23         return np.sum(np.log10(np.cosh(np.matmul(self.C,theta)-self.e)))
  24
  25     def localgrad(self, theta, idx):
  26         grad2 = (1 / np.cosh(np.inner(self.C[idx],theta[idx])-self.e[idx])) *\
  27                 np.sinh(np.inner(self.C[idx], theta[idx])-self.e[idx]) * self.C[idx]
  28         return grad2
  29
  30     def gd_grad(self, theta, dim):
  31         grad = np.zeros(dim)
  32         for i in range(self.n):
  33             grad += (1 / np.cosh(np.inner(self.C[i],theta)-self.e[i])) *\
  34                 np.sinh(np.inner(self.C[i], theta)-self.e[i]) * self.C[i]
  35         return grad
  36
  37     def networkgrad(self, theta):
  38         grad = np.zeros((self.n, self.p))
  39         for i in range(self.n):
  40             grad[i] = self.localgrad(theta, i)
  41         return grad
  42
  43     def grad(self, theta):
  44         pass
  45
  46     def local_projection(self, idx, theta):
  47         tmp = np.inner(self.A[idx],theta[idx]) - self.b[idx]
  48         if tmp < 0:
  49             return theta[idx]
  50         else:
  51             return theta[idx] - tmp*( (self.A[idx]) / (np.linalg.norm(self.A[idx]))**2 )
  52
  53     def network_projection(self, theta):
  54         proj = np.zeros((self.n, self.p))
  55         for i in range(self.n):
  56             proj[i] = self.local_projection(i, theta)
  57         return proj
  58
  59
  60
  61
  62