Coverage for bmm_multitask_learning/coalescent/inverse_wishart.py: 0%

1 

2import numpy as np 

3 

4 

5class InverseWishart: 

6 """ 

7 class to handle methods of InverseWishart distribution 

8 """ 

9 def __init__(self, d, degrees_of_freedom, init_cov=None, corr_mat=False): 

10 """ 

11 :param d: dimension of the covariance matrix 

12 :param degrees_of_freedom: degrees of freedom of the 

13 inverse Wishart distribution 

14 :param init_cov: initial covariance matrix 

15 :param corr_mat: if True, the covariance matrix is a correlation 

16 matrix (i.e., it has ones on the diagonal) 

17 """ 

18 self.d = d 

19 if init_cov is None: 

20 init_cov = np.eye(d) 

21 else: 

22 assert init_cov.shape == (d, d) 

23 

24 self.cov = init_cov 

25 self.degrees_of_freedom = degrees_of_freedom 

26 self.corr_mat = corr_mat 

27 

28 @staticmethod 

29 def cov2corr(cov): 

30 """ 

31 Conver covariance matrox to correlation matrix 

32 """ 

33 if len(cov.shape) == 0: 

34 cov = cov[None] 

35 

36 std_devs = 1/np.sqrt(np.diag(cov)) 

37 D = np.diag(std_devs) 

38 

39 d = len(cov.shape) 

40 

41 if d == 1: 

42 D = D[None] 

43 cov = cov[None] 

44 R = D @ cov @ D 

45 return R 

46 

47 def update_posterior(self, samples): 

48 """ 

49 :param samples: shape=(n x d) 

50 """ 

51 assert len(samples.shape) == 2 and samples.shape[1] == self.d 

52 K = samples.shape[0] 

53 self.degrees_of_freedom += K 

54 

55 S = samples.T @ samples # d x d 

56 self.cov += S 

57 return 

58 

59 def get_most_prob(self): 

60 """ 

61 returns the mode of the inverse Wishart distribution 

62 """ 

63 cov = self.cov 

64 cov = cov/(self.degrees_of_freedom + self.d + 1) 

65 

66 if self.corr_mat: 

67 cov = InverseWishart.cov2corr(cov) 

68 return cov 

69