def RC_drop_var(x, S, dist, df):
dist = st.wishart(df=df, scale=S)
s_logpdf = dist.logpdf(x * df)
ex_logpdf = []
for j in range(len(x[0, :])):
ex_c = np.delete(x, j, axis=0)
S_c = np.delete(S, j, axis=0)
ex_c1 = np.delete(ex_c, j, axis=1)
S_c1 = np.delete(S_c, j, axis=1)
dist_c = st.wishart(df=df, scale=S_c1)
p_ex = dist_c.logpdf(ex_c1 * df)
diff_logpdf = s_logpdf - p_ex
ex_logpdf.append(diff_logpdf)
return (np.asarray(ex_logpdf))
def RC2_en_drop(x, S, df):
dist = st.wishart(df=df, scale=S)
s_logpdf = dist.logpdf(x * df)
ex_logpdf = np.empty(x.shape)
size = range(len(x))
for i, j in it.combinations_with_replacement(size, 2):
x1 = np.delete(x, [i, j], axis=0)
x2 = np.delete(x1, [i, j], axis=1)
S1 = np.delete(S, [i, j], axis=0)
S2 = np.delete(S1, [i, j], axis=1)
dist_c = st.wishart(df=df, scale=S2)
p_ex = dist_c.logpdf(x2 * df)
diff_logpdf = s_logpdf - p_ex
ex_logpdf[j, i] = diff_logpdf
ex_logpdf[j, i] = ex_logpdf[i, j]
return (ex_logpdf)
def Gibbs_sample_slow_scipy_version(param_dict, non_inf=False):
'''Draw a sample from the Normal-Wishart model using the buit-in scipy distributions( significantly slower).
Based on Gelman et al.(2013, 3 ed., chapter 4).
Parameters
---------
param_dict: python-dict
dictionary with sufficient statistics and parameters from a multivariate data-set,
obtained through the functions 'make_param_dict' and 'update_param_dict'.
non_inf: Boolean.
This models tends weight heavily the distance from the prior to the empirical mean.
If a non-informative prior is used, it's advisable to use this option. The parametrization
will correspond to the multivariate Jeffreys prior density.
Output
--------
Returns a d-dimensional draw from the Normal-Wishart model. For multiple samples,
use the function 'Gibbs_sampler' with one of the options:
('slow'): non_inf = False
('nonInfSlow'): non_inf = True
'''
if non_inf:
Prec_m = sts.wishart(df = param_dict['n']-1., scale=param_dict['invS_m']).rvs()
Sigma_m = np.linalg.inv(Prec_m)
mu = sts.multivariate_normal(mean= param_dict['E_mu'], cov=(1./param_dict['n'])*Sigma_m).rvs()
return sts.multivariate_normal(mean=mu, cov = Sigma_m).rvs()
else:
Prec_m = sts.wishart(df = param_dict['up_v_0'], scale=param_dict['up_Prec_0']).rvs()
Sigma_m = np.linalg.inv(Prec_m)
mu = sts.multivariate_normal(mean= param_dict['up_mu_0'], cov=(1./param_dict['up_k_0'])*Sigma_m).rvs()
return sts.multivariate_normal(mean=mu, cov = Sigma_m).rvs()
def RC2_drop_var(x, S, dist, df):
dist = st.wishart(df=df, scale=S)
s_logpdf = dist.logpdf(x * df)
ex_logpdf = np.zeros((len(x[0, :]), len(x[:, 0])))
diag_indices = np.diag_indices(len(x))
diag = []
for j in range(len(x[0, :])):
for i in range(len(x[:, 0])):
if j == i: ## Diagonal element list creation, avoids a good number of
ex_c = np.delete(x, j, axis=0) ## ops and mat copies conflict
S_c = np.delete(S, j, axis=0)
ex_c1 = np.delete(ex_c, j, axis=1)
S_c1 = np.delete(S_c, j, axis=1)
dist_c = st.wishart(df=df, scale=S_c1)
p_ex = dist_c.logpdf(ex_c1 * df)
diff_logpdf = s_logpdf - p_ex
diag.append(diff_logpdf)
else:
ex_c = np.delete(x, j, axis=0)
S_c = np.delete(S, j, axis=0)
ex_c1 = np.delete(ex_c, j, axis=1)
S_c1 = np.delete(S_c, j, axis=1)
############
ex_c2 = np.delete(ex_c1, i - 1, axis=0)
S_c2 = np.delete(S_c1, i - 1, axis=0)
ex_c3 = np.delete(ex_c2, i - 1, axis=1)
S_c3 = np.delete(S_c2, i - 1, axis=1)
dist_c = st.wishart(df=df, scale=S_c3)
p_ex = dist_c.logpdf(ex_c3 * df)
diff_logpdf = s_logpdf - p_ex
ex_logpdf[j, i] = diff_logpdf
ex_logpdf[diag_indices] = diag
return (ex_logpdf)
def get_sims(self,errdist_perms,recalc=False):
""" generate simulated wishart samples """
A=self.A
if not( 'covs_sim' in self.__dict__) or recalc==True:
covs=A.get_covs()
dof=A.dof
# if sparse cov matrix
if scipy.sparse.issparse(A.get_covs()):
corrs=A.get_corrs()
cvs=covs.diagonal()[1:]
# generate random samples from wishart
(vars1,vars2,covmat1) = wishart_2(covs[0,0],covs.diagonal()[1:],corrs[0,1:].toarray(),dof,size=errdist_perms)
covs_sim=[]
for a in np.arange(vars2.shape[0]):
covmat=A.get_covs()*0
covmat[0,0]=vars1[a,0]
diag_inds=np.diag_indices(covs.shape[-1])
covmat[diag_inds[0][1:],diag_inds[1][1:]]=vars2[a,:]
covmat[0,1:]=covmat1[a,:]
covs_sim.append(covmat)
else:
# simulate underlying "true" covariance
whA=stats.wishart(dof,np.squeeze(covs)[:,:])
covs_sim=whA.rvs(10*errdist_perms)/(dof)
covs_sim=list(covs_sim)
# create simulated observed covariances from underlying true cov.
ppA=np.zeros((1,10*errdist_perms))
whA=[]
for yb in np.arange(10*errdist_perms):
whA.append(stats.wishart(dof,covs_sim[yb]))
ppA[0,yb]=whA[-1].pdf(np.squeeze(A.get_covs())*dof)
# generate sample distribution of covariances
ppA=ppA/sum(ppA)
ppA_cul=(np.dot(ppA,np.triu(np.ones(len(ppA.T)))).T) ## memory issues
rand_els = stats.uniform(0,1).rvs(errdist_perms)
els=np.sort(np.searchsorted(ppA_cul.flatten(),rand_els))
covs_sim=[]
for xb in np.arange(errdist_perms):
covs_sim.append(whA[els[xb]].rvs()/dof)
self.covs_sim=covs_sim
return(self.covs_sim)
def test_quantile_dimensions(self):
# Test that we can call the Wishart rvs with various quantile dimensions
# If dim == 1, consider x.shape = [1,1,1]
X = [
1, # scalar
[1], # iterable
np.array(1), # 0-dim
np.r_[1], # 1-dim
np.array(1, ndmin=2), # 2-dim
np.array([1], ndmin=3) # 3-dim
]
w = wishart(1,1)
density = w.pdf(np.array(1, ndmin=3))
for x in X:
assert_equal(w.pdf(x), density)
# If dim == 1, consider x.shape = [1,1,*]
X = [
[1,2,3], # iterable
np.r_[1,2,3], # 1-dim
np.array([1,2,3], ndmin=3) # 3-dim
]
w = wishart(1,1)
density = w.pdf(np.array([1,2,3], ndmin=3))
for x in X:
assert_equal(w.pdf(x), density)
# If dim == 2, consider x.shape = [2,2,1]
# where x[:,:,*] = np.eye(1)*2
X = [
2, # scalar
[2,2], # iterable
np.array(2), # 0-dim
np.r_[2,2], # 1-dim
np.array([[2,0],
[0,2]]), # 2-dim
np.array([[2,0],
[0,2]])[:,:,np.newaxis] # 3-dim
]
w = wishart(2,np.eye(2))
density = w.pdf(np.array([[2,0],
[0,2]])[:,:,np.newaxis])
for x in X:
assert_equal(w.pdf(x), density)
def test_quantile_dimensions(self):
# Test that we can call the Wishart rvs with various quantile dimensions
# If dim == 1, consider x.shape = [1,1,1]
X = [
1, # scalar
[1], # iterable
np.array(1), # 0-dim
np.r_[1], # 1-dim
np.array(1, ndmin=2), # 2-dim
np.array([1], ndmin=3) # 3-dim
]
w = wishart(1,1)
density = w.pdf(np.array(1, ndmin=3))
for x in X:
assert_equal(w.pdf(x), density)
# If dim == 1, consider x.shape = [1,1,*]
X = [
[1,2,3], # iterable
np.r_[1,2,3], # 1-dim
np.array([1,2,3], ndmin=3) # 3-dim
]
w = wishart(1,1)
density = w.pdf(np.array([1,2,3], ndmin=3))
for x in X:
assert_equal(w.pdf(x), density)
# If dim == 2, consider x.shape = [2,2,1]
# where x[:,:,*] = np.eye(1)*2
X = [
2, # scalar
[2,2], # iterable
np.array(2), # 0-dim
np.r_[2,2], # 1-dim
np.array([[2,0],
[0,2]]), # 2-dim
np.array([[2,0],
[0,2]])[:,:,np.newaxis] # 3-dim
]
w = wishart(2,np.eye(2))
density = w.pdf(np.array([[2,0],
[0,2]])[:,:,np.newaxis])
for x in X:
assert_equal(w.pdf(x), density)
def test_cholesky_transform():
d = 10
A = stats.wishart(d, np.eye(d)).rvs()
cho_factor = transform.CholeskyTransform(linalg.cho_factor(A))
U = np.triu(cho_factor.U)
iU = np.linalg.inv(U)
b = np.random.rand(d)
assert np.allclose(cho_factor * b, U @ b)
assert np.allclose(b * cho_factor, b @ U)
assert np.allclose(cho_factor.ldiv(b), iU @ b)
assert np.allclose(b / cho_factor, b @ iU)
b = np.random.rand(d, d)
assert np.allclose(cho_factor * b, U @ b)
assert np.allclose(b * cho_factor, b @ U)
assert np.allclose(cho_factor.ldiv(b), iU @ b)
assert np.allclose(b / cho_factor, b @ iU)
b = np.random.rand(d, d + 1)
assert np.allclose(cho_factor * b, U @ b)
assert np.allclose(cho_factor.ldiv(b), iU @ b)
b = np.random.rand(d + 1, d)
assert np.allclose(b * cho_factor, b @ U)
assert np.allclose(b / cho_factor, b @ iU)
def test_frozen(self):
# Test that the frozen and non-frozen Wishart gives the same answers
# Construct an arbitrary positive definite scale matrix
dim = 4
scale = np.diag(np.arange(dim)+1)
scale[np.tril_indices(dim, k=-1)] = np.arange(dim * (dim-1) // 2)
scale = np.dot(scale.T, scale)
# Construct a collection of positive definite matrices to test the PDF
X = []
for i in range(5):
x = np.diag(np.arange(dim)+(i+1)**2)
x[np.tril_indices(dim, k=-1)] = np.arange(dim * (dim-1) // 2)
x = np.dot(x.T, x)
X.append(x)
X = np.array(X).T
# Construct a 1D and 2D set of parameters
parameters = [
(10, 1, np.linspace(0.1, 10, 5)), # 1D case
(10, scale, X)
]
for (df, scale, x) in parameters:
w = wishart(df, scale)
assert_equal(w.var(), wishart.var(df, scale))
assert_equal(w.mean(), wishart.mean(df, scale))
assert_equal(w.mode(), wishart.mode(df, scale))
assert_equal(w.entropy(), wishart.entropy(df, scale))
assert_equal(w.pdf(x), wishart.pdf(x, df, scale))
def test_frozen(self):
# Test that the frozen and non-frozen Wishart gives the same answers
# Construct an arbitrary positive definite scale matrix
dim = 4
scale = np.diag(np.arange(dim) + 1)
scale[np.tril_indices(dim, k=-1)] = np.arange(dim * (dim - 1) // 2)
scale = np.dot(scale.T, scale)
# Construct a collection of positive definite matrices to test the PDF
X = []
for i in range(5):
x = np.diag(np.arange(dim) + (i + 1)**2)
x[np.tril_indices(dim, k=-1)] = np.arange(dim * (dim - 1) // 2)
x = np.dot(x.T, x)
X.append(x)
X = np.array(X).T
# Construct a 1D and 2D set of parameters
parameters = [
(10, 1, np.linspace(0.1, 10, 5)), # 1D case
(10, scale, X)
]
for (df, scale, x) in parameters:
w = wishart(df, scale)
assert_equal(w.var(), wishart.var(df, scale))
assert_equal(w.mean(), wishart.mean(df, scale))
assert_equal(w.mode(), wishart.mode(df, scale))
assert_equal(w.entropy(), wishart.entropy(df, scale))
assert_equal(w.pdf(x), wishart.pdf(x, df, scale))
def rvs(self):
"""Draw one sample. See p.582 of Gelman."""
x = stats.wishart(df=self.ν).rvs()
z = stats.multivariate_normal(mean=np.zeros(self.D),
cov=np.eye(self.D)).rvs()
A = chol(inv(self.Λ))
return self.μ + A @ z * sqrt(self.ν / x)
def _sample_gaussian_parameters(self, X, S, m_list, beta_list, nd_list,
W_list, N, K, D):
""" ガウス分布のパラメータΛ_k, μ_kをサンプルする
Parameter
X : D*N行列 入力データ
S : K*N行列 予測した各クラスの割り当て
m_list, beta_list, nd_list, W_list
N : データ数
K : クラスタ数 """
for k in range(K):
sum_s = 0.0
sum_sx = np.zeros((D, 1))
sum_sxx = np.zeros((D, D))
for n in range(N):
sum_s += S[k, n]
sum_sx += S[k, n] * X[:, n:n + 1]
sum_sxx += S[k, n] * np.dot(X[:, n:n + 1], X[:, n:n + 1].T)
# パラメータ更新，この順番に更新するのが正しい?
beta_list[k] = sum_s + self._beta
m_list[k] = (sum_sx +
np.dot(self._beta, self._m_vector)) / beta_list[k]
W_list[k] = np.linalg.inv(
sum_sxx +
np.dot(self._beta, np.dot(self._m_vector, self._m_vector.T)) -
np.dot(beta_list[k], np.dot(m_list[k], m_list[k].T)) +
np.linalg.inv(self._W))
nd_list[k] = sum_s + self._nd
# ガウス分布のパラメータをサンプリング，この順番に更新するのが正しい
wish = stats.wishart(df=nd_list[k], scale=W_list[k])
self.cov_matrixes[k] = np.linalg.inv(wish.rvs(1))
self.mu_vectors[k] = np.random.multivariate_normal(
np.squeeze(m_list[k].T),
1 / beta_list[k] * self.cov_matrixes[k])
# pdb.set_trace()
return
def __init__(self, D, a0, b0, m0, beta0):
self.D = D
self.a0 = a0
self.B0 = np.eye(D) * b0
self.m0 = m0
self.beta0 = beta0
self.precision_prior = stats.wishart(self.a0, self.B0)
def rprior(n=1,r=3,M=np.eye(2)):
""" calculate r prior: n,r,M. """
out=np.zeros((n,len(M),len(M)))
Minv=np.la.pinv(M)
for a in np.arange(n):
out[a,:,:]= scipy.linalg.cho_solve(scipy.linalg.cho_factor( stats.wishart(r,Minv).rvs()),np.eye(len(Minv)))
return(out)
def test_is_scaled_chisquared(self):
# The 2-dimensional Wishart with an arbitrary scale matrix can be
# transformed to a scaled chi-squared distribution.
# For :math:`S \sim W_p(V,n)` and :math:`\lambda \in \mathbb{R}^p` we have
# :math:`\lambda' S \lambda \sim \lambda' V \lambda \times \chi^2(n)`
np.random.seed(482974)
sn = 500
df = 10
dim = 4
# Construct an arbitrary positive definite matrix
scale = np.diag(np.arange(4)+1)
scale[np.tril_indices(4, k=-1)] = np.arange(6)
scale = np.dot(scale.T, scale)
# Use :math:`\lambda = [1, \dots, 1]'`
lamda = np.ones((dim,1))
sigma_lamda = lamda.T.dot(scale).dot(lamda).squeeze()
w = wishart(df, sigma_lamda)
c = chi2(df, scale=sigma_lamda)
# Statistics
assert_allclose(w.var(), c.var())
assert_allclose(w.mean(), c.mean())
assert_allclose(w.entropy(), c.entropy())
# PDF
X = np.linspace(0.1,10,num=10)
assert_allclose(w.pdf(X), c.pdf(X))
# rvs
rvs = w.rvs(size=sn)
args = (df,0,sigma_lamda)
alpha = 0.01
check_distribution_rvs('chi2', args, alpha, rvs)
def test_is_scaled_chisquared(self):
# The 2-dimensional Wishart with an arbitrary scale matrix can be
# transformed to a scaled chi-squared distribution.
# For :math:`S \sim W_p(V,n)` and :math:`\lambda \in \mathbb{R}^p` we have
# :math:`\lambda' S \lambda \sim \lambda' V \lambda \times \chi^2(n)`
np.random.seed(482974)
sn = 500
df = 10
dim = 4
# Construct an arbitrary positive definite matrix
scale = np.diag(np.arange(4) + 1)
scale[np.tril_indices(4, k=-1)] = np.arange(6)
scale = np.dot(scale.T, scale)
# Use :math:`\lambda = [1, \dots, 1]'`
lamda = np.ones((dim, 1))
sigma_lamda = lamda.T.dot(scale).dot(lamda).squeeze()
w = wishart(df, sigma_lamda)
c = chi2(df, scale=sigma_lamda)
# Statistics
assert_allclose(w.var(), c.var())
assert_allclose(w.mean(), c.mean())
assert_allclose(w.entropy(), c.entropy())
# PDF
X = np.linspace(0.1, 10, num=10)
assert_allclose(w.pdf(X), c.pdf(X))
# rvs
rvs = w.rvs(size=sn)
args = (df, 0, sigma_lamda)
alpha = 0.01
check_distribution_rvs('chi2', args, alpha, rvs)
def __init__(self, df, scale, size=1, preload=1, *args, **kwargs):
# Initialize parameters
self._frozen = wishart(df, scale)
self._rvs = self._frozen._wishart
self.df = self.prior_df = self._frozen.df
self.scale = self.prior_scale = self._frozen.scale
# Calculated quantities
self._inv_prior_scale = np.linalg.inv(self.prior_scale)
# Setup holder variables for posterior-related quantities
self._phi = None # (M x M)
self._lagged = None # (M x T)
self._endog = None # (M x T)
# Setup holder variables for calculated quantities
self._philagged = None
self._posterior_df = None
self._posterior_scale = None
self._posterior_cholesky = None
# Set the flag to use the prior
self._use_posterior = False
# Initialize the distribution
super(Wishart, self).__init__(None, size=size, preload=preload,
*args, **kwargs)
def test_1D_is_chisquared(self):
# The 1-dimensional Wishart with an identity scale matrix is just a
# chi-squared distribution.
# Test variance, mean, entropy, pdf
# Kolgomorov-Smirnov test for rvs
np.random.seed(482974)
sn = 500
dim = 1
scale = np.eye(dim)
df_range = np.arange(1, 10, 2, dtype=float)
X = np.linspace(0.1,10,num=10)
for df in df_range:
w = wishart(df, scale)
c = chi2(df)
# Statistics
assert_allclose(w.var(), c.var())
assert_allclose(w.mean(), c.mean())
assert_allclose(w.entropy(), c.entropy())
# PDF
assert_allclose(w.pdf(X), c.pdf(X))
# rvs
rvs = w.rvs(size=sn)
args = (df,)
alpha = 0.01
check_distribution_rvs('chi2', args, alpha, rvs)
def __init__(self, df, scale, size=1, preload=1, *args, **kwargs):
# Initialize parameters
self._frozen = wishart(df, scale)
self._rvs = self._frozen._wishart
self.df = self.prior_df = self._frozen.df
self.scale = self.prior_scale = self._frozen.scale
# Calculated quantities
self._inv_prior_scale = np.linalg.inv(self.prior_scale)
# Setup holder variables for posterior-related quantities
self._phi = None # (M x M)
self._lagged = None # (M x T)
self._endog = None # (M x T)
# Setup holder variables for calculated quantities
self._philagged = None
self._posterior_df = None
self._posterior_scale = None
self._posterior_cholesky = None
# Set the flag to use the prior
self._use_posterior = False
# Initialize the distribution
super(Wishart, self).__init__(None,
size=size,
preload=preload,
*args,
**kwargs)
def test_1D_is_chisquared(self):
# The 1-dimensional Wishart with an identity scale matrix is just a
# chi-squared distribution.
# Test variance, mean, entropy, pdf
# Kolgomorov-Smirnov test for rvs
np.random.seed(482974)
sn = 500
dim = 1
scale = np.eye(dim)
df_range = np.arange(1, 10, 2, dtype=float)
X = np.linspace(0.1, 10, num=10)
for df in df_range:
w = wishart(df, scale)
c = chi2(df)
# Statistics
assert_allclose(w.var(), c.var())
assert_allclose(w.mean(), c.mean())
assert_allclose(w.entropy(), c.entropy())
# PDF
assert_allclose(w.pdf(X), c.pdf(X))
# rvs
rvs = w.rvs(size=sn)
args = (df, )
alpha = 0.01
check_distribution_rvs('chi2', args, alpha, rvs)
def Normal_Wishart(mu_0, lamb, W, nu, seed=None):
"""Function extracting a Normal_Wishart random variable"""
# first draw a Wishart distribution:
Lambda = wishart(df=nu, scale=W, seed=seed).rvs() # NB: Lambda is a matrix.
# then draw a Gaussian multivariate RV with mean mu_0 and(lambda*Lambda)^{-1} as covariance matrix.
cov = np.linalg.inv(lamb * Lambda) # this is the bottleneck!!
mu = multivariate_normal(mu_0, cov)
return mu, Lambda, cov
def test_wishart_invwishart_2D_rvs(self):
dim = 3
df = 10
# Construct a simple non-diagonal positive definite matrix
scale = np.eye(dim)
scale[0,1] = 0.5
scale[1,0] = 0.5
# Construct frozen Wishart and inverse Wishart random variables
w = wishart(df, scale)
iw = invwishart(df, scale)
# Get the generated random variables from a known seed
np.random.seed(248042)
w_rvs = wishart.rvs(df, scale)
np.random.seed(248042)
frozen_w_rvs = w.rvs()
np.random.seed(248042)
iw_rvs = invwishart.rvs(df, scale)
np.random.seed(248042)
frozen_iw_rvs = iw.rvs()
# Manually calculate what it should be, based on the Bartlett (1933)
# decomposition of a Wishart into D A A' D', where D is the Cholesky
# factorization of the scale matrix and A is the lower triangular matrix
# with the square root of chi^2 variates on the diagonal and N(0,1)
# variates in the lower triangle.
np.random.seed(248042)
covariances = np.random.normal(size=3)
variances = np.r_[
np.random.chisquare(df),
np.random.chisquare(df-1),
np.random.chisquare(df-2),
]**0.5
# Construct the lower-triangular A matrix
A = np.diag(variances)
A[np.tril_indices(dim, k=-1)] = covariances
# Wishart random variate
D = np.linalg.cholesky(scale)
DA = D.dot(A)
manual_w_rvs = np.dot(DA, DA.T)
# inverse Wishart random variate
# Supposing that the inverse wishart has scale matrix `scale`, then the
# random variate is the inverse of a random variate drawn from a Wishart
# distribution with scale matrix `inv_scale = np.linalg.inv(scale)`
iD = np.linalg.cholesky(np.linalg.inv(scale))
iDA = iD.dot(A)
manual_iw_rvs = np.linalg.inv(np.dot(iDA, iDA.T))
# Test for equality
assert_allclose(w_rvs, manual_w_rvs)
assert_allclose(frozen_w_rvs, manual_w_rvs)
assert_allclose(iw_rvs, manual_iw_rvs)
assert_allclose(frozen_iw_rvs, manual_iw_rvs)
def test_wishart_invwishart_2D_rvs(self):
dim = 3
df = 10
# Construct a simple non-diagonal positive definite matrix
scale = np.eye(dim)
scale[0,1] = 0.5
scale[1,0] = 0.5
# Construct frozen Wishart and inverse Wishart random variables
w = wishart(df, scale)
iw = invwishart(df, scale)
# Get the generated random variables from a known seed
np.random.seed(248042)
w_rvs = wishart.rvs(df, scale)
np.random.seed(248042)
frozen_w_rvs = w.rvs()
np.random.seed(248042)
iw_rvs = invwishart.rvs(df, scale)
np.random.seed(248042)
frozen_iw_rvs = iw.rvs()
# Manually calculate what it should be, based on the Bartlett (1933)
# decomposition of a Wishart into D A A' D', where D is the Cholesky
# factorization of the scale matrix and A is the lower triangular matrix
# with the square root of chi^2 variates on the diagonal and N(0,1)
# variates in the lower triangle.
np.random.seed(248042)
covariances = np.random.normal(size=3)
variances = np.r_[
np.random.chisquare(df),
np.random.chisquare(df-1),
np.random.chisquare(df-2),
]**0.5
# Construct the lower-triangular A matrix
A = np.diag(variances)
A[np.tril_indices(dim, k=-1)] = covariances
# Wishart random variate
D = np.linalg.cholesky(scale)
DA = D.dot(A)
manual_w_rvs = np.dot(DA, DA.T)
# inverse Wishart random variate
# Supposing that the inverse wishart has scale matrix `scale`, then the
# random variate is the inverse of a random variate drawn from a Wishart
# distribution with scale matrix `inv_scale = np.linalg.inv(scale)`
iD = np.linalg.cholesky(np.linalg.inv(scale))
iDA = iD.dot(A)
manual_iw_rvs = np.linalg.inv(np.dot(iDA, iDA.T))
# Test for equality
assert_allclose(w_rvs, manual_w_rvs)
assert_allclose(frozen_w_rvs, manual_w_rvs)
assert_allclose(iw_rvs, manual_iw_rvs)
assert_allclose(frozen_iw_rvs, manual_iw_rvs)
def sample(self):
self._check_is_valid_density()
precision = wishart(df=self._get_nu(),
scale=np.linalg.inv(self._get_psi())).rvs()
L = np.linalg.cholesky(precision) * \
np.sqrt(self.natural_parameters['kappa'])
mean = self.get_mean() + \
np.linalg.solve(L.T, np.random.normal(size=self.num_dim))
return dict(mean=mean, precision=precision)
def Normal_Wishart(mu_0, lamb, W, nu, seed=None):
"""Function extracting a Normal_Wishart random variable"""
# first draw a Wishart distribution:
Lambda = wishart(df=nu, scale=W,
seed=seed).rvs() # NB: Lambda is a matrix.
# then draw a Gaussian multivariate RV with mean mu_0 and(lambda*Lambda)^{-1} as covariance matrix.
cov = np.linalg.inv(lamb * Lambda) # this is the bottleneck!!
mu = multivariate_normal(mu_0, cov)
return mu, Lambda, cov
def wishart_for_cov(dim, seed=None):
# returns positive definite matrix for dimension dim generated from the wishart distribution with designated
# degreee of freedom and identity scale matrix. recommend degree of freedom = dimension of matrix
#
if not seed is None:
numpy.random.seed(seed)
else:
pass
wishart_obj = wishart()
out = wishart.rvs(df=dim, scale=numpy.eye(dim), size=1)
return (out)
def wishart_pdf(cov,samples,dof):
""" wishart pdf given cov and dof. """
if scipy.sparse.issparse(cov):
var1=cov[0,0]
var2=cov.diagonal()[1:]
covs1=covs[0,:]
for a in len(var2):
whA.append(stats.wishart(dof,covsA_sim[b,:,:]))
ppA[0,yb]=whA[-1].pdf(A.get_covs()[xa,:,:]*dof)
return(cov)
def compute_elbo(self, phi, nu, kappa, epsilon, m, L, V, N, gamma_1, gamma_2):
"""
Function compute the evidence lower bound as defined for HRMF-DPCMM From the variational parameters.
:param phi: of shape [N,K]
:param nu: of shape [1, K]
:param kappa: of shape [1, K]
:param epsilon: of shape [1, K]
:param m: of shape [N, d]
:param L: of shape [N, d, d]
:param V: of shape [K, K]
:param N: of shape [1, K]
:param gamma_1: of shape [1, K]
:param gamma_2: of shape [1, K]
:return:
"""
hmrf_term = 0
log_likelihood_term = 0
if self.weight_prior == "Dirichelet distribution":
val = digamma(epsilon) - digamma(np.sum(epsilon))
else:
val = digamma(gamma_1) - digamma(gamma_1 + gamma_2) + cumsum_ex(
digamma(gamma_2) - digamma(gamma_1 + gamma_2))
for n in range(self.N):
for k in range(self.K):
log_likelihood_term += - 0.5 * phi[n, k] * nu[k] * np.trace(np.matmul(L[k,:,:], np.matmul(self.X[n,:].reshape(self.d, 1) - m[k].reshape(self.d, 1),
self.X[n,:].reshape(1, self.d) - m[k].reshape(1, self.d) )) ) \
- 0.5 * self.d * N[k] / kappa[k]
if self.mask[n] == 0:
log_likelihood_term += phi[n,k] * val[k]
for k in range(self.K):
log_likelihood_term += 0.5 * (nu[k] - self.d + N[k]) * (multivar_digamma(nu[k], self.d) + np.log(self.eps + LA.det(L[k, :, :]))) - 0.5 * self.d * nu[k]
for tuple in self.tuples_ml:
hmrf_term += - self.lambda_ * np.sum(np.sum(phi[tuple[0],:].reshape(self.K, 1) * phi[tuple[1],:].reshape(1,self.K) * V))
for k in range(self.K):
log_likelihood_term += wishart(nu[k],L[k,:,:]).entropy() + 0.5*(multivar_digamma(nu[k], self.d) + np.log(self.eps + LA.det(L[k, :, :]))) + 0.5*self.d*np.log(eps + kappa[k]) \
- np.sum(phi[:,k] * np.log(phi[:,k] + eps))
if self.weight_prior != "Dirichelet distribution":
log_likelihood_term += beta(gamma_1[k], gamma_2[k]).entropy()
if self.weight_prior == "Dirichelet distribution":
log_likelihood_term += dirichlet(epsilon).entropy()
elbo = log_likelihood_term + hmrf_term
return elbo/self.N, log_likelihood_term, hmrf_term
def test_scale_dimensions(self):
# Test that we can call the Wishart with various scale dimensions
# Test case: dim=1, scale=1
true_scale = np.array(1, ndmin=2)
scales = [
1, # scalar
[1], # iterable
np.array(1), # 0-dim
np.r_[1], # 1-dim
np.array(1, ndmin=2) # 2-dim
]
for scale in scales:
w = wishart(1, scale)
assert_equal(w.scale, true_scale)
assert_equal(w.scale.shape, true_scale.shape)
# Test case: dim=2, scale=[[1,0]
# [0,2]
true_scale = np.array([[1, 0], [0, 2]])
scales = [
[1, 2], # iterable
np.r_[1, 2], # 1-dim
np.array([
[1, 0], # 2-dim
[0, 2]
])
]
for scale in scales:
w = wishart(2, scale)
assert_equal(w.scale, true_scale)
assert_equal(w.scale.shape, true_scale.shape)
# We cannot call with a df < dim
assert_raises(ValueError, wishart, 1, np.eye(2))
# We cannot call with a 3-dimension array
scale = np.array(1, ndmin=3)
assert_raises(ValueError, wishart, 1, scale)
def test_scale_dimensions(self):
# Test that we can call the Wishart with various scale dimensions
# Test case: dim=1, scale=1
true_scale = np.array(1, ndmin=2)
scales = [
1, # scalar
[1], # iterable
np.array(1), # 0-dim
np.r_[1], # 1-dim
np.array(1, ndmin=2) # 2-dim
]
for scale in scales:
w = wishart(1, scale)
assert_equal(w.scale, true_scale)
assert_equal(w.scale.shape, true_scale.shape)
# Test case: dim=2, scale=[[1,0]
# [0,2]
true_scale = np.array([[1,0],
[0,2]])
scales = [
[1,2], # iterable
np.r_[1,2], # 1-dim
np.array([[1,0], # 2-dim
[0,2]])
]
for scale in scales:
w = wishart(2, scale)
assert_equal(w.scale, true_scale)
assert_equal(w.scale.shape, true_scale.shape)
# We cannot call with a df < dim
assert_raises(ValueError, wishart, 1, np.eye(2))
# We cannot call with a 3-dimension array
scale = np.array(1, ndmin=3)
assert_raises(ValueError, wishart, 1, scale)
def test_multivariate_normal_density(self):
for i in range(4):
with util.NumpySeedContext(seed=i + 8):
d = i + 2
cov = stats.wishart(df=10 + d, scale=np.eye(d)).rvs(size=1)
mean = np.random.randn(d)
X = np.random.randn(11, d)
den_estimate = density.GaussianMixture.multivariate_normal_density(
mean, cov, X)
mnorm = stats.multivariate_normal(mean=mean, cov=cov)
den_truth = mnorm.pdf(X)
np.testing.assert_almost_equal(den_estimate, den_truth)
def test_kl_between_mvn_and_std():
"""
Check that our custom implementation of KL divergence for MVN against std MVN matches KL divergence from
PyTorch's distribution module.
"""
import time
torch.manual_seed(512)
rng = np.random.RandomState(51)
batch_size = 12
num_points_per_batch = 57
# Create a distribution to check
loc = torch.randn(batch_size,
num_points_per_batch,
dtype=utils.TORCH_FLOAT_TYPE)
wishart_ = stats.wishart(seed=rng,
df=num_points_per_batch,
scale=np.eye(num_points_per_batch))
cov_samples = np.stack([wishart_.rvs() for _ in range(batch_size)])
cov_samples = torch.tensor(cov_samples, dtype=utils.TORCH_FLOAT_TYPE)
dist_1 = distributions.MultivariateNormal(loc,
covariance_matrix=cov_samples)
# Create a std normal
mn = torch.zeros(batch_size,
num_points_per_batch,
dtype=utils.TORCH_FLOAT_TYPE)
cov = torch.stack([
torch.eye(num_points_per_batch, dtype=utils.TORCH_FLOAT_TYPE)
for _ in range(batch_size)
])
std_norm = distributions.MultivariateNormal(mn, covariance_matrix=cov)
# Do the computed/expected
time_s = time.time()
computed_kl = custom_distributions.kl_mvn_and_std_norm(
dist_1).detach().numpy()
time_mid = time.time()
expected_kl = distributions.kl_divergence(dist_1,
std_norm).detach().numpy()
time_end = time.time()
print(
f"Pytorch impl: {time_end-time_mid}s; Custom imp: {time_mid-time_s}s")
# ^ not very scientific but sanity check to make sure that worth avoiding the PyTorch implementation in the
# against std normal case.
# Test!
np.testing.assert_array_almost_equal(computed_kl, expected_kl)
def set_cluster_variance(cls, dimension, tightness=5, standard=False):
"""
Randomly sample a symmetric matrix from a Wishart distribution.
:param dimension: An integer that specifies the dimension of the data space.
:param tightness: A float that controls the amount by which a covariance matrix is scaled.
:param standard: A boolean that determines whether the covariance matrix is the identity matrix or is random.
:return: A symmetric positive-definite covariance matrix.
"""
if standard:
return np.identity(dimension)
else:
w = wishart(df=dimension, scale=np.identity(dimension))
cluster_variance = w.rvs()
cluster_variance = np.random.uniform(0.1, tightness) * cluster_variance
return cluster_variance
def global_bge_prior(
graph,
total_num_variables=None,
inverse_scale_matrix=None,
degrees_freedom=None,
alpha_mu=None,
mu0=None,
size=1,
progress=False
):
p = total_num_variables
variables = graph.nodes
k = len(variables)
B = np.zeros((k, k))
V = list(variables)
scale_matrix = faster_inverse(inverse_scale_matrix)
# # Normal distribution vectorized function
# standard_normal = lambda t: np.random.normal(0, 1)
# vfunc_standard_normal = np.vectorize(standard_normal)
# if len(B[indices]) > 0:
# B[indices] = vfunc_standard_normal(B[indices])
#
# c_squared = np.zeros(k)
# for i in range(k):
# c_squared[i] = stats.chi2.rvs(df=degrees_freedom - p + i + 1)
#
# c = np.sqrt(c_squared)
# inverse_c = 1 / c
# B = np.multiply(-np.array(B), inverse_c)
# d = np.multiply(scale_matrix, c_squared)
#
# I = np.eye(len(variables))
# A = I - B.T
# inverse_sigma = A.T @ d @ A
# TODO pull directly from wishart, compare to other way
inverse_sigmas = stats.wishart(df=degrees_freedom, scale=scale_matrix).rvs(size=size)
sigmas = [faster_inverse(inverse_sigma) for inverse_sigma in inverse_sigmas]
# ipdb.set_trace()
mu_covariances = [(1 / alpha_mu) * sigma for sigma in sigmas]
# mu = stats.multivariate_normal(mean=mu0[V], cov=mu_covariance).rvs()
mus = [chol_sample(mu0[V], mu_covariance) for mu_covariance in mu_covariances]
return list(zip(inverse_sigmas, mus))
def _fit(self, X):
self.mean_ = np.mean(X, axis=0)
data = X - self.mean_
cov = np.dot(data.T, data)
w = ss.wishart(df=data.shape[1] + 1,
scale=np.matrix(
np.eye(data.shape[1]) * 3 * self.max_norm /
(2 * data.shape[0] * self.eps)))
noise = w.rvs(1, random_state=self.random_state)
cov = cov + noise
cov = cov / data.shape[0]
ev, evec = np.linalg.eig(cov)
evec = evec.T
self.components_ = evec[:self.n_components]
def updateOneComponent(X, mu, precision, muPrior, precisionPrior):
'''
X: (n,p) array of data
mu: (p,1) array of current mean
precision: (p,p) matrix of current precision
muPrior: dictionary of prior mean and precision
precisionPrior: dictionary of prior df and invScale
'''
n = X.shape[0]
An_inv = inv(muPrior['precision'] + n * precision)
Xsum = np.sum(X, axis=0)
bn = muPrior['precision'].dot(muPrior['mean']) + precision.dot(Xsum)
mu = multivariate_normal(An_inv.dot(bn), An_inv).rvs()
S_mu = np.matmul((X - mu).T, X - mu)
precision = wishart(precisionPrior['df'] + n,
inv(precisionPrior['invScale'] + S_mu)).rvs()
return mu, precision
def sampleNewComp(Knew, muPrior, precisionPrior):
'''
Knew: number of new components to generate
muPrior: dictionary of prior mean and precision
precisionPrior: dictionary of prior df and invScale
return: a list of NEW components
'''
comps = []
if Knew == 0:
return comps
muCov = inv(muPrior['precision'])
precisionScale = inv(precisionPrior['invScale'])
for k in range(Knew):
mu = rng.multivariate_normal(muPrior['mean'], muCov)
precision = wishart(precisionPrior['df'], precisionScale).rvs()
comps.append((mu, precision))
return comps
def sample(self, indices, prior_parameter):
kappa, nu, mu, psi = prior_parameter.get_kappa_nu_mu_psi()
y_indices = self.y[indices]
u_indices = self.parameter.u[indices]
uy_indices = np.outer(u_indices, np.ones(self.num_dim)) * y_indices
u_sum = np.sum(u_indices)
uy_sum = np.sum(uy_indices, axis=0)
uyy_sum = y_indices.T.dot(uy_indices)
kappa_post = kappa + u_sum
nu_post = nu + len(indices)
mu_post = (kappa * mu + uy_sum) / (kappa + u_sum)
psi_post = psi + uyy_sum + (kappa * np.outer(mu, mu) -
kappa_post * np.outer(mu_post, mu_post))
precision = wishart(df=nu_post, scale=np.linalg.inv(psi_post)).rvs()
L = np.linalg.cholesky(precision) * \
np.sqrt(kappa_post)
mean = mu_post + np.linalg.solve(L.T,
np.random.normal(size=self.num_dim))
return dict(mean=mean, precision=precision)
def test_diagonal_normal_wishart():
import numpy as np
from scipy.stats import wishart
x = np.linspace(1e-6, 20, 100)
print("Testing wishart entropy/logprob vs scipy implementation...")
for k in range(1000):
df_val = torch.randn(1).exp() + 2
scale_val = torch.randn(1).exp()
scipy_dist = wishart(df=df_val.item(), scale=scale_val.item())
torch_dist = DiagonalWishart(
scale_val.unsqueeze(-1),
df_val
)
torch_ent = torch_dist.entropy()[0]
scipy_ent = torch.FloatTensor([scipy_dist.entropy()])
if (rel_error(torch_ent, scipy_ent) > 1e-3).any():
raise ValueError(
"Entropies of torch and scipy versions doesn't match"
)
scipy_w = torch.FloatTensor(scipy_dist.logpdf(x))
torch_w = torch_dist.log_prob(torch.FloatTensor(x).unsqueeze(-1))
if (rel_error(torch_w, scipy_w) > 1e-6).any():
raise ValueError(
"Log pdf of torch and scipy versions doesn't match"
)
print("Passed")
print("Testing wishart KL divergence...")
df1, scale1 = torch.randn(32).exp() + 2, torch.randn(32).exp() + 1e-5
df2, scale2 = torch.randn(32).exp() + 2, torch.randn(32).exp() + 1e-5
init_df1, init_scale1 = df1[0].clone(), scale1[0].clone()
dist2 = DiagonalWishart(scale2.unsqueeze(-1), df2)
df1.requires_grad, scale1.requires_grad = True, True
gamma = 0.1
for k in range(10000):
dist1 = DiagonalWishart(scale1.unsqueeze(-1), df1)
loss = kl_divergence(dist1, dist2).mean()
loss.backward()
with torch.no_grad():
scale1 = scale1 - gamma * scale1.grad
df1 = df1 - gamma * df1.grad
scale1.requires_grad, df1.requires_grad = True, True
print('Distribution 1 - initial df %.3f and scale %.3f' % (
init_df1, init_scale1
))
print('Distribution 1 - trained df %.3f and scale %.3f' % (
df1[0], scale1[0]
))
print('Distribution 2 - target df %.3f and scale %.3f' % (
df2[0], scale2[0]
))
print("Passed")
print("Testing Normal Wishart Distribution...")
torch_dist = NormalDiagonalWishart(
torch.ones((100, 1, 32, 32, 1)),
torch.ones((100, 1, 32, 32, 1)),
3 * torch.ones((100, 1, 32, 32)),
3 * torch.ones((100, 1, 32, 32)),
)
ex_w = torch_dist.log_prob(
torch.ones(100, 1, 32, 32, 1),
torch.ones(100, 1, 32, 32, 1),
)
assert ex_w.shape == (100, 1, 32, 32)
print("Passed")
covsTeo, covsNum = np.array(covSuperList).transpose((1, 2, 0, 3, 4, 5)).reshape((2, 4, -1, 2, 2)) scaleNum = np.linalg.inv(covsNum) # saco la norma de frobenius de cada matriz covSuperFrob = np.linalg.norm(covSuperList, axis=(4, 5)) # %% p = 2 matt = np.eye(p) np.exp(-p/2) / spe.gamma(N/2) / 2**(N*p/2) / np.linalg.det(matt)**(p/2+0.5) sts.wishart.pdf(matt, df=N-1, scale=matt) # %% rv = sts.wishart() rv.pdf() frobQuotList = (covSuperFrob[:, 0] / covSuperFrob[:,1]).transpose((1, 0, 2)).reshape((4, -1)) plt.plot(frobQuotList) plt.hist(frobQuotList[3]) plt.violinplot(frobQuotList.T, showmeans=True, showextrema=False)