def ica(mock, ell=0, rebin=None):
''' *** TESTED ***
Test that the ICA works!
'''
Pk = NG.dataX(mock, ell=ell, rebin=rebin)
X, _ = NG.meansub(Pk)
X_w, W = NG.whiten(X) # whitened data
X_ica, W = NG.Ica(X_w)
# compare covariance?
C_X = np.cov(X.T)
C_Xica = np.cov(X_ica.T)
prettyplot()
fig = plt.figure(figsize=(20, 8))
sub = fig.add_subplot(121)
im = sub.imshow(np.log10(C_X), interpolation='none')
sub.set_title('log(Cov.) of Data')
fig.colorbar(im, ax=sub)
sub = fig.add_subplot(122)
im = sub.imshow(C_Xica, interpolation='none')
fig.colorbar(im, ax=sub)
sub.set_title('Cov. of ICA transformed Data')
# save fig
if rebin is None:
f = ''.join([UT.fig_dir(), 'tests/test.ICAcov.', mock, '.ell', str(ell), '.png'])
else:
f = ''.join([UT.fig_dir(), 'tests/test.ICAcov.', mock, '.ell', str(ell), '.rebin', str(rebin), '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def whiten(mock, ell=0, rebin=None, krange=None, method='choletsky'):
''' ***TESTED: Choletsky decomposition fails for full binned Nseries
P(k) because the precision matrix estimate is not positive definite***
test the data whitening.
'''
Pk = NG.dataX(mock, ell=ell, rebin=rebin, krange=krange)
X, _ = NG.meansub(Pk)
X_w, W = NG.whiten(X, method=method) # whitened data
prettyplot()
fig = plt.figure(figsize=(15,7))
sub = fig.add_subplot(121)
for i in range(X.shape[1]):
sub.plot(range(X_w.shape[0]), X_w[:,i])
sub.set_xlim([0, X.shape[0]])
sub.set_xlabel('$\mathtt{k}$ bins', fontsize=25)
sub.set_ylim([-7., 7.])
sub.set_ylabel('$\mathtt{W^{T} (P^i_'+str(ell)+'- \overline{P_'+str(ell)+'})}$', fontsize=25)
C_Xw = np.cov(X_w.T)
sub = fig.add_subplot(122)
im = sub.imshow(C_Xw, interpolation='none')
fig.colorbar(im, ax=sub)
if rebin is None:
f = ''.join([UT.fig_dir(), 'tests/test.whiten.', method, '.', mock, '.ell', str(ell), '.png'])
else:
f = ''.join([UT.fig_dir(), 'tests/test.whiten.', method, '.', mock, '.ell', str(ell), '.rebin', str(rebin), '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def whiten_recon(mock, ell=0, rebin=None, krange=None, method='choletsky'):
''' ***TESTED: The whitening matrices reconstruct the P(k)s***
Test whether P(k) can be reconstructed using the whitening matrix
'''
Pk, k = NG.dataX(mock, ell=ell, rebin=rebin, krange=krange, k_arr=True)
X, mu_X = NG.meansub(Pk)
X_w, W = NG.whiten(X, method=method) # whitened data
prettyplot()
fig = plt.figure(figsize=(15,7))
sub = fig.add_subplot(121)
for i in range(X.shape[0]):
sub.plot(k, Pk[i,:])
if krange is None:
sub.set_xlim([1e-3, 0.5])
else:
sub.set_xlim(krange)
sub.set_xscale('log')
sub.set_xlabel('$\mathtt{k}$', fontsize=25)
sub.set_yscale('log')
sub.set_ylim([2e3, 2.5e5])
np.random.seed(7)
sub = fig.add_subplot(122)
for i in range(X.shape[0]):
X_noise = np.random.normal(size=X_w.shape[1])
X_rec = np.linalg.solve(W.T, X_noise.T)
sub.plot(k, X_rec.T + mu_X)
if krange is None:
sub.set_xlim([1e-3, 0.5])
else:
sub.set_xlim(krange)
sub.set_xscale('log')
sub.set_xlabel('$\mathtt{k}$', fontsize=25)
sub.set_yscale('log')
sub.set_ylim([2e3, 2.5e5])
if rebin is None:
f = ''.join([UT.fig_dir(), 'tests/test.whiten_recon.', method, '.', mock, '.ell', str(ell), '.png'])
else:
f = ''.join([UT.fig_dir(), 'tests/test.whiten_recon.', method, '.', mock, '.ell', str(ell), '.rebin', str(rebin), '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def p_Xw_i_MISE(mock, ell=0, rebin=None, krange=None, method='choletsky', b=0.1):
''' Examine the pdf of X_w^i components that deviate significantly from
N(0,1) based on MISE
'''
Pk = NG.dataX(mock, ell=ell, rebin=rebin, krange=krange)
X, _ = NG.meansub(Pk)
X_w, W = NG.whiten(X, method=method) # whitened data
# calculate the chi-squared values of each p(X_w^i)
x = np.arange(-5., 5.1, 0.1)
mise = np.zeros(X_w.shape[1])
for i_bin in range(X_w.shape[1]):
mise[i_bin] = NG.MISE(X_w[:,i_bin], b=b)
# plot the most discrepant components.
prettyplot()
fig = plt.figure()
sub = fig.add_subplot(111)
i_sort = np.argsort(mise)
print 'outlier bins = ', i_sort[-5:]
print 'mise = ', mise[i_sort[-5:]]
nbin = int(10./b)
for i_bin in i_sort[-10:]:
hb_Xi, Xi_edges = np.histogram(X_w[:,i_bin], bins=nbin, range=[-5., 5.], normed=True)
p_X_w_arr = UT.bar_plot(Xi_edges, hb_Xi)
sub.plot(p_X_w_arr[0], p_X_w_arr[1])
sub.plot(x, UT.gauss(x, 1., 0.), c='k', lw=3, label='$\mathcal{N}(0,1)$')
sub.set_xlim([-2.5, 2.5])
sub.set_xlabel('$\mathtt{X^{i}_{W}}$', fontsize=25)
sub.set_ylim([0., 0.6])
sub.set_ylabel('$\mathtt{P(X^{i}_{W})}$', fontsize=25)
sub.legend(loc='upper right')
str_rebin = ''
if rebin is not None:
str_rebin = '.rebin'+str(rebin)
f = ''.join([UT.fig_dir(), 'tests/test.p_Xw_i_outlier.', method, '.', mock, '.ell', str(ell),
str_rebin, '.b', str(b), '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def p_Xw_i_outlier(mock, ell=0, rebin=None, krange=None, method='choletsky'):
''' Examine the pdf of X_w^i components that deviate significantly from
N(0,1)
'''
Pk = NG.dataX(mock, ell=ell, rebin=rebin, krange=krange)
X, _ = NG.meansub(Pk)
X_w, W = NG.whiten(X, method=method) # whitened data
# calculate the chi-squared values of each p(X_w^i)
x = np.arange(-5., 5.1, 0.1)
chi2 = np.zeros(X_w.shape[1])
for i_bin in range(X_w.shape[1]):
kern = gkde(X_w[:,i_bin]) # gaussian KDE kernel using "rule of thumb" scott's rule.
chi2[i_bin] = np.sum((UT.gauss(x, 1., 0.) - kern.evaluate(x))**2)/np.float(len(x))
# plot the most discrepant components.
prettyplot()
fig = plt.figure()
sub = fig.add_subplot(111)
i_sort = np.argsort(chi2)
print 'outlier bins = ', i_sort[-5:]
for i_bin in i_sort[-10:]:
kern = gkde(X_w[:,i_bin]) # gaussian KDE kernel using "rule of thumb" scott's rule.
sub.plot(x, kern.evaluate(x))
sub.plot(x, UT.gauss(x, 1., 0.), c='k', lw=3, label='$\mathcal{N}(0,1)$')
sub.set_xlim([-2.5, 2.5])
sub.set_xlabel('$\mathtt{X^{i}_{W}}$', fontsize=25)
sub.set_ylim([0., 0.6])
sub.set_ylabel('$\mathtt{P(X^{i}_{W})}$', fontsize=25)
sub.legend(loc='upper right')
if rebin is None:
f = ''.join([UT.fig_dir(), 'tests/test.p_Xw_i_outlier.', method, '.', mock, '.ell', str(ell), '.png'])
else:
f = ''.join([UT.fig_dir(), 'tests/test.p_Xw_i_outlier.', method, '.', mock, '.ell', str(ell), '.rebin', str(rebin), '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def p_Xw_i(mock, ell=0, rebin=None, krange=None, ica=False, pca=False):
''' Test the probability distribution function of each X_w^i
component -- p(X_w^i). First compare the histograms of p(X_w^i)
with N(0,1). Then compare the gaussian KDE of p(X_w^i).
'''
Pk = NG.dataX(mock, ell=ell, rebin=rebin, krange=krange)
X, _ = NG.meansub(Pk)
str_w = 'W'
if ica and pca:
raise ValueError
if ica: # ICA components
# ICA components do not need to be Gaussian.
# in fact the whole point of the ICA transform
# is to capture the non-Gaussianity...
X_white, _ = NG.whiten(X) # whitened data
X_w, _ = NG.Ica(X_white)
str_w = 'ICA'
if pca: # PCA components
X_w, _ = NG.whiten(X, method='pca') # whitened data
str_w = 'PCA'
if not ica and not pca: # just whitened
X_w, W = NG.whiten(X) # whitened data
# p(X_w^i) histograms
fig = plt.figure(figsize=(15,7))
sub = fig.add_subplot(121)
for i_bin in range(X_w.shape[1]):
p_X_w, edges = np.histogram(X_w[:,i_bin], normed=True)
p_X_w_arr = UT.bar_plot(edges, p_X_w)
sub.plot(p_X_w_arr[0], p_X_w_arr[1])
x = np.arange(-5., 5.1, 0.1)
sub.plot(x, UT.gauss(x, 1., 0.), c='k', lw=3, label='$\mathcal{N}(0,1)$')
sub.set_xlim([-2.5, 2.5])
sub.set_xlabel('$\mathtt{X_{'+str_w+'}}$', fontsize=25)
sub.set_ylim([0., 0.6])
sub.set_ylabel('$\mathtt{P(X_{'+str_w+'})}$', fontsize=25)
sub.legend(loc='upper right')
# p(X_w^i) gaussian KDE fits
pdfs = NG.p_Xw_i(X_w, range(X_w.shape[1]), x=x)
sub = fig.add_subplot(122)
for i_bin in range(X_w.shape[1]):
sub.plot(x, pdfs[i_bin])
sub.plot(x, UT.gauss(x, 1., 0.), c='k', lw=3, label='$\mathcal{N}(0,1)$')
sub.set_xlim([-2.5, 2.5])
sub.set_xlabel('$\mathtt{X_{W}}$', fontsize=25)
sub.set_ylim([0., 0.6])
sub.set_ylabel('$\mathtt{P(X_{W})}$', fontsize=25)
sub.legend(loc='upper right')
str_ica, str_pca = '', ''
if ica:
str_ica = '.ICA'
if pca:
str_pca = '.PCA'
if rebin is None:
f = ''.join([UT.fig_dir(), 'tests/test.p_Xw_i', str_pca, str_ica, '.', mock, '.ell', str(ell), '.png'])
else:
f = ''.join([UT.fig_dir(), 'tests/test.p_Xw_i', str_pca, str_ica, '.', mock, '.ell', str(ell), '.rebin', str(rebin), '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def p_Xwi_Xwj_outlier(mock, ell=0, rebin=None, krange=None, ica=False, pca=False):
''' Compare the joint pdfs of whitened X components (i.e. X_w^i, X_w^j)
p(X_w^i, X_w^j) to p(X_w^i) p(X_w^j) in order to test the independence
argument.
'''
Pk = NG.dataX(mock, ell=ell, rebin=rebin, krange=krange)
X, _ = NG.meansub(Pk)
if ica and pca:
raise ValueError
if ica: # ICA components
X_white, _ = NG.whiten(X) # whitened data
X_w, _ = NG.Ica(X_white)
if pca: # PCA components
X_w, _ = NG.whiten(X, method='pca') # whitened data
if not ica and not pca: # just whitened
X_w, _ = NG.whiten(X, method='choletsky') # whitened data
x, y = np.linspace(-5., 5., 50), np.linspace(-5., 5., 50)
xx, yy = np.meshgrid(x,y)
pos = np.vstack([xx.ravel(), yy.ravel()])
ij_i, ij_j = np.meshgrid(range(X_w.shape[1]), range(X_w.shape[1]))
ij = np.vstack([ij_i.ravel(), ij_j.ravel()])
# joint pdfs of X_w^i and X_w^j estimated from mocks
# i.e. p(X_w^i, X_w^j)
pdfs_2d = NG.p_Xwi_Xwj(X_w, ij, x=x, y=y)
# p(X_w^i) * p(X_w^j) estimated from mocks
pXwi = NG.p_Xw_i(X_w, range(X_w.shape[1]), x=x)
pXwj = pXwi
# calculate L2 norm difference betwen joint pdf and 2d gaussian
chi2 = np.zeros(len(pdfs_2d))
for i in range(len(pdfs_2d)):
if not isinstance(pdfs_2d[i], float):
pXwipXwj = np.dot(pXwi[ij[0,i]][:,None], pXwj[ij[1,i]][None,:]).T.flatten()
chi2[i] = np.sum((pXwipXwj - pdfs_2d[i])**2)
# ij values with the highest chi-squared
ii_out = np.argsort(chi2)[-10:]
inc = np.where(ij[0,ii_out] > ij[1,ii_out])
prettyplot()
fig = plt.figure(figsize=(len(inc[0])*10, 8))
for ii, i_sort_i in enumerate(ii_out[inc]):
sub = fig.add_subplot(1, len(inc[0]), ii+1)
# plot p(X_w^i) * p(X_w^j)
pXwipXwj = np.dot(pXwi[ij[0,i_sort_i]][:,None], pXwj[ij[1,i_sort_i]][None,:]).T
sub.contourf(xx, yy, pXwipXwj, cmap='gray_r', levels=[0.05, 0.1, 0.15, 0.2])
# p(X_w^i, X_w^j)
Z = np.reshape(pdfs_2d[i_sort_i], xx.shape)
cs = sub.contour(xx, yy, Z, colors='k', linestyles='dashed', levels=[0.05, 0.1, 0.15, 0.2])
cs.collections[0].set_label('$\mathtt{p(X_w^i, X_w^j)}$')
sub.set_xlim([-3., 3.])
sub.set_xlabel('$\mathtt{X_w^{i='+str(ij[0,i_sort_i])+'}}$', fontsize=25)
sub.set_ylim([-3., 3.])
sub.set_ylabel('$\mathtt{X_w^{j='+str(ij[1,i_sort_i])+'}}$', fontsize=25)
if ii == 0:
sub.legend(loc='upper right', prop={'size':25})
else:
sub.set_yticklabels([])
str_ica, str_pca = '', ''
if ica:
str_ica = '.ICA'
if pca:
str_pca = '.PCA'
if rebin is None:
f = ''.join([UT.fig_dir(), 'tests/test.p_Xwi_Xwj_outlier', str_ica, str_pca, '.', mock, '.ell', str(ell), '.png'])
else:
f = ''.join([UT.fig_dir(), 'tests/test.p_Xwi_Xwj_outlier', str_ica, str_pca, '.', mock, '.ell', str(ell), '.rebin', str(rebin), '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def GMF_p_Xw_i(ica=False, pca=False):
''' Test the probability distribution function of each transformed X
component -- p(X^i). First compare the histograms of p(X_w^i)
with N(0,1). Then compare the gaussian KDE of p(X_w^i).
'''
gmf = NG.X_gmf_all() # import all the GMF mocks
X, _ = NG.meansub(gmf)
str_w = 'W'
if ica and pca: raise ValueError
if ica: # ICA components
# ICA components do not need to be Gaussian.
# in fact the whole point of the ICA transform
# is to capture the non-Gaussianity...
X_white, _ = NG.whiten(X) # whitened data
X_w, _ = NG.Ica(X_white)
str_w = 'ICA'
if pca: # PCA components
X_w, _ = NG.whiten(X, method='pca') # whitened data
str_w = 'PCA'
if not ica and not pca: # just whitened
X_w, W = NG.whiten(X) # whitened data
# p(X_w^i) histograms
fig = plt.figure(figsize=(5*gmf.shape[1],4))
for icomp in range(gmf.shape[1]):
sub = fig.add_subplot(1, gmf.shape[1], icomp+1)
# histogram of X_w^i s
hh = np.histogram(X_w[:,icomp], normed=True, bins=50, range=[-5., 5.])
p_X_w_arr = UT.bar_plot(*hh)
sub.fill_between(p_X_w_arr[0], np.zeros(len(p_X_w_arr[1])), p_X_w_arr[1],
color='k', alpha=0.25)
x = np.linspace(-5., 5., 100)
sub.plot(x, UT.gauss(x, 1., 0.), c='k', lw=2, ls=':', label='$\mathcal{N}(0,1)$')
# p(X_w^i) gaussian KDE fits
t_start = time.time()
pdf = NG.p_Xw_i(X_w, icomp, x=x, method='gkde')
sub.plot(x, pdf, lw=2.5, label='Gaussian KDE')
print 'scipy Gaussian KDE ', time.time()-t_start
# p(X_w^i) SKlearn KDE fits
t_start = time.time()
pdf = NG.p_Xw_i(X_w, icomp, x=x, method='sk_kde')
sub.plot(x, pdf, lw=2.5, label='SKlearn KDE')
print 'SKlearn CV best-fit KDE ', time.time()-t_start
# p(X_w^i) statsmodels KDE fits
t_start = time.time()
pdf = NG.p_Xw_i(X_w, icomp, x=x, method='sm_kde')
sub.plot(x, pdf, lw=2.5, label='StatsModels KDE')
print 'Stats Models KDE ', time.time()-t_start
# p(X_w^i) GMM fits
pdf = NG.p_Xw_i(X_w, icomp, x=x, method='gmm', n_comp_max=20)
sub.plot(x, pdf, lw=2.5, ls='--', label='GMM')
sub.set_xlim([-3., 3.])
sub.set_xlabel('$X_{'+str_w+'}^{('+str(icomp)+')}$', fontsize=25)
sub.set_ylim([0., 0.6])
if icomp == 0:
sub.set_ylabel('$P(X_{'+str_w+'})$', fontsize=25)
sub.legend(loc='upper left', prop={'size': 15})
str_ica, str_pca = '', ''
if ica: str_ica = '.ICA'
if pca: str_pca = '.PCA'
f = ''.join([UT.fig_dir(), 'tests/test.GMF_p_Xw_i', str_pca, str_ica, '.png'])
fig.savefig(f, bbox_inches='tight')
return None
def divGMF(div_func='kl', Nref=1000, K=5, n_mc=10, n_comp_max=10, n_mocks=2000):
''' compare the divergence estimates between
D( gauss(C_gmf) || gauss(C_gmf) ), D( gmfs || gauss(C_gmf) ),
D( gmfs || p(gmfs) KDE), D( gmfs || p(gmfs) GMM),
D( gmfs || PI p(gmfs^i_ICA) KDE), and D( gmfs || PI p(gmfs^i_ICA) GMM)
'''
if isinstance(Nref, float):
Nref = int(Nref)
# read in mock GMFs from all HOD realizations (20,000 mocks)
gmfs_mock = NG.X_gmf_all()[:n_mocks]
n_mock = gmfs_mock.shape[0] # number of mocks
print("%i mocks" % n_mock)
gmfs_mock_meansub, _ = NG.meansub(gmfs_mock) # mean subtract
X_w, W = NG.whiten(gmfs_mock_meansub)
X_ica, _ = NG.Ica(X_w) # ICA transformation
C_gmf = np.cov(X_w.T) # covariance matrix
# p(gmfs) GMM
gmms, bics = [], []
for i_comp in range(1,n_comp_max+1):
gmm = GMix(n_components=i_comp)
gmm.fit(X_w)
gmms.append(gmm)
bics.append(gmm.bic(X_w))
ibest = np.array(bics).argmin()
kern_gmm = gmms[ibest]
# p(gmfs) KDE
t0 = time.time()
grid = GridSearchCV(skKDE(),
{'bandwidth': np.linspace(0.1, 1.0, 30)},
cv=10) # 10-fold cross-validation
grid.fit(X_w)
kern_kde = grid.best_estimator_
dt = time.time() - t0
print('%f sec' % dt)
# PI p(gmfs^i_ICA) GMM
kern_gmm_ica = []
for ibin in range(X_ica.shape[1]):
gmms, bics = [], []
for i_comp in range(1,n_comp_max+1):
gmm = GMix(n_components=i_comp)
gmm.fit(X_ica[:,ibin][:,None])
gmms.append(gmm)
bics.append(gmm.bic(X_ica[:,ibin][:,None]))
ibest = np.array(bics).argmin()
kern_gmm_ica.append(gmms[ibest])
# PI p(gmfs^i_ICA) KDE
kern_kde_ica = []
for ibin in range(X_ica.shape[1]):
t0 = time.time()
grid = GridSearchCV(skKDE(),
{'bandwidth': np.linspace(0.1, 1.0, 30)},
cv=10) # 10-fold cross-validation
grid.fit(X_ica[:,ibin][:,None])
kern_kde_ica.append(grid.best_estimator_)
dt = time.time() - t0
print('%f sec' % dt)
# caluclate the divergences now
div_gauss_ref, div_gauss = [], []
div_gmm, div_gmm_ica = [], []
div_kde, div_kde_ica = [], []
for i in range(n_mc):
print('%i montecarlo' % i)
t_start = time.time()
# reference divergence in order to showcase the estimator's scatter
# Gaussian distribution described by C_gmf with same n_mock mocks
gauss = mvn(np.zeros(gmfs_mock.shape[1]), C_gmf, size=n_mock)
div_gauss_ref_i = NG.kNNdiv_gauss(gauss, C_gmf, Knn=K, div_func=div_func, Nref=Nref)
div_gauss_ref.append(div_gauss_ref_i)
# estimate divergence between gmfs_white and a
# Gaussian distribution described by C_gmf
div_gauss_i = NG.kNNdiv_gauss(X_w, C_gmf, Knn=K, div_func=div_func, Nref=Nref)
div_gauss.append(div_gauss_i)
# D( gmfs || p(gmfs) GMM)
div_gmm_i = NG.kNNdiv_Kernel(X_w, kern_gmm, Knn=K, div_func=div_func,
Nref=Nref, compwise=False)
div_gmm.append(div_gmm_i)
# D( gmfs || p(gmfs) KDE)
div_kde_i = NG.kNNdiv_Kernel(X_w, kern_kde, Knn=K, div_func=div_func,
Nref=Nref, compwise=False)
div_kde.append(div_kde_i)
# D( gmfs || PI p(gmfs^i_ICA) GMM),
div_gmm_ica_i = NG.kNNdiv_Kernel(X_ica, kern_gmm_ica, Knn=K, div_func=div_func,
Nref=Nref, compwise=True)
div_gmm_ica.append(div_gmm_ica_i)
# D( gmfs || PI p(gmfs^i_ICA) KDE),
div_kde_ica_i = NG.kNNdiv_Kernel(X_ica, kern_kde_ica, Knn=K, div_func=div_func,
Nref=Nref, compwise=True)
div_kde_ica.append(div_kde_ica_i)
print('t= %f sec' % round(time.time()-t_start,2))
fig = plt.figure(figsize=(10,5))
sub = fig.add_subplot(111)
hrange = [-0.15, 0.6]
nbins = 50
divs = [div_gauss_ref, div_gauss, div_gmm, div_kde, div_gmm_ica, div_kde_ica]
labels = ['Ref.', r'$D(\{\zeta_i^{(m)}\}\parallel \mathcal{N}({\bf C}^{(m)}))$',
r'$D(\{\zeta^{(m)}\}\parallel p_\mathrm{GMM}(\{\zeta^{m}\}))$',
r'$D(\{\zeta^{(m)}\}\parallel p_\mathrm{KDE}(\{\zeta^{m}\}))$',
r'$D(\{\zeta_\mathrm{ICA}^{(m)}\}\parallel \prod_{i} p^\mathrm{GMM}(\{\zeta_{i, \mathrm{ICA}}^{m}\}))$',
r'$D(\{\zeta_\mathrm{ICA}^{(m)}\}\parallel \prod_{i} p^\mathrm{KDE}(\{\zeta_{i, \mathrm{ICA}}^{m}\}))$']
y_max = 0.
for div, lbl in zip(divs, labels):
hh = np.histogram(np.array(div), normed=True, range=hrange, bins=nbins)
bp = UT.bar_plot(*hh)
sub.fill_between(bp[0], np.zeros(len(bp[0])), bp[1], edgecolor='none',
alpha=0.5, label=lbl)
y_max = max(y_max, bp[1].max())
if (np.average(div) < hrange[0]) or (np.average(div) > hrange[1]):
print('divergence of %s (%f) is outside range' % (lbl, np.average(div)))
sub.set_xlim(hrange)
sub.set_ylim([0., y_max*1.2])
sub.legend(loc='upper left', prop={'size': 15})
# xlabels
if 'renyi' in div_func:
alpha = float(div_func.split(':')[-1])
sub.set_xlabel(r'Renyi-$\alpha='+str(alpha)+'$ divergence', fontsize=20)
elif 'kl' in div_func:
sub.set_xlabel(r'KL divergence', fontsize=20)
if 'renyi' in div_func: str_div = 'renyi'+str(alpha)
elif div_func == 'kl': str_div = 'kl'
f_fig = ''.join([UT.fig_dir(), 'tests/kNN_divergence.gmf.K', str(K), '.', str(n_mocks),
'.', str_div, '.png'])
fig.savefig(f_fig, bbox_inches='tight')
return None
def diverge(obvs,
diver,
div_func='kl',
Nref=1000,
K=5,
n_mc=10,
n_comp_max=10,
n_mocks=20000,
pk_mock='patchy.z1',
NorS='ngc',
njobs=1):
''' calculate the divergences:
- D( gauss(C_X) || gauss(C_X) )
- D( mock X || gauss(C_X))
- D( mock X || p(X) KDE)
- D( mock X || p(X) GMM)
- D( mock X || PI p(X^i_ICA) KDE)
- D( mock X || PI p(X^i_ICA) GMM)
'''
if isinstance(Nref, float): Nref = int(Nref)
if diver not in [
'ref', 'pX_gauss', 'pX_gauss_hartlap', 'pX_GMM', 'pX_GMM_ref',
'pX_KDE', 'pX_KDE_ref', 'pX_scottKDE', 'pX_scottKDE_ref',
'pXi_ICA_GMM', 'pXi_ICA_GMM_ref', 'pXi_parICA_GMM',
'pXi_parICA_GMM_ref', 'pXi_ICA_KDE', 'pXi_ICA_KDE_ref',
'pXi_parICA_KDE', 'pXi_parICA_KDE_ref', 'pXi_ICA_scottKDE',
'pXi_ICA_scottKDE_ref', 'pXi_parICA_scottKDE',
'pXi_parICA_scottKDE_ref'
]:
raise ValueError
str_obvs = ''
if obvs == 'pk': str_obvs = '.' + NorS
if 'renyi' in div_func:
alpha = float(div_func.split(':')[-1])
str_div = 'renyi' + str(alpha)
elif div_func == 'kl':
str_div = 'kl'
str_comp = ''
if 'GMM' in diver: str_comp = '.ncomp' + str(n_comp_max)
f_dat = ''.join([
UT.dat_dir(), 'diverg/', 'diverg.', obvs, str_obvs, '.', diver, '.K',
str(K), str_comp, '.Nref',
str(Nref), '.', str_div, '.dat'
])
if not os.path.isfile(f_dat):
print('-- writing to -- \n %s' % f_dat)
f_out = open(f_dat, 'w')
else:
print('-- appending to -- \n %s' % f_dat)
# read in mock data X
if obvs == 'pk':
X_mock = NG.X_pk_all(pk_mock, NorS=NorS, sys='fc')
elif obvs == 'gmf':
if n_mocks is not None:
X_mock = NG.X_gmf_all()[:n_mocks]
else:
X_mock = NG.X_gmf_all()
else:
raise ValueError("obvs = 'pk' or 'gmf'")
n_mock = X_mock.shape[0] # number of mocks
print("%i mocks" % n_mock)
X_mock_meansub, _ = NG.meansub(X_mock) # mean subtract
X_w, W = NG.whiten(X_mock_meansub)
if '_ICA' in diver:
X_ica, W_ica = NG.Ica(X_w) # ICA transformation
W_ica_inv = sp.linalg.pinv(W_ica.T)
elif '_parICA' in diver:
# FastICA transformation using parallel algorithm
X_ica, W_ica = NG.Ica(X_w, algorithm='parallel')
W_ica_inv = sp.linalg.pinv(W_ica.T)
if diver in ['pX_gauss', 'ref']:
C_X = np.cov(X_w.T) # covariance matrix
elif diver in ['pX_gauss_hartlap']:
C_X = np.cov(X_w.T) # covariance matrix
f_hartlap = (n_mock - float(X_mock.shape[1]) - 2.) / (n_mock - 1.)
print("hartlap factor = %f" % f_hartlap)
C_X = C_X / f_hartlap # scale covariance matrix by hartlap factor
elif diver in ['pX_GMM', 'pX_GMM_ref']: # p(mock X) GMM
gmms, bics = [], []
for i_comp in range(1, n_comp_max + 1):
gmm = GMix(n_components=i_comp)
gmm.fit(X_w)
gmms.append(gmm)
bics.append(gmm.bic(X_w))
ibest = np.array(bics).argmin()
kern_gmm = gmms[ibest]
elif diver in ['pX_KDE', 'pX_KDE_ref']: # p(mock X) KDE
t0 = time.time()
grid = GridSearchCV(skKDE(), {'bandwidth': np.linspace(0.1, 1.0, 30)},
cv=10,
n_jobs=njobs) # 10-fold cross-validation
grid.fit(X_w)
kern_kde = grid.best_estimator_
dt = time.time() - t0
print('%f sec' % dt)
elif diver in ['pX_scottKDE', 'pX_scottKDE_ref']: # p(mock X) KDE
# calculate Scott's Rule KDE
t0 = time.time()
kern_kde = UT.KayDE(X_w)
dt = time.time() - t0
print('%f sec' % dt)
elif diver in [
'pXi_ICA_GMM', 'pXi_ICA_GMM_ref', 'pXi_parICA_GMM',
'pXi_parICA_GMM_ref'
]:
# PI p(X^i_ICA) GMM
kern_gmm_ica = []
for ibin in range(X_ica.shape[1]):
gmms, bics = [], []
for i_comp in range(1, n_comp_max + 1):
gmm = GMix(n_components=i_comp)
gmm.fit(X_ica[:, ibin][:, None])
gmms.append(gmm)
bics.append(gmm.bic(X_ica[:, ibin][:, None]))
ibest = np.array(bics).argmin()
kern_gmm_ica.append(gmms[ibest])
elif diver in [
'pXi_ICA_KDE', 'pXi_ICA_KDE_ref', 'pXi_parICA_KDE',
'pXi_parICA_KDE_ref'
]:
# PI p(X^i_ICA) KDE
kern_kde_ica = []
for ibin in range(X_ica.shape[1]):
t0 = time.time()
grid = GridSearchCV(skKDE(),
{'bandwidth': np.linspace(0.1, 1.0, 30)},
cv=10,
n_jobs=njobs) # 10-fold cross-validation
grid.fit(X_ica[:, ibin][:, None])
kern_kde_ica.append(grid.best_estimator_)
dt = time.time() - t0
print('%f sec' % dt)
elif diver in [
'pXi_ICA_scottKDE', 'pXi_ICA_scottKDE_ref', 'pXi_parICA_scottKDE',
'pXi_parICA_scottKDE_ref'
]:
# PI p(X^i_ICA) KDE
kern_kde_ica = []
for ibin in range(X_ica.shape[1]):
kern_kde_i = UT.KayDE(X_ica[:, ibin])
kern_kde_ica.append(kern_kde_i)
# caluclate the divergences now
divs = []
for i in range(n_mc):
print('%i montecarlo' % i)
t0 = time.time()
if diver in ['pX_gauss', 'pX_gauss_hartlap']:
# estimate divergence between gmfs_white and a
# Gaussian distribution described by C_gmf
div_i = NG.kNNdiv_gauss(X_w,
C_X,
Knn=K,
div_func=div_func,
Nref=Nref,
njobs=njobs)
elif diver == 'ref':
# reference divergence in order to showcase the estimator's scatter
# Gaussian distribution described by C_gmf with same n_mock mocks
gauss = mvn(np.zeros(X_mock.shape[1]), C_X, size=n_mock)
div_i = NG.kNNdiv_gauss(gauss,
C_X,
Knn=K,
div_func=div_func,
Nref=Nref,
njobs=njobs)
elif diver == 'pX_GMM': # D( mock X || p(X) GMM)
div_i = NG.kNNdiv_Kernel(X_w,
kern_gmm,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=False,
njobs=njobs)
elif diver == 'pX_GMM_ref': # D( sample from p(X) GMM || p(X) GMM)
samp = kern_gmm.sample(n_mock)
div_i = NG.kNNdiv_Kernel(samp[0],
kern_gmm,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=False,
njobs=njobs)
elif diver in ['pX_KDE', 'pX_scottKDE']: # D( mock X || p(X) KDE)
div_i = NG.kNNdiv_Kernel(X_w,
kern_kde,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=False,
njobs=njobs)
divs.append(div_i)
elif diver in ['pX_KDE_ref', 'pX_scottKDE_ref'
]: # D( sample from p(X) KDE || p(X) KDE)
samp = kern_kde.sample(n_mock)
div_i = NG.kNNdiv_Kernel(samp,
kern_kde,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=False,
njobs=njobs)
divs.append(div_i)
elif diver in ['pXi_ICA_GMM',
'pXi_parICA_GMM']: # D( mock X || PI p(X^i_ICA) GMM),
div_i = NG.kNNdiv_Kernel(X_w,
kern_gmm_ica,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=True,
njobs=njobs,
W_ica_inv=W_ica_inv)
elif diver in ['pXi_ICA_GMM_ref', 'pXi_parICA_GMM_ref']:
# D( ref. sample || PI p(X^i_ICA) GMM),
samp = np.zeros((n_mock, X_ica.shape[1]))
for icomp in range(X_ica.shape[1]):
samp_i = kern_gmm_ica[icomp].sample(n_mock)
samp[:, icomp] = samp_i[0].flatten()
samp = np.dot(samp, W_ica_inv.T)
div_i = NG.kNNdiv_Kernel(samp,
kern_gmm_ica,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=True,
njobs=njobs,
W_ica_inv=W_ica_inv)
elif diver in [
'pXi_ICA_KDE', 'pXi_ICA_scottKDE', 'pXi_parICA_KDE',
'pXi_parICA_scottKDE'
]: # D( mock X || PI p(X^i_ICA) KDE),
div_i = NG.kNNdiv_Kernel(X_w,
kern_kde_ica,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=True,
njobs=njobs,
W_ica_inv=W_ica_inv)
elif diver in [
'pXi_ICA_KDE_ref', 'pXi_ICA_scottKDE_ref',
'pXi_parICA_KDE_ref', 'pXi_parICA_scottKDE_ref'
]:
# D( ref sample || PI p(X^i_ICA) KDE),
samp = np.zeros((n_mock, X_ica.shape[1]))
for icomp in range(X_ica.shape[1]):
samp_i = kern_kde_ica[icomp].sample(n_mock)
samp[:, icomp] = samp_i.flatten()
samp = np.dot(samp, W_ica_inv.T)
div_i = NG.kNNdiv_Kernel(samp,
kern_kde_ica,
Knn=K,
div_func=div_func,
Nref=Nref,
compwise=True,
njobs=njobs,
W_ica_inv=W_ica_inv)
print(div_i)
f_out = open(f_dat, 'a')
f_out.write('%f \n' % div_i)
f_out.close()
return None