def template_ADMM_MGL(p=100, K=5, N=1000, reg='GGL', latent=False):
"""
template for test for ADMM MGL solver with conforming variables
"""
M = 10 # M should be divisor of p
if reg == 'GGL':
Sigma, Theta = group_power_network(p, K, M)
elif reg == 'FGL':
Sigma, Theta = time_varying_power_network(p, K, M)
S, samples = sample_covariance_matrix(Sigma, N)
lambda1 = 0.05
lambda2 = 0.01
Omega_0 = get_K_identity(K, p)
sol, info = ADMM_MGL(S,
lambda1,
lambda2,
reg,
Omega_0,
tol=1e-5,
rtol=1e-5,
verbose=True,
measure=True,
latent=latent,
mu1=0.01)
assert info['status'] == 'optimal'
_, info2 = ADMM_MGL(S,
lambda1,
lambda2,
reg,
Omega_0,
tol=1e-20,
rtol=1e-20,
max_iter=2,
latent=latent,
mu1=0.01)
assert info2['status'] == 'max iterations reached'
return
def template_admm_vs_ppdna(p=50, K=3, N=1000, reg="GGL"):
M = 5 # M should be divisor of p
if reg == 'GGL':
Sigma, Theta = group_power_network(p, K, M)
elif reg == 'FGL':
Sigma, Theta = time_varying_power_network(p, K, M)
S, samples = sample_covariance_matrix(Sigma, N)
lambda1 = 0.05
lambda2 = 0.01
Omega_0 = get_K_identity(K, p)
sol, info = ADMM_MGL(S,
lambda1,
lambda2,
reg,
Omega_0,
stopping_criterion='kkt',
tol=1e-6,
rtol=1e-5,
verbose=True,
latent=False)
sol2, info2 = warmPPDNA(S,
lambda1,
lambda2,
reg,
Omega_0,
eps=1e-6,
verbose=False,
measure=True)
sol3, info3 = PPDNA(S,
lambda1,
lambda2,
reg,
Omega_0,
eps_ppdna=1e-6,
verbose=True,
measure=True)
assert_array_almost_equal(sol['Theta'], sol2['Theta'], 2)
assert_array_almost_equal(sol2['Theta'], sol3['Theta'], 2)
return
solP, infoP = warmPPDNA(S,
l1opt,
l2opt,
reg,
Omega_0,
Theta_0=Theta_0,
eps=5e-5,
verbose=True,
measure=True)
solA, infoA = ADMM_MGL(S,
l1opt,
l2opt,
reg,
Omega_0,
Theta_0=Theta_0,
eps_admm=5e-5,
verbose=True,
measure=True)
Theta_sol = solP['Theta']
Omega_sol = solP['Omega']
with sns.axes_style("white"):
fig, axs = plt.subplots(nrows=1, ncols=2, figsize=(10, 3))
draw_group_heatmap(Theta, method='truth', ax=axs[0])
draw_group_heatmap(Theta_sol, method='PPDNA', ax=axs[1])
#%%
# plotting (set save = False if plot should not be saved)
RT_PPA = np.zeros(len(vecN))
TPR = np.zeros(len(vecN))
FPR = np.zeros(len(vecN))
iA = {}
iP = {}
for j in np.arange(len(vecN)):
S, sample = sample_covariance_matrix(Sigma, vecN[j])
#start = time()
solA, infoA = ADMM_MGL(S,
l1[j],
l2[j],
reg,
Omega_0,
eps_admm=5e-5,
verbose=False,
measure=True)
#end = time()
#RT_ADMM[j] = end-start
iA[j] = infoA
TPR[j] = discovery_rate(solA['Theta'], Theta)['TPR']
FPR[j] = discovery_rate(solA['Theta'], Theta)['FPR']
#start = time()
solP, infoP = warmPPDNA(S,
l1[j],
l2[j],
reg,
l2 = 2e-2 * np.ones(len(f))
Omega_0 = get_K_identity(K,p)
TPR = np.zeros(len(vecN))
FPR = np.zeros(len(vecN))
iA = {}
iP = {}
for j in np.arange(len(vecN)):
S, sample = sample_covariance_matrix(Sigma, vecN[j])
solA, infoA = ADMM_MGL(S, l1[j], l2[j], reg , Omega_0 , tol = 5e-5, stopping_criterion = 'kkt', verbose = False, measure = True)
iA[j] = infoA
TPR[j] = discovery_rate(solA['Theta'], Theta)['TPR']
FPR[j] = discovery_rate(solA['Theta'], Theta)['FPR']
solP, infoP = warmPPDNA(S, l1[j], l2[j], reg, Omega_0, eps = 5e-5, eps_admm = 1e-2, verbose = True, measure = True)
iP[j] = infoP
#%%
# plotting
save = False
plot_runtime(iA, iP, vecN, save = save)
solPPDNA, info = warmPPDNA(S,
lambda1,
lambda2,
reg,
Omega_0,
eps=1e-5,
verbose=True,
measure=True)
#solADMM, info = ADMM_MGL(S, lambda1, lambda2, reg, Omega_0, n_samples = None, eps_admm = 1e-4 , verbose = True)
solADMM, info = ADMM_MGL(S,
lambda1,
lambda2,
reg,
Omega_0,
n_samples=None,
eps_admm=1e-4,
verbose=True,
latent=True,
mu1=.2)
#%%
# tests for the extended ADMM version
from gglasso.solver.ext_admm_solver import ext_ADMM_MGL
from gglasso.helper.ext_admm_helper import construct_trivial_G
Sdict = dict()
Omega_0 = dict()
for k in np.arange(K):
AIC = np.zeros((grid1, grid2))
BIC = np.zeros((grid1, grid2))
Omega_0 = np.zeros((K, p, p))
Theta_0 = np.zeros((K, p, p))
for g1 in np.arange(grid1):
for g2 in np.arange(grid2):
lambda1 = L1[g1, g2]
lambda2 = L2[g1, g2]
sol, info = ADMM_MGL(S,
lambda1,
lambda2,
reg,
Omega_0,
eps_admm=1e-3,
verbose=False)
Theta_sol = sol['Theta']
Omega_sol = sol['Omega']
# warm start
Omega_0 = Omega_sol.copy()
Theta_0 = Theta_sol.copy()
AIC[g1, g2] = aic(S, Theta_sol, n.mean())
BIC[g1, g2] = ebic(S, Theta_sol, n.mean(), gamma=0.1)
ix = np.unravel_index(np.nanargmin(BIC), BIC.shape)
ix2 = np.unravel_index(np.nanargmin(AIC), AIC.shape)
def template_extADMM_consistent(latent=False):
"""
tests whether the extended ADMM solver results in the same as MGL-solver for the redundant case of conforming variables
"""
p = 50
N = 1000
K = 5
M = 10
lambda1 = 0.05
lambda2 = 0.01
Sigma, Theta = group_power_network(p, K, M)
S, samples = sample_covariance_matrix(Sigma, N)
Sdict = dict()
Omega_0 = dict()
for k in np.arange(K):
Sdict[k] = S[k, :, :].copy()
Omega_0[k] = np.eye(p)
# constructs the "trivial" groups, i.e. all variables present in all instances
G = construct_trivial_G(p, K)
solext, _ = ext_ADMM_MGL(Sdict,
lambda1,
lambda2 / np.sqrt(K),
'GGL',
Omega_0,
G,
tol=1e-9,
rtol=1e-9,
verbose=True,
latent=latent,
mu1=0.01)
solext2, _ = ext_ADMM_MGL(Sdict,
lambda1,
lambda2 / np.sqrt(K),
'GGL',
Omega_0,
G,
stopping_criterion='kkt',
tol=1e-8,
verbose=True,
latent=latent,
mu1=0.01)
Omega_0_arr = get_K_identity(K, p)
solADMM, info = ADMM_MGL(S,
lambda1,
lambda2,
'GGL',
Omega_0_arr,
tol=1e-9,
rtol=1e-9,
verbose=False,
latent=latent,
mu1=0.01)
for k in np.arange(K):
assert_array_almost_equal(solext['Theta'][k],
solADMM['Theta'][k, :, :], 2)
assert_array_almost_equal(solext2['Theta'][k],
solADMM['Theta'][k, :, :], 2)
if latent:
for k in np.arange(K):
assert_array_almost_equal(solext['L'][k], solADMM['L'][k, :, :], 2)
assert_array_almost_equal(solext2['L'][k], solADMM['L'][k, :, :],
2)
return
# solve with PPDNA
Omega_0 = results.get('GLASSO').get('Theta')
Theta_0 = Omega_0.copy()
X_0 = np.zeros((K,p,p))
start = time()
sol, info = warmPPDNA(S, lambda1, lambda2, reg, Omega_0, Theta_0 = Theta_0, X_0 = X_0, eps = 5e-4 , verbose = True, measure = True)
end = time()
print(f"Running time for PPDNA was {end-start} seconds")
results['PPDNA'] = {'Theta' : sol['Theta']}
#%%
# solve with general ADMM
start = time()
sol, info = ADMM_MGL(S, lambda1, lambda2, reg, Omega_0, Theta_0 = Theta_0, X_0 = X_0, rho = 1, max_iter = 100, \
eps_admm = 5e-4, verbose = True, measure = True)
end = time()
print(f"Running time for ADMM was {end-start} seconds")
results['ADMM'] = {'Theta' : sol['Theta']}
#%%
# solve with ragin (LTGL)
start = time()
alpha = N*lambda1
beta = N*lambda2
ltgl = TimeGraphicalLasso(alpha = alpha, beta = beta , psi = 'l1', \
rho = 1., tol = 5e-4, rtol = 5e-4, max_iter = 2000, verbose = True)
ltgl = ltgl.fit(sample.transpose(0,2,1))
DFPR = np.zeros((grid1, grid2))
DTPR = np.zeros((grid1, grid2))
AIC = np.zeros((grid1, grid2))
BIC = np.zeros((grid1, grid2))
Omega_0 = get_K_identity(K,p)
Theta_0 = get_K_identity(K,p)
X_0 = np.zeros((K,p,p))
for g2 in np.arange(grid2):
for g1 in np.arange(grid1):
lambda1 = L1[g1,g2]
lambda2 = L2[g1,g2]
sol, info = ADMM_MGL(S, lambda1, lambda2, reg , Omega_0, Theta_0 = Theta_0, X_0 = X_0, tol = 1e-8, rtol = 1e-8, verbose = False, measure = False)
Theta_sol = sol['Theta']
Omega_sol = sol['Omega']
X_sol = sol['X']
# warm start
Omega_0 = Omega_sol.copy()
Theta_0 = Theta_sol.copy()
X_0 = X_sol.copy()
dr = discovery_rate(Theta_sol, Theta)
TPR[g1,g2] = dr['TPR']
FPR[g1,g2] = dr['FPR']
DTPR[g1,g2] = dr['TPR_DIFF']
DFPR[g1,g2] = dr['FPR_DIFF']
Omega_0 = get_K_identity(K, p)
Theta_0 = get_K_identity(K, p)
X_0 = np.zeros((K, p, p))
for g2 in np.arange(grid2):
for g1 in np.arange(grid1):
lambda1 = L1[g1, g2]
lambda2 = L2[g1, g2]
sol, info = ADMM_MGL(S,
lambda1,
lambda2,
reg,
Omega_0,
Theta_0=Theta_0,
X_0=X_0,
tol=1e-8,
rtol=1e-8,
verbose=False,
measure=False)
Theta_sol = sol['Theta']
Omega_sol = sol['Omega']
X_sol = sol['X']
# warm start
Omega_0 = Omega_sol.copy()
Theta_0 = Theta_sol.copy()
X_0 = X_sol.copy()
Omega_0 = get_K_identity(K, p)
Theta_0 = get_K_identity(K, p)
for g1 in np.arange(grid1):
for g2 in np.arange(grid2):
lambda1 = L1[g1, g2]
lambda2 = L2[g1, g2]
#sol, info = warmPPDNA(S_train, lambda1, lambda2, reg, Omega_0, Theta_0 = Theta_0, eps = 1e-3, verbose = False, measure = False)
sol, info = ADMM_MGL(S_train,
lambda1,
lambda2,
reg,
Omega_0,
Theta_0=Theta_0,
tol=1e-10,
rtol=1e-10,
verbose=False,
measure=False)
Theta_sol = sol['Theta']
Omega_sol = sol['Omega']
# warm start
Omega_0 = Omega_sol.copy()
Theta_0 = Theta_sol.copy()
TPR[g1, g2] = discovery_rate(Theta_sol, Theta)['TPR']
FPR[g1, g2] = discovery_rate(Theta_sol, Theta)['FPR']
DTPR[g1, g2] = discovery_rate(Theta_sol, Theta)['TPR_DIFF']
K=allK[j],
M=5,
scale=False,
nxseed=1235)
S, sample = sample_covariance_matrix(Sigma, N)
print("Shape of empirical covariance matrix: ", S.shape)
print("Shape of the sample array: ", sample.shape)
allS.append(S)
#%% initialize solvers (for numba jitting)
_, _ = ADMM_MGL(allS[0],
lambda1,
lambda2,
"GGL",
allS[0],
max_iter=10,
measure=False,
latent=False)
_, _ = ADMM_MGL(allS[0],
lambda1,
lambda2,
"FGL",
allS[0],
max_iter=10,
measure=False,
latent=False)
_ = ADMM_SGL(allS[0][0],
lambda1,
allS[0][0],