def test_stmvn_prior(method='SVD'):
ridges = np.logspace(0,1,5)
nridges = len(ridges)
ndelays = 10
delays = range(ndelays)
features_train, features_test, responses_train, responses_test = get_abc_data()
features_sizes = [fs.shape[1] for fs in features_train]
spatial_priors = [sps.SphericalPrior(features_sizes[0]),
sps.SphericalPrior(features_sizes[1], hyparams=np.logspace(-3,3,7)),
sps.SphericalPrior(features_sizes[2], hyparams=np.logspace(-3,3,7)),
]
# do not scale first. this removes duplicates
spatial_priors[0].set_hyparams(1.0)
# non-diagonal hyper-prior
W = np.random.randn(ndelays, ndelays)
W = np.dot(W.T, W)
tpriors = [tps.SphericalPrior(delays),
tps.GaussianKernelPrior(delays, hhparams=np.linspace(1,ndelays/2,ndelays)),
tps.SmoothnessPrior(delays, hhparams=np.logspace(-3,1,2)),
tps.SmoothnessPrior(delays, wishart=W, hhparams=np.logspace(-3,3,5)),
tps.SmoothnessPrior(delays, wishart=True),
tps.SmoothnessPrior(delays, wishart=False),
tps.HRFPrior([1] if delays == [0] else delays),
]
from tikreg import models
res = models.crossval_stem_wmvnp(features_train,
responses_train,
temporal_prior=tpriors[2],
feature_priors=spatial_priors,
folds=(1,5),
ridges=ridges,
verbosity=1,
method=method,
)
# find optima
cvmean = res['cvresults'].mean(0)
population_optimal = False
if population_optimal is True:
cvmean = np.nan_to_num(cvmean).mean(-1)[...,None]
for idx in range(cvmean.shape[-1]):
tpopt, spopt, ropt = models.find_optimum_mvn(cvmean[...,idx],
res['temporal'],
res['spatial'],
res['ridges'],
)
txt = "temporal=%0.03f, spatial=(%0.03f,%0.03f, %0.03f), ridge=%0.03f"
content = tuple([tpopt])+tuple(spopt)+tuple([ropt])
print(txt % content)
return res
def test_cv_api(show_figures=False, ntest=50):
# if show_figures=True, this function will create
# images of the temporal priors, and the feature prior hyparams in 3D
ridges = [0., 1e-03, 1., 10.0, 100.]
nridges = len(ridges)
ndelays = 10
delays = range(ndelays)
features_train, features_test, responses_train, responses_test = get_abc_data(
)
features_sizes = [fs.shape[1] for fs in features_train]
spatial_priors = [
sps.SphericalPrior(features_sizes[0]),
sps.SphericalPrior(features_sizes[1], hyparams=np.logspace(-3, 3, 7)),
sps.SphericalPrior(features_sizes[2], hyparams=np.logspace(-3, 3, 7)),
]
# do not scale first. this removes duplicates
spatial_priors[0].set_hyparams(1.0)
# non-diagonal hyper-prior
W = np.random.randn(ndelays, ndelays)
W = np.dot(W.T, W)
tpriors = [
tps.SphericalPrior(delays),
tps.SmoothnessPrior(delays, hhparams=np.logspace(-3, 1, 8)),
tps.SmoothnessPrior(delays, wishart=True),
tps.SmoothnessPrior(delays, wishart=False),
tps.SmoothnessPrior(delays, wishart=W, hhparams=np.logspace(-3, 3, 5)),
tps.GaussianKernelPrior(delays,
hhparams=np.linspace(1, ndelays / 2, ndelays)),
tps.HRFPrior([1] if delays == [0] else delays),
]
nfolds = (1, 5) # 1 times 5-fold cross-validation
folds = tikutils.generate_trnval_folds(responses_train.shape[0],
sampler='bcv',
nfolds=nfolds)
nfolds = np.prod(nfolds)
for ntp, temporal_prior in enumerate(tpriors):
print(temporal_prior)
all_temporal_hypers = [temporal_prior.get_hhparams()]
all_spatial_hypers = [t.get_hyparams() for t in spatial_priors]
# get all combinations of hyparams
all_hyperparams = list(
itertools.product(*(all_temporal_hypers + all_spatial_hypers)))
nspatial_hyperparams = np.prod([len(t) for t in all_spatial_hypers])
ntemporal_hyperparams = np.prod([len(t) for t in all_temporal_hypers])
population_mean = False
results = np.zeros(
(nfolds, ntemporal_hyperparams, nspatial_hyperparams, nridges,
1 if population_mean else responses_train.shape[-1]),
dtype=[
('fold', np.float32),
('tp', np.float32),
('sp', np.float32),
('ridges', np.float32),
('responses', np.float32),
])
for hyperidx, spatiotemporal_hyperparams in enumerate(all_hyperparams):
temporal_hyperparam = spatiotemporal_hyperparams[0]
spatial_hyperparams = spatiotemporal_hyperparams[1:]
spatial_hyperparams /= np.linalg.norm(spatial_hyperparams)
# get indices
shyperidx = np.mod(hyperidx, nspatial_hyperparams)
thyperidx = int(hyperidx // nspatial_hyperparams)
print(thyperidx,
temporal_hyperparam), (shyperidx, spatial_hyperparams)
this_temporal_prior = temporal_prior.get_prior(
alpha=1.0, hhparam=temporal_hyperparam)
if show_figures:
from matplotlib import pyplot as plt
if (hyperidx == 0) and (ntp == 0):
# show points in 3D
from tikreg import priors
cartesian_points = [
t[1:] / np.linalg.norm(t[1:]) for t in all_hyperparams
]
angles = priors.cartesian2polar(
np.asarray(cartesian_points))
priors.show_spherical_angles(angles[:, 0], angles[:, 1])
if hyperidx == 0:
# show priors with different hyper-priors
oldthyper = 0
plt.matshow(this_temporal_prior, cmap='inferno')
else:
if thyperidx > oldthyper:
oldthyper = thyperidx
plt.matshow(this_temporal_prior, cmap='inferno')
# only run a few
if hyperidx > ntest:
continue
Ktrain = 0.
Kval = 0.
for fdx, (fs_train, fs_test, fs_prior, fs_hyper) in enumerate(
zip(features_train, features_test, spatial_priors,
spatial_hyperparams)):
kernel_train = models.kernel_spatiotemporal_prior(
fs_train,
this_temporal_prior,
fs_prior.get_prior(fs_hyper),
delays=delays)
Ktrain += kernel_train
kernel_normalizer = tikutils.determinant_normalizer(Ktrain)
Ktrain /= float(kernel_normalizer)
# cross-validation
for ifold, (trnidx, validx) in enumerate(folds):
ktrn = tikutils.fast_indexing(Ktrain, trnidx, trnidx)
kval = tikutils.fast_indexing(Ktrain, validx, trnidx)
fit = models.solve_l2_dual(ktrn,
responses_train[trnidx],
kval,
responses_train[validx],
ridges=ridges,
verbose=False,
performance=True)
if population_mean:
cvfold = np.nan_to_num(fit['performance']).mean(-1)[...,
None]
else:
cvfold = fit['performance']
results[ifold, thyperidx, shyperidx] = cvfold
def test_general_solution(temporal_prior_name='spherical'):
tprior_names = ['spherical', 'smooth', 'hrf', 'gaussian']
normalize_kernel = False
method = 'SVD'
# make sure we can recover the ridge solution
ridges = np.round(np.logspace(1, 3, 5), 4)
nridges = len(ridges)
delays = range(10) #np.unique(np.random.randint(0,10,10))
ndelays = len(delays)
features_train, features_test, responses_train, responses_test = get_abc_data(
)
features_sizes = [fs.shape[1] for fs in features_train]
# custom effective low-rank prior
a = np.random.randn(features_train[-1].shape[-1], 3)
sigma_x = np.dot(a, a.T) + np.identity(a.shape[0])
spatial_priors = [
sps.SphericalPrior(features_sizes[0], hyparams=[1]),
sps.SphericalPrior(features_sizes[1], hyparams=[0.1, 1]),
sps.CustomPrior(sigma_x, hyparams=[0.1, 1])
]
tpriors = [
tps.SphericalPrior(delays),
tps.SmoothnessPrior(delays, wishart=False),
tps.HRFPrior(delays),
tps.GaussianKernelPrior(delays),
]
tpidx = tprior_names.index(temporal_prior_name)
temporal_prior = tpriors[tpidx]
folds = tikutils.generate_trnval_folds(
responses_train.shape[0],
sampler='bcv',
nfolds=(1, 5),
)
folds = list(folds)
res = models.crossval_stem_wmvnp(
features_train,
responses_train,
temporal_prior=temporal_prior,
feature_priors=spatial_priors,
folds=folds,
ridges=ridges,
verbosity=2,
method=method,
normalize_kernel=normalize_kernel,
)
# select a non-spherical prior
spidx = 0
sprior_ridge = res['spatial'][spidx]
newridges = res['ridges']
ridge_scale = newridges[0]
res = models.estimate_simple_stem_wmvnp(
features_train,
responses_train,
features_test=None,
responses_test=None,
temporal_prior=temporal_prior,
temporal_hhparam=1.0,
feature_priors=spatial_priors,
feature_hyparams=sprior_ridge,
weights=True,
performance=False,
predictions=False,
ridge_scale=ridge_scale,
verbosity=2,
method='SVD',
)
weights = models.dual2primal_weights(
res['weights'],
features_train,
spatial_priors,
sprior_ridge,
temporal_prior,
)
weights = np.vstack(weights)
### solve problem directly
Xx = np.hstack([tikutils.delay_signal(t.astype(np.float64), delays)\
for i,t in enumerate(features_train)])
# get scaled priors
spriors = [
sp.get_prior(param) for sp, param in zip(spatial_priors, sprior_ridge)
]
# get temporal prior
tprior = temporal_prior.get_prior(1.0)
tprior += np.identity(tprior.shape[0]) * 1e-10
# combine
from scipy import linalg as LA
prior = LA.block_diag(*[np.kron(tprior, spr) for spr in spriors])
# solve problem indirectly # dual
XSigmaXT = np.linalg.multi_dot(
[Xx, prior, Xx.T]) + (ridge_scale**2.0) * np.identity(Xx.shape[0])
alphas = np.dot(np.linalg.inv(XSigmaXT), responses_train)
assert np.allclose(alphas, res['weights'])
betas_dual = np.linalg.multi_dot([prior, Xx.T, alphas])
assert np.allclose(betas_dual, weights)
# solve problem directly # primal
penalty = np.linalg.inv(prior)
XTXSigma = np.dot(Xx.T, Xx) + (ridge_scale**2.0) * penalty
XTY = np.dot(Xx.T, responses_train)
betas = np.dot(np.linalg.inv(XTXSigma), XTY)
# check solutions
try:
assert np.allclose(betas, weights)
except AssertionError:
# numerical error with HRF because of rank
print('asserting correlation')
assert np.allclose(
np.corrcoef(betas.ravel(), weights.ravel())[0, 1], 1.0)
def test_ols():
# test we can get OLS solution
delays = [0]
ndelays = len(delays)
# make some features and signal for which we know
# the optimal ridge parameter is zero
Af = np.random.randn(150, 10)
Bf = np.random.randn(150, 20)
Cf = np.random.randn(150, 30)
A, Atest = Af[:100], Af[100:]
B, Btest = Bf[:100], Bf[100:]
C, Ctest = Cf[:100], Cf[100:]
nvox = 20
Aw = np.random.randn(Af.shape[-1], nvox)
Bw = np.random.randn(Bf.shape[-1], nvox)
Cw = np.random.randn(Cf.shape[-1], nvox)
Yf = np.dot(Af, Aw) + np.dot(Bf, Bw) + np.dot(Cf, Cw)
responses_train, responses_test = Yf[:100], Yf[100:]
features_train = [A, B, C]
features_test = [Atest, Btest, Ctest]
features_sizes = [fs.shape[1] for fs in features_train]
# solve for OLS using L2 machinery
direct_fit = models.solve_l2_primal(
tikutils.delay_signal(np.hstack(features_train), delays),
responses_train,
tikutils.delay_signal(np.hstack(features_test), delays),
responses_test,
verbose=True,
ridges=[0.],
weights=True,
performance=True,
predictions=True)
# create feature priors
spatial_priors = [
sps.SphericalPrior(features_sizes[0], hyparams=[1]),
sps.SphericalPrior(features_sizes[1], hyparams=[1]),
sps.SphericalPrior(features_sizes[2], hyparams=[1]),
]
# test all priors
tpriors = [
tps.SmoothnessPrior(delays),
tps.SmoothnessPrior(delays, wishart=True),
tps.SmoothnessPrior(delays, wishart=False),
tps.SmoothnessPrior(delays, wishart=np.eye(len(delays))),
tps.GaussianKernelPrior(delays, sigma=2.0),
tps.HRFPrior([1] if delays ==
[0] else delays), # b/c delay at 0 has no covariance
tps.SphericalPrior(delays),
]
for temporal_prior in tpriors:
print(temporal_prior)
all_temporal_hypers = [temporal_prior.get_hyparams()]
all_spatial_hypers = [[1.]] * len(spatial_priors)
# get all combinations of hyparams
all_hyperparams = itertools.product(*(all_temporal_hypers +
all_spatial_hypers))
Ktrain = 0.
Ktest = 0.
for spatiotemporal_hyperparams in all_hyperparams:
temporal_hyperparam = spatiotemporal_hyperparams[0]
spatial_hyperparams = spatiotemporal_hyperparams[1:]
this_temporal_prior = temporal_prior.get_prior(
alpha=1.0, hhparam=temporal_hyperparam)
for fdx, (fs_train, fs_test, fs_prior, fs_hyper) in enumerate(
zip(features_train, features_test, spatial_priors,
spatial_hyperparams)):
kernel_train = models.kernel_spatiotemporal_prior(
fs_train,
this_temporal_prior,
fs_prior.get_prior(fs_hyper),
delays=delays)
Ktrain += kernel_train
kernel_test = models.kernel_spatiotemporal_prior(
fs_train,
this_temporal_prior,
fs_prior.get_prior(fs_hyper),
Xtest=fs_test,
delays=delays)
Ktest += kernel_test
fit = models.solve_l2_dual(Ktrain,
responses_train,
Ktest,
responses_test,
ridges=[0., 1e-03, 1., 10.0, 100.],
verbose=True,
weights=True,
performance=True,
predictions=True)
# make sure we can predict perfectly
assert np.allclose(fit['performance'][0], 1.)
# get the feature weights from the kernel weights
weights = np.tensordot(
tikutils.delay_signal(np.hstack(features_train), delays).T,
fit['weights'], (1, 1)).swapaxes(0, 1)
if not np.allclose(this_temporal_prior, 1):
# scale weights to account for temporal hyper-prior scale
weights *= this_temporal_prior
assert np.allclose(weights[0], direct_fit['weights'])
assert np.allclose(fit['predictions'][0],
direct_fit['predictions'].squeeze())
def test_gaussian_kernel_prior():
# TODO
prior = tp.GaussianKernelPrior(delays=range(10))