def test_pca_destructor():
destructor = CompositeDestructor(destructors=[
LinearProjector(
linear_estimator=PCA(),
orthogonal=False,
),
IndependentDestructor(),
], )
assert check_destructor(destructor, is_canonical=False)
def test_random_linear_householder_destructor():
destructor = CompositeDestructor(
destructors=[
LinearProjector(
# Since n_components=1, a Householder matrix will be used
linear_estimator=RandomOrthogonalEstimator(n_components=1),
orthogonal=False,
),
IndependentDestructor(),
], )
assert check_destructor(destructor, is_canonical=False)
def _fit_and_score(data_name, destructor, destructor_name, n_train, random_state=0):
"""Simple function to fit and score a destructor."""
# Fix random state of global generator so repeatable if destructors are random
rng = check_random_state(random_state)
old_random_state = np.random.get_state()
np.random.seed(rng.randint(2 ** 32, dtype=np.uint32))
try:
# Fit destructor
start_time = time.time()
destructor.fit(X_train)
train_time = time.time() - start_time
except RuntimeError as e:
# Handle MLPACK error
if 'mlpack' not in str(e).lower():
raise e
warnings.warn('Skipping density tree destructors because of MLPACK error "%s". '
'Using dummy IndependentDestructor() instead.' % str(e))
destructor = CompositeDestructor([IndependentDestructor()]).fit(X_train)
train_time = 0
train_score = -np.inf
test_score = -np.inf
score_time = 0
else:
# Get scores
start_time = time.time()
train_score = destructor.score(X_train)
test_score = destructor.score(X_test)
score_time = time.time() - start_time
logger.debug(
'train=%.3f, test=%.3f, train_time=%.3f, score_time=%.3f, destructor=%s, data_name=%s'
% (train_score, test_score, train_time, score_time, destructor_name, data_name))
# Reset random state
np.random.set_state(old_random_state)
return dict(fitted_destructor=destructor,
destructor_name=destructor_name,
train_score=train_score,
test_score=test_score)
def _get_model(data_name, model_name, model_kwargs):
if 'is_test' not in model_kwargs:
model_kwargs['is_test'] = False
# Init destructor is shared with all models
init_destructor = CompositeDestructor(
destructors=[
_get_inverse_logit_destructor(data_name),
IndependentDestructor(independent_density=IndependentDensity(
univariate_estimators=HistogramUnivariateDensity(
bins=256, bounds=[0, 1], alpha=1)))
],
random_state=0,
)
# Setup canonical destructor for various models
if model_name == 'deep-copula':
deep_stop_tol = 0.001
canonical_destructor = _get_copula_destructor()
else:
deep_stop_tol = 0.0001
n_jobs = model_kwargs['n_jobs']
# Get pair estimators (i.e. pairs of pixels in a spiral pattern)
pair_estimators = _get_pair_estimators(data_name, n_uniq_dir=8)
# Setup the local/pair destructor
pair_canonical_destructor = _get_pair_canonical_destructor(model_name)
# Setup a list of canonical destructors that destroy in each pixel direction
canonical_destructor = [
FeatureGroupsDestructor(
groups_estimator=pair_estimator,
group_canonical_destructor=clone(pair_canonical_destructor),
n_jobs=n_jobs) for pair_estimator in pair_estimators
]
# Shared DeepDestructorCV
return DeepDestructorCV(
init_destructor=init_destructor,
canonical_destructor=canonical_destructor,
stop_tol=deep_stop_tol,
# Either n_extend or max_canonical_destructors must be None
n_extend=1,
cv=model_kwargs['cv'],
refit=model_kwargs['refit'],
silent=False,
log_prefix='',
random_state=0,
# Set maximum number of layers (None for infinite)
max_canonical_destructors=None if not model_kwargs['is_test'] else 1,
)
def _get_copula_destructor(hist_kwargs=None):
if hist_kwargs is None:
hist_kwargs = dict(bins=40, bounds=[0, 1], alpha=100)
return CompositeDestructor(
destructors=[
IndependentDestructor(independent_density=IndependentDensity(
univariate_estimators=HistogramUnivariateDensity(
**hist_kwargs))),
IndependentInverseCdf(),
BestLinearReconstructionDestructor(
linear_estimator=PCA(),
destructor=IndependentDestructor(),
)
],
random_state=0,
)
def _get_inverse_logit_destructor(data_name):
if data_name == 'mnist':
alpha = MNIST_ALPHA
elif data_name == 'cifar10':
alpha = CIFAR10_ALPHA
else:
raise ValueError('dataset should either be mnist or cifar10')
inverse_logit = CompositeDestructor(destructors=[
IndependentDestructor(independent_density=IndependentDensity(
univariate_estimators=ScipyUnivariateDensity(
scipy_rv=scipy.stats.logistic,
scipy_fit_kwargs=dict(floc=0, fscale=1)))),
IndependentDestructor(independent_density=IndependentDensity(
univariate_estimators=ScipyUnivariateDensity(
scipy_rv=scipy.stats.uniform,
scipy_fit_kwargs=dict(floc=alpha, fscale=1 - 2 * alpha))))
])
return inverse_logit
def get_rbig_model(bins="auto", bounds=0.1, alpha=1e-10):
# ================================ #
# Step I - Marginal Uniformization
# ================================ #
# Choose the Histogram estimator that converts the data X to uniform U(0,1)
univariate_estimator = HistogramUnivariateDensity(bounds=bounds,
bins=bins,
alpha=alpha)
# Marginally uses histogram estimator
marginal_uniformization = IndependentDensity(
univariate_estimators=univariate_estimator)
# Creates "Destructor" D_theta_1
uniform_density = IndependentDestructor(marginal_uniformization)
# ================================== #
# Step II - Marginal Gaussianization
# ================================== #
# Choose destructor D_theta_2 that converts data
marginal_gaussianization = IndependentInverseCdf()
# =================== #
# Step III - Rotation
# =================== #
# Choose a linear projection to rotate the features (PCA) "D_theta_3"
rotation = LinearProjector(linear_estimator=PCA())
# ==================== #
# Composite Destructor
# ==================== #
# Composite Destructor
rbig_model = CompositeDestructor([
clone(uniform_density), # Marginal Uniformization
clone(marginal_gaussianization), # Marginal Gaussianization
clone(rotation), # Rotation (PCA)
])
return rbig_model
def _get_toy_destructors_and_names():
# BASELINE SHALLOW DESTRUCTORS
gaussian_full = CompositeDestructor(
destructors=[
LinearProjector(
linear_estimator=PCA(),
orthogonal=False,
),
IndependentDestructor(),
],
)
mixture_20 = AutoregressiveDestructor(
density_estimator=GaussianMixtureDensity(
covariance_type='spherical',
n_components=20,
)
)
random_tree = CompositeDestructor(
destructors=[
IndependentDestructor(),
TreeDestructor(
tree_density=TreeDensity(
tree_estimator=RandomTreeEstimator(min_samples_leaf=20, max_leaf_nodes=50),
node_destructor=IndependentDestructor(
independent_density=IndependentDensity(
univariate_estimators=HistogramUnivariateDensity(
bins=10, alpha=10, bounds=[0, 1]
)
)
)
)
)
]
)
density_tree = CompositeDestructor(
destructors=[
IndependentDestructor(),
TreeDestructor(
tree_density=TreeDensity(
tree_estimator=MlpackDensityTreeEstimator(min_samples_leaf=10),
uniform_weight=0.001,
)
)
]
)
baseline_destructors = [gaussian_full, mixture_20, random_tree, density_tree]
baseline_names = ['Gaussian', 'Mixture', 'SingleRandTree', 'SingleDensityTree']
# LINEAR DESTRUCTORS
alpha_histogram = [10] # [1, 10, 100]
random_linear_projector = LinearProjector(
linear_estimator=RandomOrthogonalEstimator(), orthogonal=True
)
canonical_histogram_destructors = [
IndependentDestructor(
independent_density=IndependentDensity(
univariate_estimators=HistogramUnivariateDensity(bins=20, bounds=[0, 1], alpha=a)
)
)
for a in alpha_histogram
]
linear_destructors = [
DeepDestructorCV(
init_destructor=IndependentDestructor(),
canonical_destructor=CompositeDestructor(destructors=[
IndependentInverseCdf(), # Project to inf real space
random_linear_projector, # Random linear projector
IndependentDestructor(), # Project to canonical space
destructor, # Histogram destructor in canonical space
]),
n_extend=20, # Need to extend since random projections
)
for destructor in canonical_histogram_destructors
]
linear_names = ['RandLin (%g)' % a for a in alpha_histogram]
# MIXTURE DESTRUCTORS
fixed_weight = [0.5] # [0.1, 0.5, 0.9]
mixture_destructors = [
CompositeDestructor(destructors=[
IndependentInverseCdf(),
AutoregressiveDestructor(
density_estimator=FirstFixedGaussianMixtureDensity(
covariance_type='spherical',
n_components=20,
fixed_weight=w,
)
)
])
for w in fixed_weight
]
# Make deep destructors
mixture_destructors = [
DeepDestructorCV(
init_destructor=IndependentDestructor(),
canonical_destructor=destructor,
n_extend=5,
)
for destructor in mixture_destructors
]
mixture_names = ['GausMix (%.2g)' % w for w in fixed_weight]
# TREE DESTRUCTORS
# Random trees
histogram_alpha = [10] # [1, 10, 100]
tree_destructors = [
TreeDestructor(
tree_density=TreeDensity(
tree_estimator=RandomTreeEstimator(
max_leaf_nodes=4
),
node_destructor=IndependentDestructor(
independent_density=IndependentDensity(
univariate_estimators=HistogramUnivariateDensity(
alpha=a, bins=10, bounds=[0, 1]
)
)
),
)
)
for a in histogram_alpha
]
tree_names = ['RandTree (%g)' % a for a in histogram_alpha]
# Density trees using mlpack
tree_uniform_weight = [0.5] # [0.1, 0.5, 0.9]
tree_destructors.extend([
TreeDestructor(
tree_density=TreeDensity(
tree_estimator=MlpackDensityTreeEstimator(min_samples_leaf=10),
uniform_weight=w,
)
)
for w in tree_uniform_weight
])
tree_names.extend(['DensityTree (%.2g)' % w for w in tree_uniform_weight])
# Add random rotation to tree destructors
tree_destructors = [
CompositeDestructor(destructors=[
IndependentInverseCdf(),
LinearProjector(linear_estimator=RandomOrthogonalEstimator()),
IndependentDestructor(),
destructor,
])
for destructor in tree_destructors
]
# Make deep destructors
tree_destructors = [
DeepDestructorCV(
init_destructor=IndependentDestructor(),
canonical_destructor=destructor,
# Density trees don't need to extend as much as random trees
n_extend=50 if 'Rand' in name else 5,
)
for destructor, name in zip(tree_destructors, tree_names)
]
# Collect all destructors and set CV parameter
destructors = baseline_destructors + linear_destructors + mixture_destructors + tree_destructors
destructor_names = baseline_names + linear_names + mixture_names + tree_names
for d in destructors:
if 'cv' in d.get_params():
d.set_params(cv=cv)
# **** Change from notebook to make faster ****
if 'max_canonical_destructors' in d.get_params():
d.set_params(max_canonical_destructors=1)
return destructors, destructor_names