def train(self,
V,
cardinality,
L,
L_offset,
y=None,
deps=(),
init_acc=1.0,
init_deps=0.0,
init_class_prior=-1.0,
epochs=100,
step_size=None,
decay=0.99,
reg_param=0.1,
reg_type=2,
verbose=False,
truncation=10,
burn_in=50,
timer=None):
n_data = V.shape[0]
step_size = step_size or 1.0 / n_data
reg_param_scaled = reg_param / n_data
# self._process_dependency_graph(L, deps)
weight, variable, factor, ftv, domain_mask, n_edges = self._compile(
V, cardinality, L, L_offset, y, deps, init_acc,
init_deps) # , init_deps, init_class_prior)
fg = NumbSkull(n_inference_epoch=0,
n_learning_epoch=epochs,
stepsize=step_size,
decay=decay,
reg_param=reg_param_scaled,
regularization=reg_type,
truncation=truncation,
quiet=(not verbose),
verbose=verbose,
learn_non_evidence=True,
burn_in=burn_in)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
if timer is not None:
timer.start()
fg.learning(out=False)
if timer is not None:
timer.end()
self.weights = fg.factorGraphs[0].weight_value[0][:len(L)]
self.dep_weights = fg.factorGraphs[0].weight_value[0][len(L):]
self.lf_accuracy = 1. / (1. + np.exp(-self.weights[:len(L)]))
def fg_learning(compile_fg):
print('----------------------------------------------------------------')
print('Weight learning')
print('----------------------------------------------------------------')
weight, variable, factor, ftv, domain_mask, n_edges = compile_fg
fg = NumbSkull(
n_inference_epoch=1000,
n_learning_epoch=1000,
stepsize=0.01,
decay=0.95,
reg_param=1e-6,
regularization=2,
truncation=10,
quiet=(not False),
verbose=False,
learn_non_evidence=False, # need to test
sample_evidence=False,
burn_in=10,
nthreads=1)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
fg.learning(out=True)
weight_value = fg.factorGraphs[0].weight_value[0]
return weight_value
def train(self,
L,
y=None,
deps=(),
init_acc=1.0,
epochs=100,
step_size=None,
decay=0.99,
reg_param=0.1,
reg_type=2,
verbose=False,
truncation=10,
burn_in=50,
timer=None):
step_size = step_size or 1.0 / L.shape[0]
reg_param_scaled = reg_param / L.shape[0]
self._process_dependency_graph(L, deps)
weight, variable, factor, ftv, domain_mask, n_edges = self._compile(
L, y, init_acc)
fg = NumbSkull(n_inference_epoch=0,
n_learning_epoch=epochs,
stepsize=step_size,
decay=decay,
reg_param=reg_param_scaled,
regularization=reg_type,
truncation=truncation,
quiet=(not verbose),
verbose=verbose,
learn_non_evidence=True,
burn_in=burn_in)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
if timer is not None:
timer.start()
fg.learning(out=False)
if timer is not None:
timer.end()
self._process_learned_weights(L, fg)
def train(self, L, deps=(), LF_acc_prior_weights=None,
LF_acc_prior_weight_default=1, labels=None, label_prior_weight=5,
init_deps=0.0, init_class_prior=-1.0, epochs=30, step_size=None,
decay=1.0, reg_param=0.1, reg_type=2, verbose=False, truncation=10,
burn_in=5, cardinality=None, timer=None, candidate_ranges=None, threads=1):
"""
Fits the parameters of the model to a data set. By default, learns a
conditionally independent model. Additional unary dependencies can be
set to be included in the constructor. Additional pairwise and
higher-order dependencies can be included as an argument.
Results are stored as a member named weights, instance of
snorkel.learning.gen_learning.GenerativeModelWeights.
:param L: M x N csr_AnnotationMatrix-type label matrix, where there are
M candidates labeled by N labeling functions (LFs)
:param deps: collection of dependencies to include in the model, each
element is a tuple of the form
(LF 1 index, LF 2 index, dependency type),
see snorkel.learning.constants
:param LF_acc_prior_weights: An N-element list of prior weights for the
LF accuracies (log scale)
:param LF_acc_prior_weight_default: Default prior for the weight of each
LF accuracy; if LF_acc_prior_weights is unset, each LF will have
this accuracy prior weight (log scale)
:param labels: Optional ground truth labels
:param label_prior_weight: The prior probability that the ground truth
labels (if provided) are correct (log scale)
:param init_deps: initial weight for additional dependencies, except
class prior (log scale)
:param init_class_prior: initial class prior (in log scale), note only
used if class_prior=True in constructor
:param epochs: number of training epochs
:param step_size: gradient step size, default is 1 / L.shape[0]
:param decay: multiplicative decay of step size,
step_size_(t+1) = step_size_(t) * decay
:param reg_param: regularization strength
:param reg_type: 1 = L1 regularization, 2 = L2 regularization
:param verbose: whether to write debugging info to stdout
:param truncation: number of iterations between truncation step for L1
regularization
:param burn_in: number of burn-in samples to take before beginning
learning
:param cardinality: number of possible classes; by default is inferred
from the label matrix L
:param timer: stopwatch for profiling, must implement start() and end()
:param candidate_ranges: Optionally, a list of M sets of integer values,
representing the possible categorical values that each of the M
candidates can take. If a label is outside of this range throws an
error. If None, then each candidate can take any value from 0 to
cardinality.
:param threads: the number of threads to use for sampling. Default is 1.
"""
m, n = L.shape
step_size = step_size or 0.0001
# Check to make sure matrix is int-valued
element_type = type(L[0,0])
# Note: Other simpler forms of this check often don't work; still not
# sure why...
if not issubclass(element_type, np.integer):
raise ValueError("""Label matrix must have int-type elements,
but elements have type %s""" % element_type)
# Automatically infer cardinality
# Binary: Values in {-1, 0, 1} [Default]
# Categorical: Values in {0, 1, ..., K}
if cardinality is None:
# If candidate_ranges is provided, use this to determine cardinality
if candidate_ranges is not None:
cardinality = max(map(max, candidate_ranges))
else:
# This is just an annoying hack for LIL sparse matrices...
try:
lmax = L.max()
except AttributeError:
lmax = L.tocoo().max()
if lmax > 2:
cardinality = lmax
elif lmax < 2:
cardinality = 2
else:
raise ValueError(
"L.max() == %s, cannot infer cardinality." % lmax)
print("Inferred cardinality: %s" % cardinality)
self.cardinality = cardinality
# Priors for LFs default to fixed prior value
# NOTE: Setting default != 0.5 creates a (fixed) factor which increases
# runtime (by ~0.5x that of a non-fixed factor)...
if LF_acc_prior_weights is None:
LF_acc_prior_weights = [LF_acc_prior_weight_default for _ in range(n)]
else:
LF_acc_prior_weights = list(copy(LF_acc_prior_weights))
# LF weights are un-fixed
is_fixed = [False for _ in range(n)]
# If supervised labels are provided, add them as a fixed LF with prior
# Note: For large L this column stack operation could be very
# inefficient, can consider refactoring...
if labels is not None:
labels = labels.reshape(m, 1)
L = sparse.hstack([L, labels])
is_fixed.append(True)
LF_acc_prior_weights.append(label_prior_weight)
n += 1
# Reduce overhead of tracking indices by converting L to a CSR sparse matrix.
L = sparse.csr_matrix(L).copy()
# If candidate_ranges is provided, remap the values of L using
# candidate_ranges. This "scoped categorical" approach allows learning
# and inference to be efficient even with very large cardinality, as
# we only sample relevant values for each candidate. Also set
# per-candidate cardinalities according to candidate_ranges if not None,
# else as constant value.
self.cardinalities = self.cardinality * np.ones(m, dtype=np.int64)
self.candidate_ranges = candidate_ranges
if self.candidate_ranges is not None:
L, self.cardinalities, _ = self._remap_scoped_categoricals(L,
self.candidate_ranges)
# Shuffle the data points, cardinalities, and candidate_ranges
idxs = range(m)
self.rng.shuffle(idxs)
L = L[idxs, :]
if candidate_ranges is not None:
self.cardinalities = self.cardinalities[idxs]
c_ranges_reshuffled = []
for i in idxs:
c_ranges_reshuffled.append(self.candidate_ranges[i])
self.candidate_ranges = c_ranges_reshuffled
# Compile factor graph
self._process_dependency_graph(L, deps)
weight, variable, factor, ftv, domain_mask, n_edges = self._compile(
L, init_deps, init_class_prior, LF_acc_prior_weights, is_fixed, self.cardinalities)
fg = NumbSkull(
n_inference_epoch=0,
n_learning_epoch=epochs,
stepsize=step_size,
decay=decay,
reg_param=reg_param,
regularization=reg_type,
truncation=truncation,
quiet=(not verbose),
verbose=verbose,
learn_non_evidence=True,
burn_in=burn_in,
nthreads=threads
)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
if timer is not None:
timer.start()
fg.learning(out=False)
if timer is not None:
timer.end()
self._process_learned_weights(L, fg, LF_acc_prior_weights, is_fixed)
# Store info from factor graph
if self.candidate_ranges is not None:
self.cardinality_for_stats = int(max(self.cardinalities))
else:
self.cardinality_for_stats = self.cardinality
self.learned_weights = fg.factorGraphs[0].weight_value
weight, variable, factor, ftv, domain_mask, n_edges =\
self._compile(sparse.coo_matrix((1, n), L.dtype), init_deps,
init_class_prior, LF_acc_prior_weights, is_fixed,
[self.cardinality_for_stats])
variable["isEvidence"] = False
weight["isFixed"] = True
weight["initialValue"] = fg.factorGraphs[0].weight_value
fg.factorGraphs = []
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
self.fg = fg
self.nlf = n
self.cardinality = cardinality
n_learning_epoch=1000,
stepsize=0.01,
decay=0.95,
reg_param=1e-6,
regularization=2,
truncation=10,
quiet=(not False),
verbose=False,
learn_non_evidence=False, # need to test
sample_evidence=False,
burn_in=10,
nthreads=1)
subgraph = weight, variable, factor, fmap, domain_mask, edges
ns_learing.loadFactorGraph(*subgraph)
# 因子图参数学习
ns_learing.learning()
# 因子图推理
# 参数学习完成后将weight的isfixed属性置为true
for index, w in enumerate(weight):
w["isFixed"] = True
w['initialValue'] = ns_learing.factorGraphs[0].weight[index][
'initialValue']
ns_inference = NumbSkull(
n_inference_epoch=1000,
n_learning_epoch=1000,
stepsize=0.001,
decay=0.95,
reg_param=1e-6,
regularization=2,
truncation=10,
quiet=(not False),
def train(self, L, deps=(), LF_acc_prior_weights=None,
LF_acc_prior_weight_default=1, labels=None, label_prior_weight=5,
init_deps=0.0, init_class_prior=-1.0, epochs=30, step_size=None,
decay=1.0, reg_param=0.1, reg_type=2, verbose=False, truncation=10,
burn_in=5, cardinality=None, timer=None, candidate_ranges=None, threads=1):
"""
Fits the parameters of the model to a data set. By default, learns a
conditionally independent model. Additional unary dependencies can be
set to be included in the constructor. Additional pairwise and
higher-order dependencies can be included as an argument.
Results are stored as a member named weights, instance of
snorkel.learning.gen_learning.GenerativeModelWeights.
:param L: M x N csr_AnnotationMatrix-type label matrix, where there are
M candidates labeled by N labeling functions (LFs)
:param deps: collection of dependencies to include in the model, each
element is a tuple of the form
(LF 1 index, LF 2 index, dependency type),
see snorkel.learning.constants
:param LF_acc_prior_weights: An N-element list of prior weights for the
LF accuracies (log scale)
:param LF_acc_prior_weight_default: Default prior for the weight of each
LF accuracy; if LF_acc_prior_weights is unset, each LF will have
this accuracy prior weight (log scale)
:param labels: Optional ground truth labels
:param label_prior_weight: The prior probability that the ground truth
labels (if provided) are correct (log scale)
:param init_deps: initial weight for additional dependencies, except
class prior (log scale)
:param init_class_prior: initial class prior (in log scale), note only
used if class_prior=True in constructor
:param epochs: number of training epochs
:param step_size: gradient step size, default is 1 / L.shape[0]
:param decay: multiplicative decay of step size,
step_size_(t+1) = step_size_(t) * decay
:param reg_param: regularization strength
:param reg_type: 1 = L1 regularization, 2 = L2 regularization
:param verbose: whether to write debugging info to stdout
:param truncation: number of iterations between truncation step for L1
regularization
:param burn_in: number of burn-in samples to take before beginning
learning
:param cardinality: number of possible classes; by default is inferred
from the label matrix L
:param timer: stopwatch for profiling, must implement start() and end()
:param candidate_ranges: Optionally, a list of M sets of integer values,
representing the possible categorical values that each of the M
candidates can take. If a label is outside of this range throws an
error. If None, then each candidate can take any value from 0 to
cardinality.
:param threads: the number of threads to use for sampling. Default is 1.
"""
m, n = L.shape
step_size = step_size or 0.0001
# Check to make sure matrix is int-valued
element_type = type(L[0,0])
# Note: Other simpler forms of this check often don't work; still not
# sure why...
if not issubclass(element_type, np.integer):
raise ValueError("""Label matrix must have int-type elements,
but elements have type %s""" % element_type)
# Automatically infer cardinality
# Binary: Values in {-1, 0, 1} [Default]
# Categorical: Values in {0, 1, ..., K}
if cardinality is None:
# If candidate_ranges is provided, use this to determine cardinality
if candidate_ranges is not None:
cardinality = max(map(max, candidate_ranges))
else:
# This is just an annoying hack for LIL sparse matrices...
try:
lmax = L.max()
except AttributeError:
lmax = L.tocoo().max()
if lmax > 2:
cardinality = lmax
elif lmax < 2:
cardinality = 2
else:
raise ValueError(
"L.max() == %s, cannot infer cardinality." % lmax)
print("Inferred cardinality: %s" % cardinality)
self.cardinality = cardinality
# Priors for LFs default to fixed prior value
# NOTE: Setting default != 0.5 creates a (fixed) factor which increases
# runtime (by ~0.5x that of a non-fixed factor)...
if LF_acc_prior_weights is None:
LF_acc_prior_weights = [LF_acc_prior_weight_default for _ in range(n)]
else:
LF_acc_prior_weights = list(copy(LF_acc_prior_weights))
# LF weights are un-fixed
is_fixed = [False for _ in range(n)]
# If supervised labels are provided, add them as a fixed LF with prior
# Note: For large L this column stack operation could be very
# inefficient, can consider refactoring...
if labels is not None:
labels = labels.reshape(m, 1)
L = sparse.hstack([L, labels])
is_fixed.append(True)
LF_acc_prior_weights.append(label_prior_weight)
n += 1
# Reduce overhead of tracking indices by converting L to a CSR sparse matrix.
L = sparse.csr_matrix(L).copy()
# If candidate_ranges is provided, remap the values of L using
# candidate_ranges. This "scoped categorical" approach allows learning
# and inference to be efficient even with very large cardinality, as
# we only sample relevant values for each candidate. Also set
# per-candidate cardinalities according to candidate_ranges if not None,
# else as constant value.
self.cardinalities = self.cardinality * np.ones(m, dtype=np.int64)
self.candidate_ranges = candidate_ranges
if self.candidate_ranges is not None:
L, self.cardinalities, _ = self._remap_scoped_categoricals(L,
self.candidate_ranges)
# Shuffle the data points, cardinalities, and candidate_ranges
idxs = self.rng.permutation(list(range(m)))
L = L[idxs, :]
if candidate_ranges is not None:
self.cardinalities = self.cardinalities[idxs]
c_ranges_reshuffled = []
for i in idxs:
c_ranges_reshuffled.append(self.candidate_ranges[i])
self.candidate_ranges = c_ranges_reshuffled
# Compile factor graph
self._process_dependency_graph(L, deps)
weight, variable, factor, ftv, domain_mask, n_edges = self._compile(
L, init_deps, init_class_prior, LF_acc_prior_weights, is_fixed, self.cardinalities)
fg = NumbSkull(
n_inference_epoch=0,
n_learning_epoch=epochs,
stepsize=step_size,
decay=decay,
reg_param=reg_param,
regularization=reg_type,
truncation=truncation,
quiet=(not verbose),
verbose=verbose,
learn_non_evidence=True,
burn_in=burn_in,
nthreads=threads
)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
if timer is not None:
timer.start()
fg.learning(out=False)
if timer is not None:
timer.end()
self._process_learned_weights(L, fg, LF_acc_prior_weights, is_fixed)
# Store info from factor graph
if self.candidate_ranges is not None:
self.cardinality_for_stats = int(max(self.cardinalities))
else:
self.cardinality_for_stats = self.cardinality
self.learned_weights = fg.factorGraphs[0].weight_value
weight, variable, factor, ftv, domain_mask, n_edges =\
self._compile(sparse.coo_matrix((1, n), L.dtype), init_deps,
init_class_prior, LF_acc_prior_weights, is_fixed,
[self.cardinality_for_stats])
variable["isEvidence"] = False
weight["isFixed"] = True
weight["initialValue"] = fg.factorGraphs[0].weight_value
fg.factorGraphs = []
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
self.fg = fg
self.nlf = n
self.cardinality = cardinality
def train(self, L, deps=(), LF_acc_priors=None, LF_acc_features=None, LF_acc_prior_default=0.7,
labels=None, label_prior=0.99, init_deps=0.0,
init_class_prior=-1.0, epochs=30, step_size=None, decay=1.0,
reg_param=0.1, reg_type=2, verbose=False, truncation=10, burn_in=5,
cardinality=None, timer=None):
"""
Fits the parameters of the model to a data set. By default, learns a
conditionally independent model with featurized accuracies. Additional unary dependencies can be
set to be included in the constructor. Additional pairwise and
higher-order dependencies can be included as an argument.
Results are stored as a member named weights, instance of
snorkel.learning.gen_learning.GenerativeModelWeights.
:param L: M x N csr_AnnotationMatrix-type label matrix, where there are
M candidates labeled by N labeling functions (LFs)
:param deps: collection of dependencies to include in the model, each
element is a tuple of the form
(LF 1 index, LF 2 index, dependency type),
see snorkel.learning.constants
:param LF_acc_priors: An N-element list of prior probabilities for the
LF accuracies
:param LF_acc_features: An N-element list of features that determine
the accuracy of its labeling function; its labeling function has a single
feature; feature weights are coupled
:param LF_acc_prior_default: Default prior probability for each LF
accuracy; if LF_acc_priors is unset, each LF will have this prior
:param labels: Optional ground truth labels
:param label_prior: The prior probability that the ground truth labels
(if provided) are correct
:param init_deps: initial weight for additional dependencies, except
class prior (in log scale)
:param init_class_prior: initial class prior (in log scale), note only
used if class_prior=True in constructor
:param epochs: number of training epochs
:param step_size: gradient step size, default is 1 / L.shape[0]
:param decay: multiplicative decay of step size,
step_size_(t+1) = step_size_(t) * decay
:param reg_param: regularization strength
:param reg_type: 1 = L1 regularization, 2 = L2 regularization
:param verbose: whether to write debugging info to stdout
:param truncation: number of iterations between truncation step for L1
regularization
:param burn_in: number of burn-in samples to take before beginning
learning
:param cardinality: number of possible classes; by default is inferred
from the label matrix L
:param timer: stopwatch for profiling, must implement start() and end()
"""
m, n = L.shape
step_size = step_size or 0.0001
reg_param_scaled = reg_param / L.shape[0]
# Automatically infer cardinality
# Binary: Values in {-1, 0, 1} [Default]
# Categorical: Values in {0, 1, ..., K}
if cardinality is None:
# This is just an annoying hack for LIL sparse matrices...
try:
lmax = L.max()
except AttributeError:
lmax = L.tocoo().max()
if lmax > 2:
cardinality = lmax
elif lmax < 2:
cardinality = 2
else:
raise ValueError(
"L.max() == %s, cannot infer cardinality." % lmax)
print("Inferred cardinality: %s" % cardinality)
# Priors for LFs default to fixed prior value
# NOTE: Setting default != 0.5 creates a (fixed) factor which increases
# runtime (by ~0.5x that of a non-fixed factor)...
if LF_acc_priors is None:
LF_acc_priors = [LF_acc_prior_default for _ in range(n)]
else:
LF_acc_priors = list(copy(LF_acc_priors))
if LF_acc_features is None:
LF_acc_features = [str(i) for i in range(n)]
else:
LF_acc_features = list(copy(LF_acc_features))
# LF weights are un-fixed
is_fixed = [False for _ in range(n)]
# If supervised labels are provided, add them as a fixed LF with prior
# Note: For large L this column stack operation could be very
# inefficient, can consider refactoring...
if labels is not None:
labels = labels.reshape(m, 1)
L = sparse.hstack([L.copy(), labels])
is_fixed.append(True)
LF_acc_priors.append(label_prior)
n += 1
# Shuffle the data points
idxs = range(m)
np.random.shuffle(idxs)
if not isinstance(L, sparse.csr_matrix):
L = sparse.csr_matrix(L)
L = L[idxs, :]
# Compile factor graph
self._process_dependency_graph(L, deps)
weight, variable, factor, ftv, domain_mask, n_edges, feature2WoffMap =\
self._compile(L, init_deps, init_class_prior, LF_acc_priors, LF_acc_features,
is_fixed, cardinality)
fg = NumbSkull(
n_inference_epoch=0,
n_learning_epoch=epochs,
stepsize=step_size,
decay=decay,
reg_param=reg_param_scaled,
regularization=reg_type,
truncation=truncation,
quiet=(not verbose),
verbose=verbose,
learn_non_evidence=True,
burn_in=burn_in
)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
if timer is not None:
timer.start()
fg.learning(out=False)
if timer is not None:
timer.end()
self._process_learned_weights(L, fg, LF_acc_priors, LF_acc_features, feature2WoffMap, is_fixed)
# Store info from factor graph
weight, variable, factor, ftv, domain_mask, n_edges, feature2WoffMap =\
self._compile(sparse.coo_matrix((1, n), L.dtype), init_deps,
init_class_prior, LF_acc_priors, LF_acc_features, is_fixed, cardinality)
variable["isEvidence"] = False
weight["isFixed"] = True
weight["initialValue"] = fg.factorGraphs[0].weight_value
fg.factorGraphs = []
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
self.fg = fg
self.nlf = n
self.cardinality = cardinality
def train(self,
L,
deps=(),
LF_acc_priors=None,
LF_acc_features=None,
LF_acc_prior_default=0.7,
labels=None,
label_prior=0.99,
init_deps=0.0,
init_class_prior=-1.0,
epochs=30,
step_size=None,
decay=1.0,
reg_param=0.1,
reg_type=2,
verbose=False,
truncation=10,
burn_in=5,
cardinality=None,
timer=None):
"""
Fits the parameters of the model to a data set. By default, learns a
conditionally independent model with featurized accuracies. Additional unary dependencies can be
set to be included in the constructor. Additional pairwise and
higher-order dependencies can be included as an argument.
Results are stored as a member named weights, instance of
snorkel.learning.gen_learning.GenerativeModelWeights.
:param L: M x N csr_AnnotationMatrix-type label matrix, where there are
M candidates labeled by N labeling functions (LFs)
:param deps: collection of dependencies to include in the model, each
element is a tuple of the form
(LF 1 index, LF 2 index, dependency type),
see snorkel.learning.constants
:param LF_acc_priors: An N-element list of prior probabilities for the
LF accuracies
:param LF_acc_features: An N-element list of features that determine
the accuracy of its labeling function; its labeling function has a single
feature; feature weights are coupled
:param LF_acc_prior_default: Default prior probability for each LF
accuracy; if LF_acc_priors is unset, each LF will have this prior
:param labels: Optional ground truth labels
:param label_prior: The prior probability that the ground truth labels
(if provided) are correct
:param init_deps: initial weight for additional dependencies, except
class prior (in log scale)
:param init_class_prior: initial class prior (in log scale), note only
used if class_prior=True in constructor
:param epochs: number of training epochs
:param step_size: gradient step size, default is 1 / L.shape[0]
:param decay: multiplicative decay of step size,
step_size_(t+1) = step_size_(t) * decay
:param reg_param: regularization strength
:param reg_type: 1 = L1 regularization, 2 = L2 regularization
:param verbose: whether to write debugging info to stdout
:param truncation: number of iterations between truncation step for L1
regularization
:param burn_in: number of burn-in samples to take before beginning
learning
:param cardinality: number of possible classes; by default is inferred
from the label matrix L
:param timer: stopwatch for profiling, must implement start() and end()
"""
m, n = L.shape
step_size = step_size or 0.0001
reg_param_scaled = reg_param / L.shape[0]
# Automatically infer cardinality
# Binary: Values in {-1, 0, 1} [Default]
# Categorical: Values in {0, 1, ..., K}
if cardinality is None:
# This is just an annoying hack for LIL sparse matrices...
try:
lmax = L.max()
except AttributeError:
lmax = L.tocoo().max()
if lmax > 2:
cardinality = lmax
elif lmax < 2:
cardinality = 2
else:
raise ValueError("L.max() == %s, cannot infer cardinality." %
lmax)
print("Inferred cardinality: %s" % cardinality)
# Priors for LFs default to fixed prior value
# NOTE: Setting default != 0.5 creates a (fixed) factor which increases
# runtime (by ~0.5x that of a non-fixed factor)...
if LF_acc_priors is None:
LF_acc_priors = [LF_acc_prior_default for _ in range(n)]
else:
LF_acc_priors = list(copy(LF_acc_priors))
if LF_acc_features is None:
LF_acc_features = [str(i) for i in range(n)]
else:
LF_acc_features = list(copy(LF_acc_features))
# LF weights are un-fixed
is_fixed = [False for _ in range(n)]
# If supervised labels are provided, add them as a fixed LF with prior
# Note: For large L this column stack operation could be very
# inefficient, can consider refactoring...
if labels is not None:
labels = labels.reshape(m, 1)
L = sparse.hstack([L.copy(), labels])
is_fixed.append(True)
LF_acc_priors.append(label_prior)
n += 1
# Shuffle the data points
idxs = range(m)
np.random.shuffle(idxs)
if not isinstance(L, sparse.csr_matrix):
L = sparse.csr_matrix(L)
L = L[idxs, :]
# Compile factor graph
self._process_dependency_graph(L, deps)
weight, variable, factor, ftv, domain_mask, n_edges, feature2WoffMap =\
self._compile(L, init_deps, init_class_prior, LF_acc_priors, LF_acc_features,
is_fixed, cardinality)
fg = NumbSkull(n_inference_epoch=0,
n_learning_epoch=epochs,
stepsize=step_size,
decay=decay,
reg_param=reg_param_scaled,
regularization=reg_type,
truncation=truncation,
quiet=(not verbose),
verbose=verbose,
learn_non_evidence=True,
burn_in=burn_in)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
if timer is not None:
timer.start()
fg.learning(out=False)
if timer is not None:
timer.end()
self._process_learned_weights(L, fg, LF_acc_priors, LF_acc_features,
feature2WoffMap, is_fixed)
# Store info from factor graph
weight, variable, factor, ftv, domain_mask, n_edges, feature2WoffMap =\
self._compile(sparse.coo_matrix((1, n), L.dtype), init_deps,
init_class_prior, LF_acc_priors, LF_acc_features, is_fixed, cardinality)
variable["isEvidence"] = False
weight["isFixed"] = True
weight["initialValue"] = fg.factorGraphs[0].weight_value
fg.factorGraphs = []
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
self.fg = fg
self.nlf = n
self.cardinality = cardinality
def train(self,
L,
y=None,
deps=(),
init_acc=1.0,
init_deps=1.0,
init_class_prior=-1.0,
epochs=10,
step_size=None,
decay=0.99,
reg_param=0.1,
reg_type=2,
verbose=False,
truncation=10,
burn_in=5,
timer=None):
"""
Fits the parameters of the model to a data set. By default, learns a conditionally independent model.
Additional unary dependencies can be set to be included in the constructor. Additional pairwise and higher-order
dependencies can be included as an argument.
Results are stored as a member named weights, instance of snorkel.learning.gen_learning.GenerativeModelWeights.
:param L: labeling function output matrix
:param y: optional ground truth labels
:param deps: collection of dependencies to include in the model, each element is a tuple of the form
(LF 1 index, LF 2 index, dependency type), see snorkel.learning.constants
:param init_acc: initial weight for accuracy dependencies (in log scale)
:param init_deps: initial weight for additional dependencies, except class prior (in log scale)
:param init_class_prior: initial class prior (in log scale), note only used if class_prior=True in constructor
:param epochs: number of training epochs
:param step_size: gradient step size, default is 1 / L.shape[0]
:param decay: multiplicative decay of step size, step_size_(t+1) = step_size_(t) * decay
:param reg_param: regularization strength
:param reg_type: 1 = L1 regularization, 2 = L2 regularization
:param verbose: whether to write debugging info to stdout
:param truncation: number of iterations between truncation step for L1 regularization
:param burn_in: number of burn-in samples to take before beginning learning
:param timer: stopwatch for profiling, must implement start() and end()
"""
step_size = step_size or 1.0 / L.shape[0]
reg_param_scaled = reg_param / L.shape[0]
self._process_dependency_graph(L, deps)
weight, variable, factor, ftv, domain_mask, n_edges = self._compile(
L, y, init_acc, init_deps, init_class_prior)
fg = NumbSkull(n_inference_epoch=0,
n_learning_epoch=epochs,
stepsize=step_size,
decay=decay,
reg_param=reg_param_scaled,
regularization=reg_type,
truncation=truncation,
quiet=(not verbose),
verbose=verbose,
learn_non_evidence=True,
burn_in=burn_in)
fg.loadFactorGraph(weight, variable, factor, ftv, domain_mask, n_edges)
if timer is not None:
timer.start()
fg.learning(out=False)
if timer is not None:
timer.end()
self._process_learned_weights(L, fg)