def global_position(self):
#don't need to move origin to get rotation as this translation is implicit in the other measurements
#change so only executed once when the local aoa is set.
if isinf(self.local_position):
return_value = -inf
else:
# This linear algebra is nice but actually all that we need to do is add the delta between the global reading
# and the local reading to get the correct result should be less expensive in terms of calculation as well
return_value = modulo_heading(
self.local_position + self.sensor_properties.location.heading)
#local_sensor_position = array([sqrt(2) * cos(self.local_aoa), sqrt(2) * sin(self.local_aoa)])
#global_particle_position = dot(linalg.inv(self.sensor_properties.rotation_matrix), local_sensor_position)
#return_value = cartesian_to_polar(global_particle_position[0][0],global_particle_position[0][1],
# response_units=AngleUnits.radians)[1]
return return_value
def global_position(self):
#don't need to move origin to get rotation as this translation is implicit in the other measurements
#change so only executed once when the local aoa is set.
if isinf(self.local_position):
return_value = -inf
else:
# This linear algebra is nice but actually all that we need to do is add the delta between the global reading
# and the local reading to get the correct result should be less expensive in terms of calculation as well
return_value = modulo_heading(self.local_position + self.sensor_properties.location.heading)
#local_sensor_position = array([sqrt(2) * cos(self.local_aoa), sqrt(2) * sin(self.local_aoa)])
#global_particle_position = dot(linalg.inv(self.sensor_properties.rotation_matrix), local_sensor_position)
#return_value = cartesian_to_polar(global_particle_position[0][0],global_particle_position[0][1],
# response_units=AngleUnits.radians)[1]
return return_value
def local_position(self, arg_aoa):
if isinf(arg_aoa):
self._aoa = -inf
else:
self._aoa = modulo_heading(arg_aoa)
def isinf(x):
if isinstance(x, IDataDescriptor):
return _um.isinf(dd_as_py(x))
else:
return _um.isinf(x)
def get_format_func(self, elem, **options):
missing_opt = self.check_options(**options)
if missing_opt:
raise Exception("Missing options: {}".format(missing_opt))
floatmode = options['floatmode']
precision = None if floatmode == 'unique' else options['precision']
suppress_small = options['suppress_small']
sign = options['sign']
infstr = options['infstr']
nanstr = options['nanstr']
exp_format = False
pad_left, pad_right = 0, 0
# only the finite values are used to compute the number of digits
finite = umath.isfinite(elem)
finite_vals = elem[finite]
nonfinite_vals = elem[~finite]
# choose exponential mode based on the non-zero finite values:
abs_non_zero = umath.absolute(finite_vals[finite_vals != 0])
if len(abs_non_zero) != 0:
max_val = np.max(abs_non_zero)
min_val = np.min(abs_non_zero)
with np.errstate(over='ignore'): # division can overflow
if max_val >= 1.e8 or (not suppress_small and
(min_val < 0.0001
or max_val / min_val > 1000.)):
exp_format = True
# do a first pass of printing all the numbers, to determine sizes
if len(finite_vals) == 0:
trim, exp_size, unique = '.', -1, True
elif exp_format:
trim, unique = '.', True
if floatmode == 'fixed':
trim, unique = 'k', False
strs = (format_float_scientific(x,
precision=precision,
unique=unique,
trim=trim,
sign=sign == '+')
for x in finite_vals)
frac_strs, _, exp_strs = zip(*(s.partition('e') for s in strs))
int_part, frac_part = zip(*(s.split('.') for s in frac_strs))
exp_size = max(len(s) for s in exp_strs) - 1
trim = 'k'
precision = max(len(s) for s in frac_part)
# this should be only 1 or 2. Can be calculated from sign.
pad_left = max(len(s) for s in int_part)
# pad_right is only needed for nan length calculation
pad_right = exp_size + 2 + precision
unique = False
else:
trim, unique = '.', True
if floatmode == 'fixed':
trim, unique = 'k', False
strs = (format_float_positional(x,
precision=precision,
fractional=True,
unique=unique,
trim=trim,
sign=sign == '+')
for x in finite_vals)
int_part, frac_part = zip(*(s.split('.') for s in strs))
pad_left = max(len(s) for s in int_part)
pad_right = max(len(s) for s in frac_part)
exp_size = -1
if floatmode in ['fixed', 'maxprec_equal']:
precision = pad_right
unique = False
trim = 'k'
else:
unique = True
trim = '.'
# account for sign = ' ' by adding one to pad_left
if sign == ' ' and not any(np.signbit(finite_vals)):
pad_left += 1
# account for nan and inf in pad_left
if len(nonfinite_vals) != 0:
nanlen, inflen = 0, 0
if np.any(umath.isinf(nonfinite_vals)):
neginf = sign != '-' or np.any(np.isneginf(nonfinite_vals))
inflen = len(infstr) + neginf
if np.any(umath.isnan(elem)):
nanlen = len(nanstr)
offset = pad_right + 1 # +1 for decimal pt
pad_left = max(nanlen - offset, inflen - offset, pad_left)
def print_nonfinite(x):
with errstate(invalid='ignore'):
if umath.isnan(x):
ret = ('+' if sign == '+' else '') + nanstr
else: # isinf
infsgn = '-' if x < 0 else '+' if sign == '+' else ''
ret = infsgn + infstr
return ' ' * (pad_left + pad_right + 1 - len(ret)) + ret
if exp_format:
def print_finite(x):
return format_float_scientific(x,
precision=precision,
unique=unique,
trim=trim,
sign=sign == '+',
pad_left=pad_left,
exp_digits=exp_size)
else:
def print_finite(x):
return format_float_positional(x,
precision=precision,
unique=unique,
fractional=True,
trim=trim,
sign=sign == '+',
pad_left=pad_left,
pad_right=pad_right)
def fmt(x):
if umath.isfinite(x):
return print_finite(x)
else:
return print_nonfinite(x)
return fmt
def isinf(x):
if isinstance(x, IDataDescriptor):
return _um.isinf(dd_as_py(x))
else:
return _um.isinf(x)
def isinf(x):
if isinstance(x, DDesc):
return _um.isinf(ddesc_as_py(x))
else:
return _um.isinf(x)
def sample(self, num_iterations):
"""
for num_iters:
for each customer
remove him from his old_table and update the table params.
if old_table is empty:
remove table
Calculate prior and likelihood for this customer sitting at each table
sample for a table index
if new_table is equal to old_table
don't have to update the parameters
else update params of the old table.
"""
if self.show_topics is not None:
print("Topics after initialization")
print(self.format_topics())
# Compute the overall usage of topics across the training corpus
topic_props = self.table_counts_per_doc.sum(axis=1).astype(
np.float64)
topic_props /= topic_props.sum()
print("Words using topics: {}".format(", ".join(
"{}={:.1f}%".format(i, prop)
for i, prop in enumerate(topic_props * 100.))))
topic_doc_props = (self.table_counts_per_doc > 0).astype(
np.float64).sum(axis=1)
topic_doc_props /= self.num_documents
print("Docs using topics: {}".format(", ".join(
"{}={:.1f}%".format(i, prop)
for i, prop in enumerate(topic_doc_props * 100.))))
with VoseAliasUpdater(
self.aliases,
self.vocab_embeddings,
self.prior.kappa,
self.prior.nu,
self.table_counts,
self.table_means,
self.table_cholesky_ltriangular_mat,
self.log_determinants,
das_normalization=self.das_normalization,
) as alias_updater:
for iteration in range(num_iterations):
stats = SamplingDiagnostics()
self.log.info("Iteration {}".format(iteration))
alias_updater.unpause()
pbar = get_progress_bar(len(self.corpus),
title="Sampling",
show_progress=self.show_progress)
for d, doc in enumerate(pbar(self.corpus)):
if self.show_topics is not None and self.show_topics > 0 and d % self.show_topics == 0:
print("Topics after {:,} docs".format(d))
print(self.format_topics())
for w, cust_id in enumerate(doc):
x = self.vocab_embeddings[cust_id]
# Remove custId from his old_table
old_table_id = self.table_assignments[d][w]
self.table_assignments[d][
w] = -1 # Doesn't really make any difference, as only counts are used
with self.table_counts.lock:
self.table_counts.np[old_table_id] -= 1
self.table_counts_per_doc[old_table_id, d] -= 1
# Update vector means etc
self.sum_squared_table_customers[
old_table_id] -= np.outer(x, x)
# Topic 'old_table_id' now has one member fewer
# Just update params for this customer
self.update_table_params(old_table_id,
cust_id,
is_removed=True)
# Under the alias method, we only do the full likelihood computation for topics
# that already have a non-zero count in the current document
non_zero_tables = np.where(
self.table_counts_per_doc[:, d] > 0)[0]
if len(non_zero_tables) == 0:
# If there's only one word in a doc, there are no topics to compute the full posterior for
no_non_zero = True
else:
no_non_zero = False
# We only compute the posterior for these topics
log_priors = np.log(
self.table_counts_per_doc[non_zero_tables, d])
log_likelihoods = np.zeros(len(non_zero_tables),
dtype=np.float32)
for nz_table, table in enumerate(non_zero_tables):
log_likelihoods[
nz_table] = self.log_multivariate_tdensity(
x, table)
log_posterior = log_priors + log_likelihoods
# To prevent overflow, subtract by log(p_max)
max_log_posterior = log_posterior.max()
scaled_posterior = log_posterior - max_log_posterior
if self.das_normalization:
# Not doing this now, but following what the Java impl does, however odd that seems
psum = np.sum(np.exp(scaled_posterior))
else:
# Java impl subtracts max before computing psum, but this seems to be wrong
# We still subtract first, but then multiply by the max prob afterwards
psum = np.exp(
np.log(np.sum(np.exp(scaled_posterior))) +
max_log_posterior)
# Now just use the scaled log posterior in the same way as in the Java impl
# They have a bin-search method for sampling from the cumulative dist,
# but we simply normalize and use Numpy to sample
unnormed_posterior = np.exp(scaled_posterior)
normed_posterior = unnormed_posterior / unnormed_posterior.sum(
)
# Don't let the alias parameters get updated in the middle of the sampling
self.aliases.lock.acquire_read(cust_id)
select_pr = psum / (
psum + self.alpha *
self.aliases.likelihood_sum.np[cust_id])
# MHV to draw new topic
# Take a number of Metropolis-Hastings samples
current_sample = old_table_id
# Calculate the true likelihood of this word under the current sample,
# for calculating acceptance prob
current_sample_log_prob = self.log_multivariate_tdensity(
x, current_sample)
for r in range(self.mh_steps):
# 1. Flip a coin
if not no_non_zero and np.random.sample(
) < select_pr:
# Choose from the exactly computed posterior dist, only allowing
# topics already sampled in the doc
temp = np.random.choice(len(non_zero_tables),
p=normed_posterior)
new_sample = non_zero_tables[temp]
stats.log_select_pr(True, select_pr)
else:
# Choose from the alias, allowing any topic but using slightly
# out-of-date likelihoods
new_sample = self.aliases.sample_vose(cust_id)
stats.log_select_pr(False, select_pr)
if new_sample != current_sample:
# 2. Find acceptance probability
new_sample_log_prob = self.log_multivariate_tdensity(
x, new_sample)
# This can sometimes generate an overflow warning from Numpy
# We don't care, though: in that case acceptance > 1., so we always accept
with np.errstate(over="ignore"):
# From my reading of:
# Li et al. (2014): Reducing the sampling complexity of topic models
# the acceptance probability should be as follows:
acceptance = \
(self.table_counts_per_doc[new_sample, d] + self.alpha) / \
(self.table_counts_per_doc[current_sample, d] + self.alpha) * \
np.exp(new_sample_log_prob - current_sample_log_prob) * \
(self.table_counts_per_doc[current_sample, d]*np.exp(current_sample_log_prob) +
self.alpha*np.exp(self.aliases.log_likelihoods.np[cust_id, current_sample])) / \
(self.table_counts_per_doc[new_sample, d]*np.exp(new_sample_log_prob) +
self.alpha*np.exp(self.aliases.log_likelihoods.np[cust_id, new_sample]))
# The Java implementation, however, does this:
#acceptance = \
# (self.table_counts_per_doc[new_table_id, d] + self.alpha) / \
# (self.table_counts_per_doc[current_sample, d] + self.alpha) * \
# np.exp(new_prob - old_prob) * \
# (self.table_counts_per_doc[current_sample, d]*old_log_prob +
# self.alpha*alias.w.np[current_sample]) / \
# (self.table_counts_per_doc[new_table_id, d]*new_log_prob +
# self.alpha*alias.w.np[new_table_id])
# The difference is the Java impl doesn't exp the log likelihood in the last
# fraction, i.e. it uses a log prob instead of a prob
# 3. Compare against uniform[0,1]
# If the acceptance prob > 1, we always accept: this means the new sample
# has a higher probability than the old
if isinf(
acceptance
) or acceptance >= 1. or np.random.sample(
) < acceptance:
# No need to sample if acceptance >= 1
# If the acceptance prob < 1, sample whether to accept or not, such that
# the more likely the new sample is compared to the old, the more likely we
# are to keep it
current_sample = new_sample
current_sample_log_prob = new_sample_log_prob
stats.log_acceptance(True, acceptance)
else:
stats.log_acceptance(False, acceptance)
# NOTE: There seems to be a small error in the Java implementation here
# On the last MH step, it doesn't make any difference whether we accept the
# sample or not - we always end up using it
self.aliases.lock.release_read()
if current_sample == old_table_id:
stats.log_sampled_same()
else:
stats.log_sampled_different()
# Now have a new assignment: add its counts
self.table_assignments[d][w] = current_sample
with self.table_counts.lock:
self.table_counts.np[current_sample] += 1
self.table_counts_per_doc[current_sample, d] += 1
self.sum_squared_table_customers[
current_sample] += np.outer(x, x)
self.update_table_params(current_sample, cust_id)
# Pause the alias updater until we start the next iteration
alias_updater.pause()
# Output some useful stats about sampling
if stats.acceptance_used():
self.log.info(
"Acceptance rate = {:.2f}%, mean acceptance: {:.2f} ({:,} samples draw)"
.format(stats.acceptance_rate() * 100.,
stats.mean_acceptance(),
stats.acceptance_samples()))
else:
self.log.info("No new samples drawn")
self.log.info(
"Prior select rate = {:.2f}%, mean select_pr = {:.2f}".
format(stats.select_pr_rate() * 100.,
stats.mean_select_pr()))
self.log.info("Chose new sample: {:.2f}%".format(
stats.sample_change_rate() * 100.))
if self.show_topics is not None:
print("Topics after iteration {}".format(iteration))
print(self.format_topics())
# Compute the overall usage of topics across the training corpus
topic_props = self.table_counts_per_doc.sum(axis=1).astype(
np.float64)
topic_props /= topic_props.sum()
print("Words using topics: {}".format(", ".join(
"{}={:.1f}%".format(i, prop)
for i, prop in enumerate(topic_props * 100.))))
topic_doc_props = (self.table_counts_per_doc > 0).astype(
np.float64).sum(axis=1)
topic_doc_props /= self.num_documents
print("Docs using topics: {}".format(", ".join(
"{}={:.1f}%".format(i, prop)
for i, prop in enumerate(topic_doc_props * 100.))))
if self.save_path is not None:
self.log.info("Saving model")
self.save()
def local_position(self, arg_aoa):
if isinf(arg_aoa):
self._aoa = -inf
else:
self._aoa = modulo_heading(arg_aoa)
def isinf(x):
if isinstance(x, DDesc):
return _um.isinf(ddesc_as_py(x))
else:
return _um.isinf(x)
