Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix epsilon according to dtype in LdaModel #1770

Merged
merged 5 commits into from
Dec 7, 2017
Merged

Fix epsilon according to dtype in LdaModel #1770

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
db1ce0f
add type check + adopt eps for different dtypes
menshikh-iv Dec 7, 2017
5f56a1c
fix typo
menshikh-iv Dec 7, 2017
488dfd1
improve readability
menshikh-iv Dec 7, 2017
e4e96e4
move mapping to module level, remove duplication
menshikh-iv Dec 7, 2017
3d46db8
fix typo + remove duplication
menshikh-iv Dec 7, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 16 additions & 3 deletions gensim/models/ldamodel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,12 @@

logger = logging.getLogger('gensim.models.ldamodel')

DTYPE_TO_EPS = {
np.float16: 1e-5,
np.float32: 1e-35,
np.float64: 1e-100,
}


def logsumexp(x):
"""Log of sum of exponentials
Expand Down Expand Up @@ -275,6 +281,7 @@ def __init__(self, corpus=None, num_topics=100, id2word=None,
`callbacks` a list of metric callbacks to log/visualize evaluation metrics of topic model during training.

`dtype` is data-type to use during calculations inside model. All inputs are also converted to this dtype.
Available types: `numpy.float16`, `numpy.float32`, `numpy.float64`.

Example:

Expand All @@ -286,6 +293,11 @@ def __init__(self, corpus=None, num_topics=100, id2word=None,
>>> lda = LdaModel(corpus, num_topics=50, alpha='auto', eval_every=5) # train asymmetric alpha from data

"""
if dtype not in DTYPE_TO_EPS:
raise ValueError(
"Incorrect 'dtype', please choose one of {}".format(
", ".join("numpy.{}".format(tp.__name__) for tp in sorted(DTYPE_TO_EPS))))

self.dtype = dtype

# store user-supplied parameters
Expand Down Expand Up @@ -497,8 +509,9 @@ def inference(self, chunk, collect_sstats=False):

# The optimal phi_{dwk} is proportional to expElogthetad_k * expElogbetad_w.
# phinorm is the normalizer.
# TODO treat zeros explicitly, instead of adding 1e-100?
phinorm = np.dot(expElogthetad, expElogbetad) + 1e-100
# TODO treat zeros explicitly, instead of adding epsilon?
eps = DTYPE_TO_EPS[self.dtype]
phinorm = np.dot(expElogthetad, expElogbetad) + eps

# Iterate between gamma and phi until convergence
for _ in xrange(self.iterations):
Expand All @@ -509,7 +522,7 @@ def inference(self, chunk, collect_sstats=False):
gammad = self.alpha + expElogthetad * np.dot(cts / phinorm, expElogbetad.T)
Elogthetad = dirichlet_expectation(gammad)
expElogthetad = np.exp(Elogthetad)
phinorm = np.dot(expElogthetad, expElogbetad) + 1e-100
phinorm = np.dot(expElogthetad, expElogbetad) + eps
# If gamma hasn't changed much, we're done.
meanchange = np.mean(abs(gammad - lastgamma))
if meanchange < self.gamma_threshold:
Expand Down