def _do_chain(args):
"""
Generate a synthetic documen set from the teaching model.
Return the documents and the acceptance rate of the sampler
"""
docs_r = args[0]
docs_t = args[1] # start state for teaching
topics = args[2] # target topics
α = args[3]
β = args[4]
ρ = args[5]
n_samples = args[6]
n_iter = args[7]
seed = args[8]
chain = args[9]
n_lda_steps = 500
n_words = len(topics[0])
docs_t = _jitter_docs(docs_t, 1.0, n_words)
n_acpt = 0
acr = 0
lp = teach_lda_pgis(docs_t, topics, alpha=α, beta=β, n_samples=n_samples)
# lp = teach_lda_exact(docs_t, topics, alpha=α, beta=β)
for i in range(n_iter):
docs_p = _jitter_docs(docs_t, ρ, n_words)
lp_p = teach_lda_pgis(docs_p, topics, alpha=α, beta=β,
n_samples=n_samples)
# lp_p = teach_lda_exact(docs_p, topics, alpha=α, beta=β)
if log(random.random()) < lp_p - lp:
n_acpt += 1
docs_t = docs_p
lp = lp_p
acr = n_acpt/float(i+1)
err_t = _get_err(docs_t, topics, n_words, α, β, n_lda_steps, seed)
err_r = _get_err(docs_r, topics, n_words, α, β, n_lda_steps, seed)
print(".", end="")
sys.stdout.flush()
return docs_t, acr, err_t, err_r, chain
def evidence(docs, n_topics, n_words, beta=.1, n_samples=10000, seed=1337):
""" estimate the evidence using sequental importance sampling """
# generate fake topics. Remember, teach_lda_XXX calcualtes the teaching
# probability; we're onlt concerned with the evidence
# TODO: Pass m as an argument
m = 50.
alpha = m / n_topics
if USE_HARMONIC_MEAN:
evidence = evidence_lda_hm(docs,
n_topics,
n_words,
alpha=alpha,
beta=beta,
n_samples=n_samples,
stop_itr=1000,
seed=None)
else:
topics = np.ones((n_topics, n_words)) / n_words
_, _, evx = teach_lda_pgis(docs,
topics,
alpha=alpha,
beta=beta,
n_samples=n_samples,
seed=seed,
return_parts=True)
# the evidence isn't normalized in pgis because the normalizer cancels
# out of the ratio. Normalize here.
evidence = logsumexp(evx) - log(n_samples)
return evidence
def expitr(n_topics, n_vocab, n_words, α, β, n_docs, tol, seed, tasknum):
random.seed(seed)
print('Running %d topics, %d n_vocab, and %d words to ε=%1.2f.' %
(n_topics, n_vocab, n_words, tol,))
docs, topics = gen_docs(n_docs, n_topics, n_vocab, n_words, α, β)
i = 0
t_pgis = 0
relerr = float('Inf')
weights = []
while relerr > tol and i < MAX_ITR:
i += 1
seed = random.randrange(2**31)
t_start = time.time()
_, _, lml = teach_lda_pgis(docs, topics, alpha=α, beta=β, n_samples=1,
seed=seed, return_parts=True)
t_end = time.time()
t_pgis += t_end - t_start
weights.append(lml)
if i > 1:
relerr = isrelerr(weights)
if i > MAX_ITR:
print('Warning: MAX_ITR reached. Breaking.')
return t_pgis, i, n_topics, n_vocab, n_words, α, β, n_docs
def _calc_docset_logp_inner(docs, topics, topics_list, α, β, n_samples, seed):
pt = teach_lda_pgis(docs, topics, topics_list=topics_list, alpha=α, beta=β,
n_samples=n_samples, seed=seed, return_parts=True)
prior, numer, denom = pt
lp = prior + logsumexp(numer) - logsumexp(denom)
ll = logsumexp(numer) - np.log(len(numer))
print(".", end="")
sys.stdout.flush()
return lp, ll
def _calc_docset_logp_inner(docs, topics, topics_list, α, β, n_samples, seed):
pt = teach_lda_pgis(docs,
topics,
topics_list=topics_list,
alpha=α,
beta=β,
n_samples=n_samples,
seed=seed,
return_parts=True)
prior, numer, denom = pt
lp = prior + logsumexp(numer) - logsumexp(denom)
ll = logsumexp(numer) - np.log(len(numer))
print(".", end="")
sys.stdout.flush()
return lp, ll
def _expitr(docs, dtpc, alpha, beta, n_samples, seed):
logsamples = log(n_samples)
n_topics = len(dtpc)
n_words = len(dtpc[0])
exact = logsumexp(
teach_lda_exact(docs, dtpc, alpha=alpha, beta=beta,
return_parts=True)[-1])
print("1")
unis = logsumexp(
teach_lda_unis(docs,
dtpc,
alpha=alpha,
beta=beta,
n_samples=n_samples,
seed=seed,
return_parts=True)[-1]) - logsamples
print("2")
unis += sum(len(d['w']) for d in docs) * log(n_topics)
pgis = logsumexp(
teach_lda_pgis(docs,
dtpc,
alpha=alpha,
beta=beta,
n_samples=n_samples,
seed=seed + 1,
return_parts=True)[-1]) - logsamples
print("3")
hm = evidence_lda_hm(docs,
n_topics,
n_words,
alpha=alpha,
beta=beta,
n_samples=n_samples,
stop_itr=100,
seed=seed + 2)
print("4")
print(seed)
return {'Exact': exact, 'Uniform': unis, 'SIS': pgis, 'HM': hm}
def expitr(n_topics, n_vocab, n_words, α, β, n_docs, tol, seed, tasknum):
random.seed(seed)
print('Running %d topics, %d n_vocab, and %d words to ε=%1.2f.' % (
n_topics,
n_vocab,
n_words,
tol,
))
docs, topics = gen_docs(n_docs, n_topics, n_vocab, n_words, α, β)
i = 0
t_pgis = 0
relerr = float('Inf')
weights = []
while relerr > tol and i < MAX_ITR:
i += 1
seed = random.randrange(2**31)
t_start = time.time()
_, _, lml = teach_lda_pgis(docs,
topics,
alpha=α,
beta=β,
n_samples=1,
seed=seed,
return_parts=True)
t_end = time.time()
t_pgis += t_end - t_start
weights.append(lml)
if i > 1:
relerr = isrelerr(weights)
if i > MAX_ITR:
print('Warning: MAX_ITR reached. Breaking.')
return t_pgis, i, n_topics, n_vocab, n_words, α, β, n_docs
def evidence(docs, n_topics, n_words, beta=.1, n_samples=10000, seed=1337):
""" estimate the evidence using sequental importance sampling """
# generate fake topics. Remember, teach_lda_XXX calcualtes the teaching
# probability; we're onlt concerned with the evidence
# TODO: Pass m as an argument
m = 50.
alpha = m/n_topics
if USE_HARMONIC_MEAN:
evidence = evidence_lda_hm(docs, n_topics, n_words, alpha=alpha,
beta=beta, n_samples=n_samples,
stop_itr=1000, seed=None)
else:
topics = np.ones((n_topics, n_words))/n_words
_, _, evx = teach_lda_pgis(docs, topics, alpha=alpha, beta=beta,
n_samples=n_samples, seed=seed,
return_parts=True)
# the evidence isn't normalized in pgis because the normalizer cancels
# out of the ratio. Normalize here.
evidence = logsumexp(evx) - log(n_samples)
return evidence
def _expitr(docs, dtpc, alpha, beta, n_samples, seed):
n_topics = len(dtpc)
n_words = len(dtpc[0])
unis = np.array(teach_lda_unis(docs, dtpc, alpha=alpha, beta=beta,
n_samples=n_samples, seed=seed,
return_parts=True)[-1])
unis += sum([len(d['w']) for d in docs])*log(n_topics)
pgis = np.array(teach_lda_pgis(docs, dtpc, alpha=alpha, beta=beta,
n_samples=n_samples, seed=seed+1,
return_parts=True)[-1])
hm = evidence_lda_hm(docs, n_topics, n_words, alpha=alpha, beta=beta,
n_samples=n_samples, stop_itr=200, seed=seed+2,
return_samples=True)
lsxaccum = lambda x: logsumexp(x) - log(len(x))
hmaccum = lambda x: -(logsumexp(x) - log(len(x)))
data = []
steps = np.unique(np.array(np.exp(np.linspace(
log(2), log(n_samples-1), 250))))
for i in steps:
n_samples = float(i+1)
data.append({'Samples': n_samples,
'Type': 'Uniform',
'Estimate': lsxaccum(unis[:i]),
'Relative error': isrelerr(unis[:i]), })
data.append({'Samples': n_samples,
'Type': 'SIS',
'Estimate': lsxaccum(pgis[:i]),
'Relative error': isrelerr(pgis[:i]), })
data.append({'Samples': n_samples,
'Type': 'Harmonic mean',
'Estimate': hmaccum(hm[:i]),
'Relative error': isrelerr(hm[:i]), })
return data