def test_lda_inference_gd():
rng = np.random.RandomState(2)
n_topics = rng.randint(15, 20)
n_words = rng.randint(400, 500)
mean_words = 10
min_words = 3
doc_topic_prior = 1.
topic_word_prior = 1.
n_doc = rng.randint(100, 200)
gen = LdaSampleGenerator(n_topics=n_topics,
n_words=n_words,
min_doc_size=min_words,
mean_doc_size=mean_words,
doc_topic_prior=doc_topic_prior,
topic_word_prior=topic_word_prior,
random_state=0)
_, doc_word_mtx = gen.generate_documents(n_doc)
alpha = gen.doc_topic_prior_
beta = gen.topic_word_distr_
docs_distr_uniform = np.ones((n_doc, n_topics)) / n_topics
ll_uniform = doc_likelihood(doc_word_mtx, docs_distr_uniform, alpha, beta)
inference_1 = lda_inference_gd(doc_word_mtx, alpha, beta, max_iter=5)
ones = np.ones(n_doc)
assert_array_almost_equal(ones, inference_1.sum(axis=1))
ll_inference_1 = doc_likelihood(doc_word_mtx, inference_1, alpha, beta)
inference_2 = lda_inference_gd(doc_word_mtx, alpha, beta, max_iter=1000)
assert_array_almost_equal(ones, inference_2.sum(axis=1))
ll_inference_2 = doc_likelihood(doc_word_mtx, inference_2, alpha, beta)
assert_greater(ll_inference_2, ll_inference_1)
assert_greater(ll_inference_1, ll_uniform)
def test_doc_likelihood():
rng = np.random.RandomState(2)
n_topics = rng.randint(15, 20)
n_words = rng.randint(400, 500)
mean_words = 10
min_words = 3
doc_topic_prior = 1.
topic_word_prior = 1.
n_doc = rng.randint(100, 200)
gen = LdaSampleGenerator(n_topics=n_topics,
n_words=n_words,
min_doc_size=min_words,
mean_doc_size=mean_words,
doc_topic_prior=doc_topic_prior,
topic_word_prior=topic_word_prior,
random_state=0)
docs_distr, doc_word_mtx = gen.generate_documents(n_doc)
alpha = gen.doc_topic_prior_
beta = gen.topic_word_distr_
doc_ll_true = doc_likelihood(doc_word_mtx, docs_distr, alpha, beta)
# uniform distribution
docs_distr_uniform = np.ones((n_doc, n_topics)) / n_topics
doc_ll_uniform = doc_likelihood(doc_word_mtx, docs_distr_uniform, alpha,
beta)
assert_greater(doc_ll_true, doc_ll_uniform)
def test_lda_inference_vi():
rng = np.random.RandomState(4)
n_topics = rng.randint(15, 20)
n_words = rng.randint(400, 500)
mean_words = 10
min_words = 3
doc_topic_prior = 1.
topic_word_prior = 1.
smooth_param = 0.01
n_doc = rng.randint(100, 200)
gen = LdaSampleGenerator(n_topics=n_topics,
n_words=n_words,
min_doc_size=min_words,
mean_doc_size=mean_words,
doc_topic_prior=doc_topic_prior,
topic_word_prior=topic_word_prior,
random_state=2)
_, doc_word_mtx = gen.generate_documents(n_doc)
alpha = gen.doc_topic_prior_
beta = gen.topic_word_distr_
smooth_beta = (1. - smooth_param) * beta + (smooth_param / n_words)
docs_distr_uniform = np.ones((n_doc, n_topics)) / n_topics
ll_uniform = doc_likelihood(doc_word_mtx, docs_distr_uniform, alpha, beta)
inference_1 = lda_inference_vi(doc_word_mtx,
alpha,
smooth_beta,
max_iter=5)
inference_1 /= inference_1.sum(axis=1)[:, np.newaxis]
ll_inference_1 = doc_likelihood(doc_word_mtx, inference_1, alpha, beta)
inference_2 = lda_inference_vi(doc_word_mtx,
alpha,
smooth_beta,
max_iter=1000)
inference_2 /= inference_2.sum(axis=1)[:, np.newaxis]
ll_inference_2 = doc_likelihood(doc_word_mtx, inference_2, alpha, beta)
assert_greater(ll_inference_2, ll_inference_1)
assert_greater(ll_inference_1, ll_uniform)