def train(self, n_iterations=100, verbose=1, **kwargs):
"""
Takes an optional argument, `n_iterations` and updates the model
`n_iterations` times.
:param n_iterations: Number of iterations. Default is 100.
:type n_iterations: int, optional
:param verbose: If 1, current number of iterations
are printed out to notify the user. Default is 1.
:type verbose: int, optional
:param kwargs: For compatability with calls to LdaCgsMulti.
:type kwargs: optional
"""
random_state = np.random.RandomState(self.seed)
random_state.set_state(self._mtrand_state)
if verbose > 0:
print ('Begin LDA training for {0} iterations'\
.format(n_iterations))
start = time.time()
t = start
# Training loop
stop = self.iteration + n_iterations
pbar = ProgressBar(widgets=[Percentage(), Bar()], maxval=n_iterations).start()
#print("Stop ", stop)
for itr in xrange(self.iteration , stop):
results = cgs_update(self.iteration, self.corpus, self.word_top,
self.inv_top_sums, self.top_doc, self.Z,
self.indices, self._mtrand_state[0],
self._mtrand_state[1], self._mtrand_state[2],
self._mtrand_state[3], self._mtrand_state[4])
lp = results[4]
self.log_probs.append((self.iteration, lp))
if verbose == 2:
itr_time = np.around(time.time()-t, decimals=1)
t = time.time()
if verbose > 1 or itr==stop-1:
print ('\nIteration {0} complete: log_prob={1}, time={2}'
.format(self.iteration, lp, itr_time))
if verbose == 1:
#print("Self iteration", self.iteration)
pbar.update(self.iteration - (stop - n_iterations))
time.sleep(0.01)
self.iteration += 1
self._mtrand_state = results[5:]
pbar.finish();
if verbose > 1:
print '-'*60, ('\n\nWalltime per iteration: {0} seconds'
.format(np.around((t-start)/n_iterations, decimals=2)))
def update((docs, doc_indices, mtrand_state)):
"""
For LdaCgsMulti
"""
start, stop = docs[0][0], docs[-1][1]
corpus = np.frombuffer(_corpus, dtype='i')[start:stop]
Z = np.frombuffer(_Z, dtype='i')[start:stop].copy()
gbl_word_top = np.frombuffer(_word_top, dtype='d')
gbl_word_top = gbl_word_top.reshape(_V.value, _K.value)
loc_word_top = gbl_word_top.copy()
inv_top_sums = np.frombuffer(_inv_top_sums, dtype='d').copy()
top_doc = np.frombuffer(_top_doc, dtype='d')
top_doc = top_doc.reshape(_K.value, top_doc.size/_K.value)
top_doc = top_doc[:, doc_indices[0]:doc_indices[1]].copy()
log_p = 0
log_wk = np.log(gbl_word_top * inv_top_sums[np.newaxis, :])
log_kc = np.log(top_doc / top_doc.sum(0)[np.newaxis, :])
indices = np.array([(j - start) for (i,j) in docs], dtype='i')
results = cgs_update(_iteration.value,
corpus,
loc_word_top,
inv_top_sums,
top_doc,
Z,
indices,
mtrand_state[0],
mtrand_state[1],
mtrand_state[2],
mtrand_state[3],
mtrand_state[4])
(loc_word_top, inv_top_sums, top_doc, Z, log_p, mtrand_str, mtrand_keys,
mtrand_pos, mtrand_has_gauss, mtrand_cached_gaussian) = results
loc_word_top -= gbl_word_top
return (Z, top_doc, loc_word_top, log_p,
mtrand_str, mtrand_keys, mtrand_pos,
mtrand_has_gauss, mtrand_cached_gaussian)
