def train_model():
updates = []
log_lik = []
# Train model
print "Parsing dir.."
docs, vocab, word_idx, sent_idx, sentences = data.parse_dir("./data/small/")
l = MgLda(n_local_topics,
n_global_topics,
docs, vocab, word_idx, sentences, sent_idx)
print "Done! Running %d iterations..."%iterations
for i in range(0, iterations):
start = time.time()
l.update()
duration = time.time() - start
updates.append(l.updates)
log_lik.append(l.log_lik)
times.append(duration)
print "Iteration %d, Duration: %f"%(i, duration)
# Save model and data
dir = os.path.join("models/", datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))
os.makedirs(dir)
with open(dir+'/mglda_model_2.pkl', 'wb') as output:
pickle.dump(l, output, -1)
updates = np.asarray(updates)
log_lik = np.asarray(log_lik)
np.savetxt(dir+'/updates.txt', updates)
np.savetxt(dir+'/log_lik.txt', log_lik)
np.savetxt(dir+'/times.txt', np.asarray(times))
print "Model and data saved!"
# Add some plots
print "Adding some plots.."
x = np.arange(0, iterations)
fig, ax = plt.subplots( nrows=1, ncols=1 )
ax.plot(x, log_lik/100000)
plt.xlabel("Iterations")
plt.ylabel("Log likelihood x 10^5")
plt.title("MG-LDA training")
plt.grid(True)
fig.savefig(dir + "/log_lik.png")
plt.close(fig)
# Plot log updates
fig, ax = plt.subplots( nrows=1, ncols=1 )
ax.plot(x, updates)
plt.xlabel("Iterations")
plt.ylabel("Updates")
plt.title("MG-LDA training")
plt.grid(True)
fig.savefig(dir + "/updates.png")
plt.close(fig)
# Return the model
return l
new_z = np.random.multinomial(1, p_z).argmax()
loglikelihood += log(p_z[new_z])
self.assignment[i][w] = new_z
self.nwz[cur_w, new_z] += 1
self.ndz[i, new_z] += 1
self.nz[new_z] += 1
print "Loglikelihood %f" % loglikelihood
self.loglikelihood.append(loglikelihood)
# EXAMPLE
print "Parsing dir.."
docs, vocab, word_idx, _, _ = data.parse_dir("./data/all/")
# docs, vocab, word_idx, _, _ = data.parse_dir("/Users/jeisses/Documents/datasets/nlp/movie/review_polarity/txt_sentoken/all/")
print "Done"
print " ====== "
print "Setting up LDA.."
l = Lda(10, docs, vocab, word_idx)
print "Training for 200 iterations..."
for i in range(0, 200):
l.update()
if i % 10 == 0:
print " -- iteration %d ---" % i
print "Done. Top words for reach topic:"
l.top_words(10)
loglikelihood = [item / 100000 for item in l.loglikelihood]
