G[0, :, :] = bf2f.np.eye(d+1)
G[:, -1, :] = 0
G[:, -1, -1] = 1
initial_params = (C, G, V)
if options['diagnostics_rate'] > 0:
DIAGNOSTICS = True
else:
DIAGNOSTICS = False
# --- initialise parameters --- #
pp = bf2f.params(initial_params, options, vocab)
# --- ll before --- #
if options['calc_ll']:
pre_ll = bf2f.log_likelihood(pp, train_data)
print 'pre ll:', pre_ll
# --- start the logfile --- #
if DIAGNOSTICS:
logf = open(output_root+'_logfile.txt','w')
logf.write('n\ttime\tll\tdata_energy\tmodel_energy\tvaliset_energy\trandom_energy\tperm_energy\tC_lens\tG_lens\tV_lens\n')
logf.close()
# ---- TRAIN! --- #
for epoch in xrange(options['n_epochs']):
print 'epoch:', epoch
cProfile.runctx('vali_set = bf2f.train(train_data, pp, options)', None, locals())
if options['online']:
train_data = dstream.acquire_all(SHUFFLE=True)
# --- actually get the training data --- #
if ONLINE:
train_data = dstream
else:
train_data = dstream.acquire_all()
# --- initialise parameters --- #
pp = bf2f.params((C, G, V),
vocab,
fix_words=fix_words,
fix_relas=fix_relas,
trans_rela=trans_rela)
# --- ll before --- #
if CALC_LL:
print 'pre ll:', bf2f.log_likelihood(pp, train_data)
# --- start the logfile --- #
if DIAGNOSTICS:
logf = open(fname + '_logfile.txt', 'w')
logf.write(
'n\ttime\tll\tdata_energy\tmodel_energy\tvaliset_energy\trandom_energy\tperm_energy\tC_lens\tG_lens\tV_lens\n'
)
logf.close()
else:
print 'WARNING: no diagnostics.'
# ---- TRAIN! --- #
for epoch in xrange(n_epochs):
print 'epoch:', epoch
cProfile.runctx(
relas = map(str, range(R))
vocab = {'words':words, 'relas': relas}
# --- actually get the training data --- #
if ONLINE:
train_data = dstream
else:
train_data = dstream.acquire_all()
# --- initialise parameters --- #
pp = bf2f.params((C, G, V), vocab, fix_words=fix_words, fix_relas=fix_relas, trans_rela=trans_rela)
# --- ll before --- #
if CALC_LL:
print 'pre ll:', bf2f.log_likelihood(pp, train_data)
# --- start the logfile --- #
if DIAGNOSTICS:
logf = open(fname+'_logfile.txt','w')
logf.write('n\ttime\tll\tdata_energy\tmodel_energy\tvaliset_energy\trandom_energy\tperm_energy\tC_lens\tG_lens\tV_lens\n')
logf.close()
else:
print 'WARNING: no diagnostics.'
# ---- TRAIN! --- #
for epoch in xrange(n_epochs):
print 'epoch:', epoch
cProfile.runctx('vali_set = bf2f.train(train_data, pp, options, EXACT, PERSISTENT, NOISE)', None, locals())
if ONLINE:
# (the purpose of this is to shuffle the training data)
