def transfer(model, decoder, sess, args, vocab, data0, data1, out_path):
batches, order0, order1 = get_batches(data0, data1, vocab.word2id,
args.batch_size)
data0_tsf, data1_tsf = [], []
losses = Losses(len(batches))
for batch in batches:
ori, tsf = decoder.rewrite(batch)
half = batch['size'] / 2
data0_tsf += tsf[:half]
data1_tsf += tsf[half:]
loss, loss_g, loss_d, loss_d0, loss_d1 = sess.run(
[
model.loss, model.loss_g, model.loss_d, model.loss_d0,
model.loss_d1
],
feed_dict=feed_dictionary(model, batch, args.rho, args.gamma_min))
losses.add(loss, loss_g, loss_d, loss_d0, loss_d1)
n0, n1 = len(data0), len(data1)
data0_tsf = reorder(order0, data0_tsf)[:n0]
data1_tsf = reorder(order1, data1_tsf)[:n1]
if out_path:
write_sent(data0_tsf, out_path + '.0' + '.tsf')
write_sent(data1_tsf, out_path + '.1' + '.tsf')
return losses
def transfer(model, sess, args, vocab, data0, data1, out_path):
batches, order0, order1 = get_batches(data0, data1, vocab.word2id,
args.batch_size)
data0_tsf, data1_tsf = [], []
losses = Losses(len(batches))
for batch in batches:
ori, tsf, loss, loss_g, loss_d, loss_d0, loss_d1 = rewrite(
model, sess, args, vocab, batch)
half = batch['size'] / 2
data0_tsf += tsf[:half]
data1_tsf += tsf[half:]
losses.add(loss, loss_g, loss_d, loss_d0, loss_d1)
n0, n1 = len(data0), len(data1)
data0_tsf = reorder(order0, data0_tsf)[:n0]
data1_tsf = reorder(order1, data1_tsf)[:n1]
if out_path:
write_sent(data0_tsf, out_path + '.0' + '.tsf')
write_sent(data1_tsf, out_path + '.1' + '.tsf')
return losses
loss_d1, _ = sess.run([model.loss_d1, model.optimizer_d1],
feed_dict=feed_dict)
# do not back-propagate from the discriminator
# when it is too poor
if loss_d0 < 1.2 and loss_d1 < 1.2:
optimizer = model.optimizer_all
else:
optimizer = model.optimizer_ae
loss, loss_g, loss_d, _ = sess.run(
[model.loss, model.loss_g, model.loss_d, optimizer],
feed_dict=feed_dict)
step += 1
losses.add(loss, loss_g, loss_d, loss_d0, loss_d1)
if step % args.steps_per_checkpoint == 0:
losses.output('step %d, time %.0fs,' %
(step, time.time() - start_time))
losses.clear()
if args.dev:
dev_losses = transfer(model, decoder, sess, args, vocab,
dev0, dev1,
args.output + '.epoch%d' % epoch)
dev_losses.output('dev')
if dev_losses.loss < best_dev:
best_dev = dev_losses.loss
print 'saving model...'
model.saver.save(sess, args.model)