reference.append(sents.replace("\n", ""))
for epoch in range(1, args.epochs + 1):
g_loss, c_loss = train(epoch)
data_test = list()
for i in range(2):
test_noise = torch.Tensor(
np.random.normal(0, 1,
(250, args.latent_size))).to(args.device)
test_z = generator(test_noise).data
new_sent = rollout_test(model_decoder, test_z, tokenizer_decoder,
args.max_seq_length, 250, 0, 1)
data_test.extend(new_sent)
p_reference = random.sample(reference, 500)
bleu = calc_blue_parallel_func(p_reference, data_test, 2, 500)
b_bleu = calc_blue_parallel_func(data_test, p_reference, 2, 500)
logger.info("Bleu-2:{:0.3f} | B-Bleu-2:{:0.3f}".format(bleu, b_bleu))
if (bleu + b_bleu) > best_bleu:
best_bleu = bleu + b_bleu
logger.info(
'* Saving. Best Score:{:0.3f} | Bleu-2:{:0.3f} | B-Bleu-2:{:0.3f}'
.format(best_bleu, bleu, b_bleu))
torch.save(
generator.state_dict(), args.output_dir + '/generator_' +
str(args.gloabl_step_eval) + '.th')
torch.save(
critic.state_dict(), args.output_dir + '/critic_' +
str(args.gloabl_step_eval) + '.th')
reference.append(sents.replace("\n", ""))
for epoch in range(1, args.epochs + 1):
g_loss, c_loss = train(epoch)
data_test = list()
test_lab = torch.LongTensor([0]*100 + [1]*100 + [2]*100 + [3]*100 + [4]*100).to(args.device)
for i in range(5):
test_noise = torch.Tensor(np.random.normal(0, 1, (100, args.latent_size))).to(args.device)
test_z = generator(test_noise, test_lab[100*i:100*(i+1)]).data
new_sent = rollout_test(model_decoder, test_z, tokenizer_decoder, args.max_seq_length, 100, 0, 1)
data_test.extend(new_sent)
p_reference = random.sample(reference, 500)
data_test = [str(lab)+" "+str(sen) for lab,sen in zip(test_lab.tolist(), data_test)]
bleu = calc_blue_parallel_func(p_reference, data_test, 2, 500, True)
b_bleu = calc_blue_parallel_func(data_test, p_reference, 2, 500, True)
logger.info("Bleu-2:{:0.3f} | B-Bleu-2:{:0.3f}".format(bleu, b_bleu))
if (bleu+b_bleu) > best_bleu:
best_bleu = bleu + b_bleu
logger.info('* Saving. Best Score:{:0.3f} | Bleu-2:{:0.3f} | B-Bleu-2:{:0.3f}'.format(best_bleu, bleu, b_bleu))
torch.save(generator.state_dict(), args.output_dir+'/generator_'+str(args.gloabl_step_eval)+'.th')
torch.save(critic.state_dict(), args.output_dir+'/critic_'+str(args.gloabl_step_eval)+'.th')
torch.save(classifier.state_dict(), args.output_dir+'/classifier_'+str(args.gloabl_step_eval)+'.th')
if args.finetune_decoder:
logger.info("Loading generator")
generator.load_state_dict(torch.load(args.output_dir+'/generator_'+str(args.gloabl_step_eval)+'.th'))
model_decoder.train()
generator.eval()