def train(self):
generator = Generator(self.embedding, self.vocab_size,
config.att_embedding_size, 2, config.ber_prob)
generator = generator.to(config.device)
discriminator = Discriminator(2, config.dec_hidden_size,
config.enc_hidden_size)
discriminator = discriminator.to(config.device)
self.g_optim = optim.Adam(params=generator.parameters(),
lr=config.g_lr)
self.d_optim = optim.Adam(params=discriminator.parameters(),
lr=config.d_lr)
criterion = nn.CrossEntropyLoss(ignore_index=0)
num_step = 0
best_loss = 1e10
batch_nb = len(self.train_loader)
for epoch in range(1, config.num_epochs + 1):
start = time.time()
for i, train_data in enumerate(self.train_loader):
batch_idx = i + 1
num_step += 1
recon_loss, errG, errD = self.step(generator, discriminator,
criterion, train_data)
generator_loss = recon_loss + errG * config.loss_lambda
# generator_loss = recon_loss
msg = "{}/{} {} - ETA : {} - recon : {:.4f}, loss G: {:.4f}, loss D: {:.4f}".format(
batch_idx, batch_nb, progress_bar(batch_idx, batch_nb),
eta(start, batch_idx, batch_nb), recon_loss, errG, errD)
print(msg)
# only see reconstruction loss
dev_loss = self.evaluate(generator)
msg = "Epoch {} took {} - final loss : {:.4f} - validation loss : {:.4f}" \
.format(epoch, user_friendly_time(time_since(start)), generator_loss, dev_loss)
print(msg)
if dev_loss < best_loss:
best_loss = dev_loss
generator.save(config.save_dir, epoch, num_step)
# log data for visualization later
log_parameters = train_utils.log_epoch_stats(
log_parameters,
epoch,
train_loss,
train_r_sq,
train_accu,
val_loss,
val_r_sq,
val_accu
)
# print loss
if epoch % PRINT_EVERY == 0:
print('='*20 + '\nEpoch %d / %d\n' % (epoch, NUM_EPOCHS) + '='*20)
print('[%s (%d %.1f%%)]' % (train_utils.time_since(START), epoch, float(epoch) / NUM_EPOCHS * 100))
print('[%s %0.5f, %s %0.5f, %s %0.5f]'% ('Train Loss: ', train_loss, ' R-sq: ', train_r_sq, ' Accu:', train_accu))
print('[%s %0.5f, %s %0.5f, %s %0.5f]'% ('Valid Loss: ', val_loss, ' R-sq: ', val_r_sq, ' Accu:', val_accu))
# save model if best validation loss
if val_loss < best_val_loss:
n = file_info + '_best'
train_utils.save(n, perf_model)
best_val_loss = val_loss
print("Saving...")
train_utils.save(file_info, perf_model, log_parameters)
except KeyboardInterrupt:
print("Saving before quit...")
train_utils.save(file_info, perf_model, log_parameters)
print("----------------")
print()
# 학습 시작
try:
for iter in range(1, n_iter + 1):
input, target = get_batch_set()
loss = train(input, target)
total_loss += loss
# 현재 학습 과정 출력
if iter % print_every == 0:
avg_loss = total_loss / print_every
sys.stdout.write(
"%d %d%% (%s) %.4f\n" %
(iter, iter / n_iter * 100, time_since(start), avg_loss))
losses.append(avg_loss)
total_loss = 0
lyrics = generate_lyrics(['사랑', '발라드'])
print(lyrics)
print()
sys.stdout.write("학습이 완료되었습니다.\n")
# 중단 시그널 핸들링
except KeyboardInterrupt:
print("학습이 중단되었습니다.")
pass
# 손실 그래프 출력
plt.figure()
perf_model, criterion, perf_optimizer, aug_training_data,
aug_validation_data, METRIC, MTYPE, CTYPE)
# adjut learning rate
# train_utils.adjust_learning_rate(perf_optimizer, epoch, ADJUST_EVERY)
# log data for visualization later
log_value('train_loss', train_loss, epoch)
log_value('val_loss', val_loss, epoch)
log_value('train_r_sq', train_r_sq, epoch)
log_value('val_r_sq', val_r_sq, epoch)
log_value('train_accu', train_accu, epoch)
log_value('val_accu', val_accu, epoch)
log_value('train_accu2', train_accu2, epoch)
log_value('val_accu2', val_accu2, epoch)
# print loss
if epoch % PRINT_EVERY == 0:
print('[%s (%d %.1f%%)]' % (train_utils.time_since(START), epoch,
float(epoch) / NUM_EPOCHS * 100))
print('[%s %0.5f, %s %0.5f, %s %0.5f %0.5f]' %
('Train Loss: ', train_loss, ' R-sq: ', train_r_sq, ' Accu:',
train_accu, train_accu2))
print('[%s %0.5f, %s %0.5f, %s %0.5f %0.5f]' %
('Valid Loss: ', val_loss, ' R-sq: ', val_r_sq, ' Accu:',
val_accu, val_accu2))
# save model if best validation loss
if val_loss < best_val_loss:
n = file_info + '_best'
train_utils.save(n, perf_model)
best_val_loss = val_loss
print("Saving...")
train_utils.save(file_info, perf_model)
except KeyboardInterrupt: