def main():
parser = argparse.ArgumentParser(description='training script')
# data load
parser.add_argument('--data', type=str, default='blizzard', help='blizzard / nancy')
parser.add_argument('--batch_size', type=int, default=32, help='batch size')
parser.add_argument('--text_limit', type=int, default=1000, help='maximum length of text to include in training set')
parser.add_argument('--wave_limit', type=int, default=1400, help='maximum length of spectrogram to include in training set')
parser.add_argument('--trunc_size', type=int, default=700, help='used for truncated-BPTT when memory is not enough.')
parser.add_argument('--shuffle_data', type=int, default=1, help='whether to shuffle data loader')
parser.add_argument('--load_queue_size', type=int, default=8, help='maximum number of batches to load on the memory')
parser.add_argument('--n_workers', type=int, default=2, help='number of workers used in data loader')
# model
parser.add_argument('--charvec_dim', type=int, default=256, help='')
parser.add_argument('--hidden_size', type=int, default=128, help='')
parser.add_argument('--dec_out_size', type=int, default=80, help='decoder output size')
parser.add_argument('--post_out_size', type=int, default=1025, help='should be n_fft / 2 + 1(check n_fft from "input_specL" ')
parser.add_argument('--num_filters', type=int, default=16, help='number of filters in filter bank of CBHG')
parser.add_argument('--r_factor', type=int, default=5, help='reduction factor(# of multiple output)')
parser.add_argument('--dropout', type=float, default=0.5, help='')
# optimization
parser.add_argument('--max_epochs', type=int, default=100000, help='maximum epoch to train')
parser.add_argument('--grad_clip', type=float, default=1, help='gradient clipping')
parser.add_argument('--learning_rate', type=float, default=1e-3, help='2e-3 from Ito, I used to use 5e-4')
parser.add_argument('--lr_decay_every', type=int, default=25000, help='decay learning rate every...')
parser.add_argument('--lr_decay_factor', type=float, default=0.5, help='decay learning rate by this factor')
parser.add_argument('--teacher_forcing_ratio', type=float, default=1, help='value between 0~1, use this for scheduled sampling')
# loading
parser.add_argument('--init_from', type=str, default='', help='load parameters from...')
parser.add_argument('--resume', type=int, default=0, help='1 for resume from saved epoch')
# misc
parser.add_argument('--exp_no', type=int, default=0, help='')
parser.add_argument('--print_every', type=int, default=-1, help='')
parser.add_argument('--plot_every', type=int, default=-1, help='')
parser.add_argument('--save_every', type=int, default=-1, help='')
parser.add_argument('--save_dir', type=str, default='checkpoint', help='')
parser.add_argument('--pinned_memory', type=int, default=1, help='1 to use pinned memory')
parser.add_argument('--gpu', type=int, nargs='+', help='index of gpu machines to run')
# debug
parser.add_argument('--debug', type=int, default=0, help='1 for debug mode')
args = parser.parse_args()
torch.manual_seed(0)
# set dataset option
if args.data == 'blizzard':
args.dir_bin = '/home/lyg0722/TTS_corpus/blizzard/segmented/bin/'
elif args.data == 'etri':
args.dir_bin = '/data2/lyg0722/TTS_corpus/etri/bin/'
else:
print('no dataset')
return
if args.gpu is None:
args.use_gpu = False
args.gpu = []
else:
args.use_gpu = True
torch.cuda.manual_seed(0)
torch.cuda.set_device(args.gpu[0])
loader = DataLoader(args)
# set misc options
args.vocab_size = loader.get_num_vocab()
if args.print_every == -1:
args.print_every = loader.iter_per_epoch
if args.plot_every == -1:
args.plot_every = args.print_every
if args.save_every == -1:
args.save_every = loader.iter_per_epoch * 10 # save every 10 epoch by default
model = Tacotron(args)
model_optim = optim.Adam(model.parameters(), args.learning_rate)
criterion_mel = nn.L1Loss(size_average=False)
criterion_lin = nn.L1Loss(size_average=False)
start = time.time()
plot_losses = []
print_loss_total = 0 # Reset every print_every
plot_loss_total = 0 # Reset every plot_every
start_epoch = 0
iter = 1
if args.init_from:
checkpoint = torch.load(args.init_from, map_location=lambda storage, loc: storage)
model.load_state_dict(checkpoint['state_dict'])
if args.resume != 0:
start_epoch = checkpoint['epoch']
plot_losses = checkpoint['plot_losses']
print('loaded checkpoint %s (epoch %d)' % (args.init_from, start_epoch))
model = model.train()
if args.use_gpu:
model = model.cuda()
criterion_mel = criterion_mel.cuda()
criterion_lin = criterion_lin.cuda()
print('Start training... (1 epoch = %s iters)' % (loader.iter_per_epoch))
while iter < args.max_epochs * loader.iter_per_epoch + 1:
if loader.is_subbatch_end:
prev_h = (None, None, None) # set prev_h = h_0 when new sentences are loaded
enc_input, target_mel, target_lin, wave_lengths, text_lengths = loader.next_batch('train')
max_wave_len = max(wave_lengths)
enc_input = Variable(enc_input, requires_grad=False)
target_mel = Variable(target_mel, requires_grad=False)
target_lin = Variable(target_lin, requires_grad=False)
prev_h = loader.mask_prev_h(prev_h)
model_optim.zero_grad()
pred_mel, pred_lin, prev_h = model(enc_input, target_mel[:, :-1], wave_lengths, text_lengths, prev_h)
loss_mel = criterion_mel(pred_mel, target_mel[:, 1:])\
.div(max_wave_len * args.batch_size * args.dec_out_size)
loss_linear = criterion_lin(pred_lin, target_lin[:, 1:])\
.div(max_wave_len * args.batch_size * args.post_out_size)
loss = torch.sum(loss_mel + loss_linear)
loss.backward()
nn.utils.clip_grad_norm(model.parameters(), args.grad_clip) # gradient clipping
model_optim.step()
print_loss_total += loss.data[0]
plot_loss_total += loss.data[0]
if iter % args.print_every == 0:
print_loss_avg = print_loss_total / args.print_every
print_loss_total = 0
print('%s (%d %d%%) %.4f' % (timeSince(start, iter / args.max_epochs),
iter, iter / args.max_epochs * 100, print_loss_avg))
if iter % args.plot_every == 0:
plot_loss_avg = plot_loss_total / args.plot_every
plot_losses.append(plot_loss_avg)
plot_loss_total = 0
save_name = '%s/%dth_exp_loss.png' % (args.save_dir, args.exp_no)
savePlot(plot_losses, save_name)
if iter % args.save_every == 0:
epoch = start_epoch + iter // loader.iter_per_epoch
save_name = '%s/%d_%dth.t7' % (args.save_dir, args.exp_no, epoch)
state = {
'epoch': epoch,
'args': args,
'state_dict': model.state_dict(),
'optimizer': model_optim.state_dict(),
'plot_losses': plot_losses
}
torch.save(state, save_name)
print('model saved to', save_name)
# if is_best: # TODO: implement saving best model.
# shutil.copyfile(save_name, '%s/%d_best.t7' % (args.save_dir, args.exp_no))
iter += 1
def main():
parser = argparse.ArgumentParser(description='training script')
# data load
parser.add_argument('--data', type=str, default='blizzard', help='blizzard / nancy')
parser.add_argument('--batch_size', type=int, default=6, help='batch size')
parser.add_argument('--text_limit', type=int, default=1500, help='maximum length of text to include in training set')
parser.add_argument('--wave_limit', type=int, default=800, help='maximum length of spectrogram to include in training set')
parser.add_argument('--shuffle_data', type=int, default=0, help='whether to shuffle data loader')
parser.add_argument('--batch_idx', type=int, default=0, help='n-th batch of the dataset')
parser.add_argument('--load_queue_size', type=int, default=1, help='maximum number of batches to load on the memory')
parser.add_argument('--n_workers', type=int, default=1, help='number of workers used in data loader')
# generation option
parser.add_argument('--exp_no', type=int, default=0, help='')
parser.add_argument('--out_dir', type=str, default='generated', help='')
parser.add_argument('--init_from', type=str, default='', help='load parameters from...')
parser.add_argument('--caption', type=str, default='', help='text to generate speech')
parser.add_argument('--teacher_forcing_ratio', type=float, default=0, help='value between 0~1, use this for scheduled sampling')
# audio related option
parser.add_argument('--n_fft', type=int, default=2048, help='fft bin size')
parser.add_argument('--sample_rate', type=int, default=16000, help='sampling rate')
parser.add_argument('--frame_len_inMS', type=int, default=50, help='used to determine window size of fft')
parser.add_argument('--frame_shift_inMS', type=int, default=12.5, help='used to determine stride in sfft')
parser.add_argument('--num_recon_iters', type=int, default=50, help='# of iteration in griffin-lim recon')
# misc
parser.add_argument('--gpu', type=int, nargs='+', help='index of gpu machines to run')
parser.add_argument('--seed', type=int, default=0, help='random seed')
new_args = vars(parser.parse_args())
# load and override some arguments
checkpoint = torch.load(new_args['init_from'], map_location=lambda storage, loc: storage)
args = checkpoint['args']
for i in new_args:
args.__dict__[i] = new_args[i]
torch.manual_seed(args.seed)
# set dataset option
if args.data == 'blizzard':
args.dir_bin = '/data2/lyg0722/TTS_corpus/blizzard/segmented/bin/'
elif args.data == 'etri':
args.dir_bin = '/data2/lyg0722/TTS_corpus/etri/bin/'
else:
print('no dataset')
return
if args.gpu is None:
args.use_gpu = False
args.gpu = []
else:
args.use_gpu = True
torch.cuda.manual_seed(0)
torch.cuda.set_device(args.gpu[0])
model = Tacotron(args)
criterion_mel = nn.L1Loss(size_average=False)
criterion_lin = nn.L1Loss(size_average=False)
window_len = int(np.ceil(args.frame_len_inMS * args.sample_rate / 1000))
hop_length = int(np.ceil(args.frame_shift_inMS * args.sample_rate / 1000))
if args.init_from:
model.load_state_dict(checkpoint['state_dict'])
print('loaded checkpoint %s' % (args.init_from))
model = model.eval()
if args.use_gpu:
model = model.cuda()
criterion_mel = criterion_mel.cuda()
criterion_lin = criterion_lin.cuda()
if args.caption:
text_raw = args.caption
if args.data == 'etri':
text_raw = decompose_hangul(text_raw) # For Korean dataset
vocab_dict = torch.load(args.dir_bin + 'vocab.t7')
enc_input = [vocab_dict[i] for i in text_raw]
enc_input = enc_input + [0] # null-padding at tail
text_lengths = [len(enc_input)]
enc_input = Variable(torch.LongTensor(enc_input).view(1,-1))
dec_input = torch.Tensor(1, 1, args.dec_out_size).fill_(0) # null-padding for start flag
dec_input = Variable(dec_input)
wave_lengths = [args.wave_limit] # TODO: use <EOS> later...
prev_h = (None, None, None) # set prev_h = h_0 when new sentences are loaded
if args.gpu:
enc_input = enc_input.cuda()
dec_input = dec_input.cuda()
_, pred_lin, prev_h = model(enc_input, dec_input, wave_lengths, text_lengths, prev_h)
# start generation
wave = spectrogram2wav(
pred_lin.data.view(-1, args.post_out_size).cpu().numpy(),
n_fft=args.n_fft,
win_length=window_len,
hop_length=hop_length,
num_iters=args.num_recon_iters
)
# write to file
outpath1 = '%s/%s_%s.wav' % (args.out_dir, args.exp_no, args.caption)
outpath2 = '%s/%s_%s.png' % (args.out_dir, args.exp_no, args.caption)
librosa.output.write_wav(outpath1, wave, 16000)
saveAttention(text_raw, torch.cat(model.attn_weights, dim=-1).squeeze(), outpath2)
else:
loader = DataLoader(args)
args.vocab_size = loader.get_num_vocab()
for iter in range(1, loader.iter_per_epoch + 1):
if loader.is_subbatch_end:
prev_h = (None, None, None) # set prev_h = h_0 when new sentences are loaded
for i in range(args.batch_idx):
loader.next_batch('train')
enc_input, target_mel, target_lin, wave_lengths, text_lengths = loader.next_batch('train')
enc_input = Variable(enc_input, volatile=True)
target_mel = Variable(target_mel, volatile=True)
target_lin = Variable(target_lin, volatile=True)
prev_h = loader.mask_prev_h(prev_h)
if args.gpu:
enc_input = enc_input.cuda()
target_mel = target_mel.cuda()
target_lin = target_lin.cuda()
pred_mel, pred_lin, prev_h = model(enc_input, target_mel[:, :-1], wave_lengths, text_lengths, prev_h)
loss_mel = criterion_mel(pred_mel, target_mel[:, 1:]) \
.div(max(wave_lengths) * args.batch_size * args.dec_out_size)
loss_linear = criterion_lin(pred_lin, target_lin[:, 1:]) \
.div(max(wave_lengths) * args.batch_size * args.post_out_size)
loss = torch.sum(loss_mel + loss_linear)
print('loss:' , loss.data[0])
attentions = torch.cat(model.attn_weights, dim=-1)
# write to file
for n in range(enc_input.size(0)):
wave = spectrogram2wav(
pred_lin.data[n].view(-1, args.post_out_size).cpu().numpy(),
n_fft=args.n_fft,
win_length=window_len,
hop_length=hop_length,
num_iters=args.num_recon_iters
)
outpath1 = '%s/%s_%s_%s.wav' % (args.out_dir, args.exp_no, n, args.caption)
librosa.output.write_wav(outpath1, wave, 16000)
outpath2 = '%s/%s_%s_%s.png' % (args.out_dir, args.exp_no, n, args.caption)
saveAttention(None, attentions[n], outpath2)
# showPlot(plot_losses)
break
